CN101140622A - Method and device for extracting signal - Google Patents

Abstract

There are included: an initial sub-signal creation section which creates, from an original signal, sub-signals of shorter length than the original signal; a created sub-signal selection section which, for the sub-signals produced by the initial sub-signal creation section, prunes the created sub-signal candidates to those for which the amount of data is less than for the original signal; a sub-signal re-creation section which determines a created sub-signal which is actually to be used, using the created sub-signal candidates produced by the created sub-signal selection section; a compression mapping determination section which determines, from the sub-signals produced by the sub-signal re-creation section, a mapping for calculation of a compressed signal; and a signal compression section which calculates a compressed signal corresponding to the sub-signals obtained by the sub-signal re-creation section, based upon the mapping obtained by the compression mapping determination section.

Description

Signal retrieval method and device

The application is a divisional application of the following applications:

the invention name is as follows: signal compression method, signal retrieval method, signal compression apparatus, signal retrieval apparatus, signal compression program, and signal retrieval program

Application date: 12/month and 4/2003

Application No.: 200310121679.1

Technical Field

The present invention relates to a signal compression method, a signal compression apparatus, a program and a recording medium suitable for expressing a large number of signal series with a small amount of information, and a signal search method, a signal search apparatus, a program and a recording medium suitable for searching out a signal location similar to a previously registered signal from the large number of signal series using the signal compression method.

Background

As a conventional signal search method for finding a signal location similar to an input signal from a huge signal series, a high-speed signal search method for finding a similar signal location at a high speed is known (for example, patent document 1: japanese patent No. 3065314, patent document 2: japanese patent laid-open No. 2001-0924).

As a method for reducing the calculation cost of feature verification in which signal feature compression is not performed every time, a method for effectively reducing the dimension (sub-element) of a feature by using a line-form map for division of time-series signal continuity is known (for example, non-patent document 3). The method does not construct the image by uniformly dividing the signal according to the signal property. Further, there is known a method of changing the division length according to the signal property (for example, non-patent document 1.

As a method of omitting a search locally, a time-series active search method is known in which a lower limit value of a distance at an unverified time is calculated using a histogram property as a search feature, and a search at a time not exceeding a predetermined threshold is omitted (for example, patent document 1: japanese patent No. 3065314, patent document 2: japanese patent application laid-open No. 2001-092486).

As a signal compression technique used in the above-described signal search method, a signal compression technique is known in which the number of dimensions of a feature series extracted from an acoustic signal is reduced, the feature series is stored in a small amount of memory, and similarity determination between signals can be performed at high speed. The signal search method, in which the signal compression technique is used to detect the time when a specific music piece is played from a reproduced acoustic signal and the like, and to find out a signal location similar to a signal registered in advance from a signal sequence, enables the high-speed execution. Further, as a specific signal compression method used in this technique, a signal compression method is known in which a previously prepared original signal is divided and compressed (for example, non-patent document 3: woodcun wu, etc.; search for high-speed video signal based on line mapping, news technical report, society of electronic information and communications, in 14 years and 2 months, vol.101 No.653, p.75 to 80).

Here, the high-speed signal search methods of patent documents 1 and 2 have a problem that similar signals cannot be found in a very short time for extremely large signal sequences, and a problem that the methods cannot be applied to features other than the histogram.

Further, the signal compression methods of patent documents 1 and 2 have a problem that signal compression can be performed at a high compression ratio, and a huge processing is required for the processing of determining the optimum division.

In the signal compression method described in non-patent document 3, the partial signals are configured not by dividing the original signal into equal parts according to the signal properties, but the division length may be changed according to the signal properties. Therefore, this method, which can compress a signal at a higher compression ratio, has a problem that a huge amount of processing is required for determining a function to be compressed.

Disclosure of Invention

The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to provide a signal compression method, a signal compression apparatus, a signal compression program, a recording medium, and a signal search method using the signal compression method, which can perform signal compression processing more suitable for signal properties than the conventional signal compression method while avoiding a huge amount of prior processing, and can express a signal sequence with a smaller amount of information, and a signal search method, a signal search apparatus, a signal search program, and a recording medium, which are higher in computational efficiency than the conventional method, can guarantee the same search result, and can perform a higher-speed signal search without extremely increasing a temporary storage capacity required for the search.

The present invention is a signal compression method for converting an original signal prepared in advance into a compressed signal, the method including: an initial partial signal forming process of forming a partial signal having a length shorter than that of the original signal from the original signal; a partial signal composition selection step of inserting partial signals, which are candidates for partial signal composition and have a smaller data amount than the original signals, into the partial signals derived in the initial partial signal composition step; using the partial signal reconstruction candidates derived in the partial signal reconstruction selection process to determine a partial signal reconstruction process using the partial signal reconstruction candidates derived in the partial signal reconstruction selection process; a compressed image determining step of determining an image for calculating a compressed signal from each partial signal obtained in the partial signal reconstructing step; and a signal compression method characterized by calculating a signal compression process of a compressed signal corresponding to each partial signal obtained in the partial signal reconstruction process, based on the map obtained in the compressed map determination process.

In the signal compression process, the invention comprises the following steps: a signal mapping process for mapping each partial signal obtained in the partial signal reconstruction process according to the map obtained in the compressed map determination process; a projection distance calculation process of calculating a distance between the partial signal after the image derived in the compressed image determination process and the partial signal obtained in the partial signal reconstruction process; and according to the projection distance derived in the projection distance calculation process and the partial signals after the image derived in the signal image mapping process, the compression characteristic forming process for forming the compression signals is characterized.

In the initial partial signal forming process, the present invention is characterized in that the original signal is divided from the beginning and each divided partial signal is not taken as a partial signal.

In the partial signal configuration selecting process and the partial signal reconstructing process, the present invention is characterized in that the division boundary is determined from the beginning of the original signal.

In the partial signal configuration selecting process and the partial signal reconstructing process, the present invention is characterized in that a predetermined range of possible movement of the division boundary is set, and the division boundary obtained in the initial partial signal configuration process is used as a reference, and the actually used division boundary is determined in the range of possible movement of the division boundary having the range of possible movement of the division boundary in front and in back.

In the partial signal composition selection process, the present invention is characterized in that the compression efficiency is calculated for a plurality of points moving the division boundary, and the range in which the actually used division boundary can exist is selected based on the result.

In the partial signal configuration selecting process, the present invention is characterized in that the number of compression efficiency calculations in the partial signal configuration selecting process is automatically obtained by reducing the number of compression efficiency calculations in the partial signal configuration selecting process and the partial signal reconstructing process.

In the initial partial signal configuration selection process, the present invention is characterized in that features are extracted from the original signal, and features expressed as a multidimensional vector series are reused as the original signal.

The present invention is a signal search method for calculating a distance from an arbitrary point of an accumulated signal, which is a pre-registered original signal, to a reference signal, which is a set target signal, and finding a point similar to the reference signal from the accumulated signal, the method including: a process in the signal compression method; a reference feature extraction process for deriving a series of features from the reference signal; setting a gazing window for the accumulated signals, and deriving an accumulated characteristic extraction process of a characteristic series from the signals in the gazing window; a reference feature compression step of compressing the reference feature series derived in the reference feature extraction step, based on the map derived in the compressed map determination step; a feature checking process of calculating a distance between the reference compressed signal derived in the reference feature compressing process and the cumulative compressed signal derived in the signal compressing process by reusing the feature series derived in the cumulative feature extracting process; and a signal search method for repeating a process according to the feature search process and the signal detection determination process while shifting a window of interest, as a feature, in a signal detection determination process for determining whether or not a reference signal is present at a point of an accumulated signal by comparing a distance derived in the feature search process with a search threshold which is a threshold corresponding to the distance.

In the signal search method of the present invention, a distance recalculation step of calculating a distance between the feature series derived in the reference feature extraction step and the feature series derived in the cumulative feature extraction step at the location having the database signal determined to be the presence of the query signal in the signal detection determination step; and a signal detection re-determination step of determining whether the query signal is present at the location of the database signal by comparing the distance derived in the distance re-calculation step with a search threshold, and repeating the processing according to the feature check step, the signal detection determination step, the distance re-calculation step, and the signal detection re-determination step while shifting the viewing window, thereby calculating the distance to the query signal at a plurality of locations of the database signal, and determining whether the query signal is present at the location of the database signal as a feature.

In the signal retrieval method of the present invention, the jump width of the gaze window is calculated based on the distance calculated in the feature check process, and the process of calculating the jump width of the gaze window is repeated while shifting the gaze window only by the jump width, and the distance from the query signal is calculated for several locations of the database signal, and it is determined whether the query signal exists at the location of the database signal.

The present invention is a signal compression apparatus for compressing an original signal prepared in advance and converting the original signal into a compressed signal, and includes: an initial partial signal forming section for forming a partial signal having a length shorter than that of the original signal from the original signal; a partial signal configuration selecting unit configured to squeeze a candidate partial signal configuration having a smaller data amount than the original signal into each partial signal derived by the initial partial signal configuring unit; a partial signal reconstruction unit for determining a partial signal configuration to be actually used, using the partial signal configuration candidates derived by the partial signal configuration selection unit; a compressed image determining unit for determining an image for calculating a compressed signal from the partial signals obtained by the partial signal reconstructing unit; and a signal compression unit for calculating a compressed signal corresponding to each partial signal obtained by the partial signal configuration selection unit based on the map obtained by the compressed map determination unit.

The present invention is a signal search device for calculating a distance from a reference signal as a set target signal to an arbitrary point of an accumulated signal as an original signal registered in advance, and finding a point similar to the reference signal from the accumulated signal, the signal search device including: a component provided in the signal compression device; a reference feature extracting unit that derives a feature from the reference signal; an accumulated feature extraction unit that sets a gaze window for the accumulated signal and derives a feature from the signal in the gaze window; a reference feature compressing unit that compresses the reference feature derived by the reference feature extracting unit, based on the map derived by the compressed map determining unit; a feature checking unit for calculating a distance between the reference compressed signal derived by the reference feature compressing unit and the accumulated compressed signal derived by the accumulated feature compressing unit by reusing a feature series derived by repeating the accumulated feature compressing unit while shifting a gaze window; a signal detection determining unit that determines whether or not the reference signal is present at the point of the accumulated signal by comparing the distance derived by the feature checking unit with a search threshold value that is a threshold value corresponding to the distance; and a signal search device for repeatedly operating the feature check means and the signal detection determination means as features while shifting the injection window.

The signal search device of the present invention includes: a distance recalculating unit that recalculates a distance between the feature series derived by the reference feature extracting unit and the feature series derived by the cumulative feature extracting unit at the location of the database signal determined by the signal detection determining unit to be the query signal; and a signal detection re-determination unit for comparing the distance derived by the distance re-calculation unit with a search threshold and determining whether the query signal is present at the location of the data signal, wherein the processing by the characteristic-based checking unit, the signal detection determination unit, the distance re-calculation unit, and the signal detection re-determination unit is repeated while shifting the viewing window, and the distance to the query signal is calculated for several locations of the database signal, and whether the query signal is present at the location of the database signal is determined.

The present invention is characterized in that the signal search device includes a jump width calculation unit that calculates a jump width of the gaze window based on the distance calculated by the feature check unit, and moves the gaze window only by the jump width, and repeats the processing by the feature check unit, the signal detection determination unit, and the jump width calculation unit while shifting the gaze window, and calculates the distance to the query signal for several locations of the database signal, and determines whether the query signal is present at the location of the database signal.

The present invention is a signal compression program for compressing an original signal prepared in advance into a compressed signal, the signal compression program being for executing on a computer: an initial partial signal composing process of composing a partial signal having a shorter length than the original signal from the original signal; forming a selection process of partial signal composition candidates for partial signals derived in the initial partial signal composition process, the partial signals having a smaller data amount than the original signals; partial signal reconstruction processing for constructing candidates of partial signals derived in the selection processing by using the partial signals and determining partial signals constructed by the partial signals to be actually used; compressed image determining processing for determining an image for calculating a compressed signal from each partial signal obtained in the partial signal reconstructing processing; and a signal compression program for signal compression processing for calculating compressed signals corresponding to the respective partial signals obtained in the partial signal reconstruction processing, based on the map obtained in the compressed map determination processing.

The present invention is a signal search program for calculating a distance from a reference signal, which is a target signal, to an arbitrary point of an accumulated signal, which is an original signal registered in advance, and finding a point similar to the reference signal from the accumulated signal, the program being configured to execute, on a computer: processing included within the signal compression program; a reference feature extraction process of deriving a feature from the reference signal; setting a note window for the accumulated signal, and deriving accumulated characteristic extraction processing of the characteristic from the signal in the note window; a reference feature compression process of calculating a reference feature derived in the reference feature extraction process based on the map derived in the compressed map determination process; a feature check process of calculating a reference compressed signal derived in the reference feature compression process, and repeating the cumulative feature extraction process while shifting the observation window, thereby obtaining a distance from the cumulative compressed signal derived in the signal compression process; a signal detection determination process of determining whether or not a reference signal is present at the point of the integrated signal by comparing the distance derived in the feature check process with a search threshold value which is a threshold value corresponding to the distance; and a signal search program for repeatedly executing the feature check processing and the signal detection determination processing while shifting the annotation window.

A distance recalculation process of calculating a distance between the feature series derived in the reference feature extraction process and the feature series derived in the distance recalculation process for the location of the database signal determined to be the query signal in the signal detection determination process, in addition to the process included in the signal retrieval program; and a signal search program for comparing the distance derived in the distance recalculation processing with a search threshold value, and performing a signal detection re-determination processing for determining whether or not the query signal is present at the location of the database signal, repeating the feature check processing, the signal detection determination processing, the distance recalculation processing, and the signal detection re-determination processing while shifting the observation window, calculating distances to the query signal at a plurality of locations of the database signal, and performing a processing for determining whether or not the query signal is present at the location of the database signal.

The present invention is a signal retrieval program characterized by comprising, in addition to the processing included in the signal retrieval program, a step width calculation process for calculating a step width of a gaze window based on the distance calculated in the feature check process, and a step width calculation process for moving the gaze window only by the step width, repeating the feature check process, the signal detection determination process, and the step width calculation process while shifting the gaze window, calculating distances to query signals at several locations of a database signal, and performing a process for determining whether or not the query signal is present at the location of the database signal.

The present invention is a computer-readable recording medium recording a signal compression program for compressing an original signal prepared in advance into a compressed signal, the computer-readable recording medium recording a program for executing: an initial partial signal composing process for composing, from an original signal, a partial signal having a length shorter than that of the original signal; a partial signal configuration selection process of inserting partial signals, which are candidates for partial signal configuration and have a smaller data amount than the original signals, into the partial signals derived in the initial partial signal configuration process; partial signal reconstruction processing for determining a partial signal configuration to be actually used by using the partial signal configuration candidates derived in the partial signal configuration selection processing; a compressed image determining process for determining an image for calculating a compressed signal from each partial signal obtained in the partial signal reconstructing process; and a computer-readable recording medium storing a signal compression program for signal compression processing for calculating compressed signals corresponding to the partial signals obtained in the partial signal reconstruction processing, based on the map obtained in the compressed map determination processing.

The present invention is a computer-readable recording medium storing a signal search program for calculating a distance from an arbitrary point of an accumulated signal as a pre-registered original signal to a reference signal as a set target signal and for finding a point similar to the reference signal from the accumulated signal, the program being recorded in a computer for: processing included within the signal compression program; a reference feature extraction process for deriving a feature from the reference signal; setting a gazing window for the accumulated characteristics, and extracting and processing the accumulated characteristics of the characteristics from the signals in the gazing window; a reference feature compression process of compressing the reference feature derived in the reference feature process on the basis of the map derived in the compressed map determination process; a feature matching process of calculating a reference compressed signal derived by the reference feature compression process, and repeatedly performing the cumulative feature extraction process while shifting a gaze window, thereby matching a distance between the reference compressed signal and the cumulative compressed signal derived by the signal compression process; a signal detection determination process of determining whether or not a reference signal is present at the point of the accumulated signal by comparing the distance derived in the feature check process with a search threshold value which is a threshold value corresponding to the distance; and a computer-readable recording medium storing a signal search program for repeatedly executing the feature check processing and the signal detection determination processing while shifting a window of interest.

A distance recalculation process of calculating a distance between the feature series derived in the reference feature extraction process and the feature series derived in the cumulative feature extraction process for the location where the database signal determined to be the query signal in the signal detection determination process is present, in addition to the processes included in the signal search program; and a computer-readable recording medium storing a signal search program for performing a process of comparing the distance derived in the distance recalculation process with a search threshold, and determining whether or not the query signal is present at the location of the database signal, and repeating the feature check process, the signal detection determination process, the distance recalculation process, and the signal detection re-determination process while shifting the window, thereby calculating the distance to the query signal for several locations of the database signal, and determining whether or not the query signal is present at the location of the database signal.

The present invention is a computer-readable recording medium recording a signal search program characterized by comprising a step of calculating a jump width of a gaze window based on a distance calculated in the feature check step, a step of moving the jump width calculation step of the gaze window only by the jump width, a step of repeating the feature check step, the signal detection judgment step, and the jump width calculation step while shifting the gaze window, a step of calculating a distance to a query signal for a plurality of locations of a database signal, and a step of determining whether the query signal is present at the location of the database signal.

According to the present invention, by using a new signal compression method or signal compression apparatus having a partial signal configuration process and a partial signal reconstruction process, the partial signal length is changed in accordance with the signal properties while avoiding enormous prior processing, thereby compressing the original signal prepared in advance more than the conventional method, and obtaining an effect that the signal series can be represented with a smaller amount of information.

In addition, in a signal search method and a signal search device which search for a portion similar to a target reference signal from a previously registered accumulated signal, by using the signal compression method and the signal compression device, it is possible to compress characteristic information, thereby achieving a higher search speed and achieving an effect of reducing the amount of accumulated information.

The present invention is a signal search method for finding out a portion similar to a target inquiry signal (reference signal) from a database signal (accumulated signal) registered in advance, and the method is a inquiry feature extraction process (reference feature extraction process) having a feature derived from the inquiry signal; a database feature extraction step (cumulative feature extraction step) of setting a gazing window for a database signal and deriving features from the signal in the gazing window; a database feature extraction process for extracting a plurality of feature series from the database features in a sequence of the input data; extracting representative features from the feature series obtained after the database feature distinguishing process, and deriving a database feature sparse process of the representative feature series consisting of fewer features; a feature region extraction step of deriving a region in which a feature included in the discrimination derived in the database feature discrimination step exists; a feature checking process of calculating a distance between the feature series derived in the query feature extraction process and the representative feature series derived in the database feature thinning process; a distance correction step of correcting the distance calculated in the feature check step using the region derived in the feature region extraction step; and a signal search method for calculating distances to the query signals for several locations of the database signals, and determining whether the query signals are present at the location of the database signals as a feature by comparing the corrected distances derived in the distance correction process with a search threshold which is a threshold corresponding to the distances, and repeating the processing from the feature search process to the signal detection determination process while shifting the annotation window.

According to the present invention, it is possible to ensure the same search result as compared with the conventional signal search method, and it is possible to perform a higher-speed search without increasing the storage capacity required for the search to some extent, since unnecessary search is partially omitted.

In addition, the signal retrieval method of the present invention is characterized in that any one of the features in the predetermined division is used as a representative feature in the thinning process of the database features.

In addition, the signal retrieval method is characterized in that the center of gravity of the features in the specified division is taken as a representative feature in the sparse process among the features of the database.

The signal search method according to the present invention is characterized in that the database feature extraction process is repeated while shifting the gaze window, and the feature series derived by the database feature extraction process is equally divided by a predetermined length in the database feature division process.

The signal search method according to the present invention is characterized in that the feature series derived by repeating the database feature extraction process while shifting the attention window in the database feature classification process is divided so that the region in which the feature derived in the feature region extraction process exists is reduced to be smaller than a predetermined maximum region.

The signal searching method comprises the following steps: a segment extraction step of extracting segments (partial signals) as a partial series by dividing a feature series derived by repeating the database feature extraction step while shifting a gaze window, a compressed image determination step of determining each segment obtained from the segment extraction step to calculate a feature image lower than the feature dimension, a database feature compression step (signal compression step) of calculating features lower than the feature dimension corresponding to the segment obtained in the segment extraction step based on the image obtained in the compressed image determination step, and a query feature compression step (reference feature compression step) of calculating features lower than the feature dimension corresponding to the features obtained in the query feature extraction step based on the image obtained in the compressed image determination step; in the database feature thinning process, the compressed feature series derived in the database feature compression process is used as a new feature series derived representative feature series; in the feature check process, the compressed features derived in the query feature compression process are checked as new features, and further, the processes from the feature check process to the signal detection and determination process are repeated while the gaze window is shifted, and the distances from the query signal to several locations of the database signal are calculated, and it is determined whether the query signal exists in the location of the database signal as a feature.

According to the present invention, by reducing the number of features added to an index significantly, it is possible to perform a search with a smaller storage capacity.

The invention is to have in the said database characteristic compression process of the said signal retrieval method, according to the said compressed mapping decision process in the mapping, the said segmentation draws the database characteristic mapping process that obtains the segmentation in the course, to the compressed characteristic series that the characteristic mapping process of the said database derives, calculate the database projection calculation process with the characteristic series distance that the characteristic series derives in the course of extracting the said database characteristic, and form the database compressed characteristic composing process of the new compressed characteristic series by the compressed characteristic series that the said database characteristic mapping process derives and the projected distance that the said database projects and derives in the calculation process of the distance; in the query feature compression process, a query feature mapping process of mapping the features obtained in the query feature extraction process according to the map obtained in the compressed map determination process, a query projection distance calculation process of calculating a distance from the features derived in the query feature extraction process for the compressed features derived in the query feature mapping process, and a query compressed feature formation process of forming new compressed features from the compressed features derived in the query feature mapping process and the projection distances derived in the query projection distance calculation process are provided as features.

In the compressed image determining step of the signal search method, the present invention is characterized in that the representative feature is extracted in accordance with K arhunen-Loeve expansion.

The present invention is the signal search method described above, which includes a distance recalculation step of calculating a distance between a feature derived in the query feature extraction step and a feature series derived in the database feature extraction step for the location of the database signal determined to be the presence of the query signal in the signal detection determination step; and a signal detection re-determination process for determining whether the query signal is present at the location of the database signal by comparing the distance derived in the distance re-calculation process with a search threshold, and repeating the processing according to the characteristic checking process, the distance correction process, the signal detection determination process, the distance re-calculation process, and the signal detection re-determination process while shifting the viewing window.

The invention is based on the signal retrieval method, and the database feature classification process comprises classifying each feature derived in the database feature extraction process by staggering a note window and repeatedly performing the classification representative feature according to the predefined distance; a selection threshold setting process for calculating a selection threshold defining a distance in the database feature classification process from a predetermined search threshold; and selecting, as a feature, a database feature selection process having a feature included in the representative feature classification that satisfies a condition derived from the calculated selection threshold in the selection threshold setting process, with respect to a distance between the classification derived in the database feature classification process and the feature derived in the query feature extraction process.

According to the present invention, by narrowing the search range, the projected image can be narrowed down less for the signal to be searched. Therefore, the calculation cost can be reduced.

In the database feature classification process of the signal retrieval method, the features are classified according to a vector quantization algorithm, and Euclidean distance is used as a distance scale.

And the present invention is characterized in that the distance is calculated from one of a manhattan distance or a euclidean distance in the feature verification process of the signal retrieval method.

And the invention is characterized in that the calculation of the distance is based on one of the manhattan distance or the euclidean distance in the calculation process of the database projection distance of the signal retrieval method.

And the present invention is characterized in that the distance recalculation process of the signal retrieval method calculates the distance based on one of a manhattan distance or a euclidean distance.

The present invention classifies the features according to a predetermined method by the query feature extraction process and the database feature extraction process of the signal retrieval method, creates a histogram as a frequency distribution table for each classification, and outputs the histogram as a new feature.

The present invention is characterized in that the signal retrieval method calculates the jump width of the gaze window based on the distance calculated in the distance correction process, moves the jump width calculation process of the gaze window only by the jump width, repeats the processing of the characteristic checking process, the distance correction process, the signal detection determination process, and the jump width calculation process while shifting the gaze window, calculates the distance to the query signal for several locations of the database signal, and determines whether the query signal exists at the location of the database signal.

The present invention is a signal retrieval method for finding out a portion similar to a target query signal from a pre-registered database signal, and the method includes: a query feature extraction process of deriving a feature from the query signal; setting a gazing window for a database signal, and deriving a database characteristic extraction process of the characteristics from the signal in the gazing window; classifying each feature derived in the database feature extraction process while staggering a watching window according to a predefined distance, and determining a database feature classification process of the classification representative feature; a selection threshold setting process of calculating a selection threshold for a distance defined in the database feature classification process from a predetermined search threshold; a database feature selection step of selecting a database feature having a feature included in the representative feature classification which satisfies a condition derived from the calculated selection threshold in the selection threshold setting step, from among the classifications derived in the database feature classification step, the distance from the feature derived in the query feature extraction step; extracting a segmentation extraction process which is a segmentation of a partial series by a method of segmenting features derived by repeating the database feature extraction process while staggering the annotation windows; a compressed image determining step of determining a feature image for calculating a feature image lower than the feature dimension from each segment obtained in the segment extracting step; a database feature compression process of calculating features lower than the feature dimension corresponding to the segments obtained in the segment extraction process, based on the map obtained in the compressed map determination process; a query feature compression process of calculating a feature having a lower dimension than the feature dimension, which corresponds to the feature obtained in the query feature extraction process, based on the map obtained in the compressed map determination process; a feature checking process of calculating a distance between the compression feature series derived in the database feature compression process and the compression feature derived in the query feature extraction process; and a signal search method for determining whether or not the query signal exists at the location of the database signal by comparing the distance calculated in the feature check process with a search threshold value which is a threshold value corresponding to the distance, repeating the processing from the feature check process to the signal detection determination process while shifting the attention window, calculating the distance from the query signal for several locations of the database signal, and determining whether or not the query signal exists at the location of the database signal.

According to the present invention, by narrowing the search range, it is possible to narrow the map used for the signal intended to be searched less. Therefore, the calculation cost can be cut.

The present invention is the signal retrieval method, which comprises a distance recalculation step of calculating the distance between the feature derived in the query feature extraction step and the feature series derived in the database feature extraction step for the location of the database signal determined to be the presence of the query signal in the signal detection and determination step; and a signal detection re-determination process for determining whether the query signal is present at the location of the database signal by comparing the distance derived in the distance re-calculation process with the search threshold, and repeating the processing according to the characteristic checking process, the signal detection determination process, the distance re-calculation process, and the signal detection re-determination process while shifting the viewing window, thereby calculating the distance to the query signal at several locations of the database signal and determining whether the query signal is present at the location of the database signal as a characteristic.

The present invention is characterized in that the signal retrieval method includes a jump width calculation step of calculating a jump width of the window based on the distance calculated in the feature check-up step and a jump width calculation step of moving the window only by the jump width, and the distance from the query signal is calculated for several locations of the database signal by repeating the processing according to the feature check-up step, the signal detection judgment step and the jump width calculation step while shifting the window, and whether the query signal is present at the location of the database signal is determined.

The present invention is a signal search device for finding out a portion similar to a target query signal from a pre-registered database signal, and the signal search device is provided with a query feature extraction means (reference feature extraction means) for deriving a feature from the query signal; a database feature extraction unit (cumulative feature extraction unit) for setting a gazing window for a database signal and deriving a feature from a signal in the gazing window; a database feature extracting unit that extracts a feature series from the database feature sequence by repeating the process of the database feature extracting unit while shifting the focus window; a database feature thinning-out unit that extracts representative features from the classified feature series obtained by the database feature classifying unit and derives a representative feature series composed of a smaller number of features; a feature region extracting unit that derives a feature existence region included in the classification derived by the database feature classifying unit; a feature searching means for calculating a distance between the feature series derived by the query feature extracting means and the representative feature series derived by the database feature thinning means; a distance correcting unit for correcting the distance calculated by the feature matching unit using the area derived by the feature area extracting unit; and a signal detection determining unit for determining whether or not the query signal is present at the location of the database signal by comparing the corrected distance derived by the distance correcting unit with a search threshold value which is a threshold value corresponding to the distance, and a signal retrieval device for calculating distances to the query signal for several locations of the database signal by repeating the processing from the characteristic checking unit to the signal detection determining unit while shifting the observation window, and determining whether or not the query signal is present at the location of the database signal.

The signal search device of the present invention includes: a database feature compression unit (signal compression unit) that extracts segments as a partial series, extracts each segment obtained from the segment extraction unit, and determines a compressed image determination unit for calculating a feature image lower than the feature dimension, calculates features lower than the feature dimension corresponding to the segments obtained by the segment extraction unit, based on the image obtained by the compressed image determination unit, and a query feature compression unit (reference feature compression unit) that calculates features lower than the feature dimension corresponding to the features obtained by the query feature extraction unit, based on the image obtained by the compressed image determination unit, by repeating the segmentation while shifting the observation window; in the database feature thinning-out unit, the compressed feature series derived in the database feature compressing unit is used as a new feature series derived representative feature series; in the feature check means, the compressed feature derived by the query feature compression means is checked as a new feature, and further, the distance to the query signal is calculated for several locations of the database signal while repeating the processing from the feature check means to the signal detection and determination means while shifting the attention window, and whether or not the query signal is present at the location of the database signal is determined as a feature.

The signal search device of the present invention includes: a distance recalculating unit that calculates a distance between the feature series derived by the reference feature extracting unit and the feature series derived by the database feature extracting unit for the location of the database signal determined by the signal detection determining unit to be the presence of the query signal; and a signal detection re-determination unit for comparing the distance derived by the distance re-calculation unit with a search threshold and determining whether or not the query signal is present at the location of the data signal, wherein the processing by the feature check unit, the distance correction unit, the signal detection determination unit, the distance re-calculation unit, and the signal detection re-determination unit is repeated while shifting the gaze window, and the distance to the query signal is calculated for several locations of the database signal, and it is determined whether or not the query signal is present at the location of the database signal.

The present invention is the signal search device, which includes a database feature classification unit that classifies, in the database feature extraction unit, features derived by repeating the process of shifting a gaze window according to a predetermined distance, and determines a representative feature of the classification; a selection threshold setting section for calculating a selection threshold for a defined distance in the database feature classification section from a predetermined search threshold; and a database feature selecting unit that selects, as a feature, a database feature having a feature included in the representative feature classification that satisfies a condition derived from the calculated selection threshold in the selection threshold setting unit, with respect to a distance between the classification derived by the database feature classifying unit and the feature derived by the query feature extracting unit.

The present invention is characterized in that the signal search device includes a jump width calculation unit that calculates a jump width of the gaze window based on the distance calculated by the distance correction unit, and moves the gaze window only by the jump width, and repeats the processing by the feature check unit, the distance correction unit, the signal detection determination unit, and the jump width calculation unit while shifting the gaze window, and calculates the distance to the query signal for several locations of the database signal, and determines whether the query signal is present at the location of the database signal.

The present invention is a signal search device for finding out a portion classified as a target query signal from a preregistered database signal, including: a query feature extraction section that derives a feature from the query signal; a database feature extraction unit for setting a gazing window for a database signal and deriving features from the signal in the gazing window; a database feature classification unit that classifies, according to a predetermined distance, each feature derived by repeating the process of the database feature extraction unit while shifting a gaze window, and determines a representative feature of the classification; a selection threshold setting unit that calculates a selection threshold for a defined distance in the database feature classification unit from a predetermined search threshold; a database feature selection unit that selects a database feature having a feature included in the representative feature classification that satisfies a condition derived from the calculated selection threshold by the selection threshold setting unit, from among the classifications derived by the database feature classification unit, a distance from the feature derived by the query feature extraction unit; a segment extraction unit that extracts segments of a partial sequence by dividing a feature sequence derived by repeating the division while shifting a focus window in the database feature extraction unit; a compressed image determining unit configured to determine a feature image for calculating a feature image lower than the feature dimension for each segment obtained by the segment extracting unit; a database feature compression unit for calculating features lower than the feature dimension corresponding to the segments obtained by the segment extraction unit, based on the map obtained by the compressed map determination unit; a query feature compression unit for calculating a feature having a lower dimension than the feature dimension, which corresponds to the feature obtained by the query feature extraction unit, based on the map obtained by the compressed map determination unit; a feature check unit that calculates a distance between the compressed feature series derived by the database feature compression unit and the compressed feature derived by the query feature extraction unit; and a signal detection determining unit for determining whether or not the query signal is present at the location of the database signal by comparing the distance calculated by the feature checking unit with a search threshold value which is a threshold value corresponding to the distance.

The signal search device of the present invention includes: a distance recalculating unit that calculates a distance between the feature series derived by the query feature extracting unit and the feature series derived by the database feature extracting unit at the location of the database signal determined to be the query signal by the signal detection determining unit; and a signal detection re-determination unit for comparing the distance derived by the distance re-calculation unit with a search threshold and determining whether or not the query signal is present at the location of the database signal, wherein the processing by the characteristic-based checking unit, the signal detection determination unit, the distance re-calculation unit, and the signal detection re-determination unit is repeated while shifting the observation window, and the distance from the query signal is calculated for several locations of the database signal, thereby determining whether or not the query signal is present at the location of the database signal.

The present invention is characterized in that the signal search device calculates a jump width of the gaze window based on the distance calculated by the feature check means, moves the jump width calculation means of the gaze window only by the jump width, repeats the processing by the feature check means, the signal detection determination means, and the jump width calculation means while shifting the gaze window, calculates the distance to the query signal for several locations of the database signal, and determines whether or not the query signal is present at the location of the database signal.

The present invention is a program to be executed in a computer of a signal search device for finding out a portion similar to a target inquiry signal from a pre-registered database signal, the program causing the computer to execute: query feature extraction processing (reference feature extraction processing) for deriving a feature from a query signal; a database feature extraction process (cumulative feature extraction process) of setting a gazing window for a database signal and deriving a feature from a signal in the gazing window; database feature discrimination processing for discriminating the derived feature series by repeating the database feature extraction processing while skipping the gaze window; extracting representative features from the classified feature series obtained in the database feature classification processing, and deriving database feature thinning processing of a representative feature series composed of a smaller number of features; a feature region extraction process of deriving a feature existing region included in the distinction derived in the database feature distinction process; feature collation processing of calculating a distance between a feature derived in the query feature extraction processing and a representative feature series derived in the database feature thinning processing; distance correction processing for correcting the distance calculated in the feature check processing by using the region derived by the feature region extraction processing; and a program for performing a signal detection determination process of determining whether or not the query signal is present at the location of the database signal by comparing the corrected distance derived in the distance correction process with a search threshold value that is a threshold value corresponding to the distance, repeating the signal detection determination process from the feature checking process while shifting the observation window, calculating the distance to the query signal for several locations of the database signal, and determining whether or not the query signal is present at the location of the database signal.

In the present invention, a computer is caused to execute a segment extraction process for extracting segments as a partial series of segments, a compressed image determination process for determining each segment obtained from the segment extraction process, a database feature compression process (signal compression process) for determining a feature image having a lower dimension than the feature dimension corresponding to the segment obtained in the segment extraction process, and a query feature compression process (reference feature compression process) for calculating a feature having a lower dimension than the feature dimension corresponding to the feature obtained in the query feature extraction process, based on the image obtained in the compressed image determination process, by dividing a feature derived by repeating the database feature extraction process while shifting a gazing window, in addition to the processes included in the program; in the database feature thinning processing, the compressed feature series derived in the database feature compression processing is used as a new feature series derived representative feature series; and a program for performing a process of checking the compression feature derived in the query feature compression process as a new feature, repeating the process from the feature checking process to the signal detection determination process while shifting the observation window, calculating distances to the query signal for several locations of the database signal, and determining whether or not the query signal is present at the location of the database signal.

In addition to the processing included in the program, the present invention is a distance recalculation processing of calculating a distance between a feature derived in the query feature extraction processing and a feature series derived in the database feature extraction processing for the location of the database signal determined to be the query signal in the signal detection determination processing; and a program for comparing the distance derived in the distance recalculation processing with a search threshold value, and determining whether or not the query signal is present in the database signal at the position of the database signal by signal detection and re-determination processing, repeating the feature check processing, the distance correction processing, the signal detection and determination processing, the distance recalculation processing, and the signal detection and re-determination processing while shifting the window of interest, calculating the distance to the query signal at several positions of the database signal, and determining whether or not the query signal is present at the position of the database signal.

The present invention is a database feature classification process for causing a computer to execute, in addition to processes included in the program, a database feature classification process for classifying, based on a predetermined distance, each feature derived by repeating the database feature extraction process while shifting a gaze window, and determining a representative feature of the classification; calculating a selection threshold setting process of a selection threshold defining a distance in the database feature classification process from a predetermined search threshold; and a program for selecting a database feature selection process having a feature included in the representative feature classification that satisfies a condition derived from the calculated selection threshold in the selection threshold setting process, from among the classifications derived in the database feature classification process, a distance from the feature derived in the query feature extraction process.

The present invention is a method for calculating jump width of a gaze window based on a distance calculated in the distance correction processing and moving the gaze window only by the jump width, in addition to the processing included in the program; and a program for repeating the characteristic checking process, the distance correcting process, the signal detection judging process and the jump amplitude calculating process while shifting the watching window, calculating the distance from the inquiry signal for several positions of the database signal, and determining whether the inquiry signal exists at the position of the database signal.

The present invention is a computer-implemented program for a signal search device for extracting a portion classified as a target query signal from a pre-registered database signal, the program causing a computer to execute: query feature extraction processing for deriving a feature from the query signal; setting a gazing window for a database signal, and deriving a database characteristic extraction process of the characteristic from the signal in the gazing window; classifying, based on a predefined distance, each feature derived by repeating the database feature extraction process while shifting the gaze window, and determining a database feature classification process for classifying representative features of the feature; calculating a selection threshold setting process of a selection threshold defining a distance in the database feature classification process from a predetermined search threshold; a database feature selection process of selecting a database feature having a feature included in the representative feature classification that satisfies a condition derived from the calculated selection threshold in the selection threshold setting process, from among the classifications derived in the database feature classification process, a distance from the feature derived in the query feature extraction process; extracting a segment extraction process which is a segment of a partial series by dividing a feature series derived by repeating the database feature extraction process while shifting a gaze window; a compressed image determination process for determining a feature image for calculating a feature image lower than the feature dimension from each segment obtained by the segment extraction process; calculating a database feature compression process of features lower than the feature dimension corresponding to the segments obtained in the segment extraction process, based on the map obtained in the compressed map determination process; a query feature compression process of calculating a feature having a lower dimension than the feature dimension, which corresponds to the feature obtained by the query feature extraction process, from the map obtained by the compression map determination process; feature matching processing of calculating a distance between the series of compressed features derived in the database feature compression processing and the compressed features derived in the query feature extraction processing; and a processing unit configured to compare the distance calculated in the feature check processing with a search threshold value corresponding to the distance, determine whether or not the query signal is present at the location of the database signal, repeat the feature check processing and the signal detection determination processing while shifting the observation window, calculate distances to the query signal for several locations of the database signal, and determine whether or not the query signal is present at the location of the database signal.

In addition to the processing included in the program, the present invention is directed to a distance recalculation process of calculating a distance between a feature derived in the query feature extraction process and a feature series derived in the database feature extraction process for a location of a database signal determined to be a query signal present in the signal detection determination process; and a program for performing a signal detection re-determination process for determining whether or not the query signal is present at the location of the database signal by comparing the distance derived in the distance re-calculation process with a search threshold, and repeating the feature check process, the signal detection determination process, the distance re-calculation process, and the signal detection re-determination process while shifting the annotation window, thereby calculating distances to the query signal at several locations of the database signal and determining whether or not the query signal is present at the location of the database signal.

The present invention is a jump amplitude calculation process for causing a computer to execute, in addition to the processes included in the program, a jump amplitude calculation process for calculating a jump amplitude of a gaze window based on the distance calculated in the characteristic check process and moving the gaze window by the jump amplitude only; and a program for repeating the characteristic checking process, the signal detection judging process and the jump amplitude calculating process while shifting the watching window, calculating the distance between the inquiry signal and several positions of the database signal, and determining whether the inquiry signal exists at the position of the database signal.

The present invention is a computer-readable recording medium recording a program executed by a computer of a signal search device for finding a portion similar to a target inquiry signal from a pre-registered database signal, the program causing the computer to execute: query feature extraction processing for deriving a feature from the query signal; setting a gazing window for a database signal, and deriving a database characteristic extraction process of the characteristic from the signal in the gazing window; database feature extraction processing is repeated while the gaze window is staggered, and database feature classification processing for classifying the derived feature series is performed; extracting representative features from the characteristic series after the distinguishing obtained in the database feature distinguishing process, and deriving a database feature sparse process of the representative feature series composed of fewer features; a feature region extraction process of deriving a feature existing region included in the classification derived in the database feature region classification process; a feature collation process of calculating a distance between the feature derived in the query feature extraction process and the representative feature series derived in the database feature thinning process; distance correction processing for correcting the distance calculated in the feature check processing by using the region derived by the feature region extraction processing; and a computer-readable recording medium storing a program for performing a process of comparing the corrected distance derived in the distance correction process with a search threshold as a threshold corresponding to the distance, determining whether or not the query signal is present at the location of the database signal, repeating the process from the feature checking process to the signal detection determination process while shifting the observation window, calculating the distance to the query signal for several locations of the database signal, and determining whether or not the query signal is present at the location of the database signal.

The present invention is a computer program for recording a query feature compression process for causing a computer to execute a feature series derived by repeating the database feature extraction process while shifting a focus window, in addition to processes included in the program, extracting a segment extraction process as a partial series segment, each segment obtained from the segment extraction process, determining a compressed image determination process for calculating a feature image lower than the feature dimension, calculating a database feature compression process for calculating features lower than the feature dimension corresponding to the segment obtained in the segment extraction process based on the image obtained in the compressed image determination process, and calculating features lower than the feature dimension corresponding to the features obtained in the query feature extraction process based on the image obtained in the compressed image determination process; in the database feature thinning processing, the compressed feature series derived in the database feature compression processing is taken as a new feature series derived representative feature series; in the feature check processing, the compressed feature derived in the query feature compression processing is checked as a new feature, and further, the processing from the feature check processing to the signal detection determination processing is repeated while shifting the gaze window, and the distance to the query signal is calculated for several locations of the database signal, and whether or not the query signal is present at the location of the database signal is determined.

In addition, the present invention is a distance recalculation process for recording a distance between a feature derived in the query feature extraction process and a feature series derived in the database feature extraction process at a location of the database signal determined to be the presence of the query signal in the signal detection determination process, in addition to the processes included in the program; and a computer-readable recording medium storing a program for performing a process of comparing the distance derived in the distance recalculation process with a search threshold value, and determining whether or not the query signal is present at the location of the database signal, and repeating the feature check process, the distance correction process, the signal detection determination process, the distance recalculation process, and the signal detection re-determination process while leaving the window of interest, and calculating the distance to the query signal at a plurality of locations of the database signal, thereby determining whether or not the query signal is present at the location of the database signal.

The present invention is a database feature classification method for classifying, based on a predetermined distance, each feature derived by repeating the database feature extraction process while shifting a viewing window, in addition to the processes included in the program, and determining a representative feature of the classification; a selection threshold setting process of calculating a selection threshold for a distance defined in the database feature classification process from a predetermined search threshold; and a computer-readable recording medium for recording a program for database feature selection processing for selecting a database feature having a feature included in a classification of the representative feature that satisfies a condition derived from the calculated selection threshold in the selection threshold setting processing, from among the classifications derived in the database feature classification processing, a distance from the feature derived in the query feature extraction processing.

The present invention is a computer program for causing a computer to execute a jump width calculation process of calculating a jump width of a gaze window based on a distance calculated in the distance correction process, and moving the gaze window by only the jump width, in addition to the processes included in the program; and a computer-readable recording medium storing a program for repeating the feature-based checking process, the distance correcting process, the signal detection judging process, and the jump width calculating process while shifting the viewing window, calculating distances from the query signal at several locations of the database signal, and determining whether the query signal is present at the location of the database signal.

The present invention is a computer-readable recording medium recording a program executed by a computer of a signal search device for finding out a portion similar to a target inquiry signal from a pre-registered database signal, the program causing the computer to execute: query feature extraction processing for deriving a feature from the query signal; setting a gazing window for a database signal, and deriving a database characteristic extraction process of the characteristic from the signal in the gazing window; classifying, based on a predefined distance, the features derived from the database feature extraction process while shifting the gaze window, and determining a database feature classification process for classifying representative features of the features; a selection threshold setting process of calculating a selection threshold for a distance defined in the database feature classification process from a predetermined search threshold; a database feature selection process of selecting a database feature having a feature included in the representative feature classification that satisfies a condition derived from the calculated selection threshold in the selection threshold setting process, from among the classifications derived in the database feature classification calculation process, a distance from the feature derived in the query feature extraction process; a segmentation extraction process of extracting segments as a partial series by segmenting features derived by repeating the database feature extraction process while shifting the gaze window; determining a compressed image determination process for calculating a feature image lower than the feature dimension from each segment obtained in the segment extraction process; calculating a database feature compression process of features having a dimension lower than the feature dimension, which corresponds to the segment obtained in the segment extraction process, based on the map obtained in the compressed map determination process; a query feature compression process of calculating a feature having a dimension lower than the feature dimension, which corresponds to the feature obtained by the query feature extraction process, from the map obtained by the compression map determination process; feature check processing of calculating a distance between the compressed feature series derived in the database feature compression processing and the compressed feature derived in the query feature extraction processing; and a computer-readable recording medium storing a program for performing a process of comparing the distance calculated in the feature check process with a search threshold value which is a threshold value corresponding to the distance, determining whether or not the query signal is present at the location of the database signal, repeating the feature check process and the signal detection determination process while shifting the observation window, calculating the distance to the query signal for several locations of the database signal, and determining whether or not the query signal is present at the location of the database signal.

A distance recalculation process of calculating a distance between a feature derived in the query feature extraction process and a feature series derived in the database feature extraction process for the location of the database signal determined to be the presence of the query signal in the signal detection determination process, in addition to the processes included in the program; and a computer-readable recording medium storing a program for performing a process of comparing the distance derived in the distance recalculation process with a search threshold value, and determining whether or not the query signal is present at the location of the database signal, and repeating the feature check process, the signal detection determination process, the distance recalculation process, and the signal detection re-determination process while shifting the window of view, and determining whether or not the query signal is present at the location of the database signal by calculating the distance to the query signal at a plurality of locations of the database signal.

The present invention is a computer-readable recording medium recording a program for causing a computer to execute processing for calculating a jump width of a gaze window based on a distance calculated in the feature check processing in addition to processing included in the program and moving the gaze window only by the jump width, and repeating the feature check processing, the signal detection determination processing, and the jump width calculation processing while shifting the gaze window, and for calculating a distance to a query signal for several locations of a database signal, and determining whether the query signal is present at the location of the database signal.

Drawings

Fig. 1 is a block diagram showing the structure of embodiment 1 of the present invention.

Fig. 2 is a block diagram showing the structure of embodiment 2 of the present invention.

Fig. 3 is a block diagram showing the structure of embodiment 3 of the present invention.

Fig. 4 is a block diagram showing the structure of embodiment 4 of the present invention.

Fig. 5 is a flowchart showing the processing operation of the signal compression apparatus according to embodiment 1.

Fig. 6 is a flowchart showing the operation of the initial partial signal composing unit 1 shown in fig. 1.

Fig. 7 is a flowchart showing an operation of the partial signal configuration selecting unit 2 shown in fig. 1.

Fig. 8 is an overall flowchart showing the average dimension calculation process shown in fig. 7.

Fig. 9 is a flowchart showing a basic extraction process used in the average dimension calculation process shown in fig. 8.

Fig. 10 is a flowchart showing the operation of the compressed image determination unit 4 shown in fig. 1.

Fig. 11 is a flowchart showing the operation of the partial signal reconstructing unit 3 shown in fig. 1.

Fig. 12 is a flowchart showing the overall operation of the signal compression unit 5 shown in fig. 1.

Fig. 13 is a flowchart showing a procedure of the signal mapping process shown in fig. 12.

Fig. 14 is a flowchart showing a procedure of the projection distance calculating process shown in fig. 12.

Fig. 15 is a program flowchart showing the compression feature configuration process shown in fig. 12.

Fig. 16 is a flowchart showing the processing operation of the signal search device according to embodiments 2, 3 and 4.

Fig. 17 is a flowchart showing the operation of the reference feature extracting unit 6 shown in fig. 2.

Fig. 18 is a flowchart showing the operation of the cumulative feature extracting unit 7 shown in fig. 2.

Fig. 19 is a flowchart showing the operation of the reference feature compressing unit 8 shown in fig. 2.

Fig. 20 is a flowchart showing a procedure of the reference signal mapping process shown in fig. 19.

Fig. 21 is a flowchart showing a procedure of the reference projection distance calculating process shown in fig. 19.

Fig. 22 is a flowchart showing a procedure of the reference compression characteristic configuration process shown in fig. 19.

Fig. 23 is a flowchart showing the operation of the feature collating unit 9 shown in fig. 2.

Fig. 24 is a flowchart showing the operation of the signal detection determining unit 10 shown in fig. 2.

Fig. 25 is a flowchart showing the operation of the jump width calculating unit 11 shown in fig. 3.

Fig. 26 is a flowchart showing the operation of the distance recalculating unit 12 shown in fig. 4.

Fig. 27 is a flowchart showing the operation of the signal detection re-determination unit 13 shown in fig. 4.

FIG. 28 is a figure 1 for explaining the experimental results.

FIG. 29 is a 2 nd explanatory view showing the experimental results.

Fig. 30 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 5 is applied.

Fig. 31 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 6 is applied.

Fig. 32 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 7 is applied.

Fig. 33 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 8 is applied.

Fig. 34 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 9 is applied.

Fig. 35 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 10 is applied.

Fig. 36 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 11 is applied.

Fig. 37 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 12 is applied.

Fig. 38 is a flowchart showing the processing of the database feature distinguishing unit.

Fig. 39 is a flowchart showing a process of the database inter-feature thinning section.

Fig. 40 is a flowchart showing the processing of the feature region extracting unit.

Fig. 41 is a flow chart 2 showing the processing of the feature collating part.

Fig. 42 is a flowchart showing the processing of the distance correction unit.

Fig. 43 is a flowchart showing the processing of the segment extracting unit.

Fig. 44 is a flowchart showing the processing of the database feature classification unit.

Fig. 45 is a flowchart showing a process of selecting the threshold setting unit.

Fig. 46 is a flowchart showing the processing of the database feature selecting unit.

Fig. 47 is a view showing a state where a histogram space is cut.

FIG. 48 is a 3 rd explanatory view showing an experimental result.

FIG. 49 is a 4 th explanatory view showing an experimental result.

FIG. 50 is a 5 th explanatory view showing the experimental results.

FIG. 51 is a 6 th explanatory view showing the experimental results.

Fig. 52 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 13 is applied.

Fig. 53 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 14 is applied.

Fig. 54 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 15 is applied.

Fig. 55 is a flow chart showing the processing of the feature collating part 3.

FIG. 56 is a 7 th explanatory view showing an experimental result.

FIG. 57 is an explanatory view of FIG. 8 showing the experimental results.

Detailed Description

Hereinafter, a signal compression apparatus according to an embodiment of the present invention will be described with reference to the drawings. In the present invention, various processing target signals can be used, but here, as an example of the processing target signal (original signal), it is assumed that a histogram series which is a kind of multidimensional vector system is created from a video signal and used. The histogram is a classification frequency distribution table obtained by extracting features from a video signal and classifying the extracted features by a predetermined method. Further, as long as the features of the original signal can be extracted as a multidimensional vector series, the original signal is not limited to the histogram, and the multidimensional vector series may be used.

[ 1 st embodiment ]

Fig. 1 is a block diagram showing the structure of embodiment 1. In fig. 1, reference numeral 1 denotes an initial partial signal forming section which forms a partial signal having a shorter length than the original signal from the original signal. Reference numeral 2 denotes a partial signal configuration selecting unit which squeezes partial signals having a smaller data amount than the original signal into partial signal configuration candidates for the partial signals output from the initial partial signal configuring unit 1. Reference numeral 3 denotes a partial signal reconstruction unit 3 which uses the partial signal configuration candidates output from the partial signal configuration selection unit 2 to determine the partial signal configuration to be actually used. Reference numeral 4 denotes a compressed image determination unit which determines an image for calculating a compressed signal from each partial signal output from the partial signal reconstruction unit 3. Reference numeral 5 denotes a signal compression unit which calculates compressed signals corresponding to the partial signals output from the partial signal reconstruction unit 3, based on the video output from the compressed video determination unit 4.

Further, the signal compression unit 5 includes: a signal mapping unit 51 for projecting each histogram in the partial signal to a partial space formed by the partial signal by using the partial signal outputted from the partial signal reconstructing unit 3 and the linear image set outputted from the compressed image determining unit 4; a projection distance calculating unit 5 for calculating the distance between each histogram and the corresponding compressed histogram by using the partial signal outputted from the partial signal reconstructing unit 3, the linear image set outputted from the compressed image determining unit 4, and the compressed histogram series set outputted from the signal mapping unit 51; and a compression feature configuration unit 53 for calculating a compression feature series by using the set of the compression histogram series output from the signal mapping unit 51 and the projection distance output from the projection distance calculation unit 5.

The signal compression apparatus shown in fig. 1 receives a histogram series extracted from an original signal, i.e., a video signal to be compressed, and outputs a compressed histogram series obtained by compressing a compressed signal, i.e., the histogram series extracted from the video signal.

Next, the operation of the signal compression device shown in fig. 1 will be described with reference to fig. 5 to 15. First, the overall operation flow of the signal compression apparatus will be described with reference to a flowchart showing the processing operation of the signal compression apparatus of the present embodiment shown in fig. 5. The detailed operation of each process will be described later.

In fig. 5, first, the initial partial signal composing section 1 reads the supplied original signal (step S1). Then, the initial partial signal composing unit 1 performs an initial partial signal composing process (step S2). Next, the partial signal configuration selecting unit 2 performs a partial signal configuration selecting process (step S3). Next, the partial signal reconstruction unit 3 performs a partial signal reconstruction process (step S4). Next, the compressed image determination unit 4 performs a compressed image determination process (step S5). Further, the signal compression unit 5 performs signal compression processing (step S6). Then, the signal compression unit 5 outputs a compressed signal of the original signal (step S7).

Next, referring to fig. 6, the initial partial signal configuration processing (step S2) shown in fig. 5 will be described in detail. Fig. 6 is a flowchart showing the operation of the initial partial signal composing unit 1.

First, the initial partial signal composing unit 1 reads a histogram series as an original signal (step S11). Then, the initial partial signal composing section 1 equally divides the read histogram series by a predetermined number of segments (step S12). Then, the initial partial signal composing unit 1 outputs a set of segments as a histogram series after division (step S13).

Next, referring to fig. 7, the partial signal configuration selection process (step S3) shown in fig. 5 will be described in detail. Fig. 7 is a flowchart showing the operation of the partial signal configuration selecting unit 2.

In fig. 7, first, the partial signal configuration selecting unit 2 reads a set of fixed-length segments (partial signals) output from the initial partial signal configuring unit 1 (step S15). Next, a split boundary movable range which is a range in which the split boundary can be moved from the present position is given in advance, and as for each split boundary, a partial range of the split boundary movable range is set as a split boundary movable range from before and after the present split boundary position (step S16). Next, the first division boundary among the division boundaries of the segments is focused on (step S17). Then, for 2 segments sharing the segment boundary, the number of compressed signal dimensions at the current position of the segment boundary is calculated, and an average value normalized by the segment length is calculated (step S18).

Next, for 2 segments sharing the division boundary, the compressed signal dimensions when the division boundary is at both ends of the movable range of the division boundary are calculated, and an average value normalized by the segment length is calculated (step S19).

Further, the average value of the dimensions is calculated similarly at several points in the range in which the division boundary may move. The number of calculations is determined from the average of the dimensions of the 3 point-division boundaries. Then, based on the obtained number of calculations, the average value of the dimensions is calculated from the top program of the split boundary movable range so as to be at equal intervals in the split boundary movable range (steps S20, S21, S22).

Then, it is determined whether or not there is a dimensional change in the histogram series of the common division boundary (step S23), and when there is a dimensional change in the histogram series of the common division boundary (Y ES in step S23), the range position from the current calculation point to one preceding calculation point is stored as an optimum division boundary candidate and the range position is completely stored (step S24). In addition, a method of calculating the dimension average value will be described in detail later with reference to the drawings.

The number x of places to be calculated on one side of the range in which the divided boundary is movable is obtained as follows.

First, in the processing of the partial signal configuration selecting unit 2 and the partial signal reconstructing unit 3 described later, the number of sites f (x) for calculating the required dimension average value to determine one division boundary is given by the following equation.

Is given by the following formula.

Where Δ is the partition boundary movable amplitude, C _LL 、C _LC 、C _LR A segment dimension near the head, C, representing the division boundary of the gaze at the front, initial position, and rear of the range of possible movement of the division boundary _{R L} 、C _RC 、C _RR The division boundaries indicating the gaze are located at the front end, the initial position, and the rear end of the range of possible movement of the division boundaries, respectively.

The f (x) is when

Then, take the minimum value. The number of points to be calculated is set to the nearest integer to x.

Since the number of times of calculation is obtained in this way, if the calculation is not completed for all the calculation points (NO in step S25), the boundary is moved to the next calculation point (step S26), and the operations from step S22 to step S24 are repeated.

Then, if the calculation is completed for all the calculation points (YES in step S25), it is determined whether or not the operation is completed for all the division boundaries (step S27), and if the operation is not completed for all the division boundaries (NO in step S27), the division boundary is changed to the next division boundary (step S2), and the operations from step S18 to step S26 are repeated. Then, at the time point when the operation for all the divided boundaries is finished (YES at step S27), the partial signal configuration selecting unit 2 outputs the kept divided boundary candidate set (step S29).

Next, a method of calculating the dimension average value will be described with reference to fig. 8 and 9. Fig. 8 is an overall flowchart showing the average dimension calculation process, and fig. 9 is a flowchart showing a basic extraction process for the average dimension calculation process. Specifically, the dimension average is calculated as follows.

In fig. 8, first, 2 segments sharing a division boundary are read (step S31). Next, from the given 2 segments, a partial spatial basis is extracted that well represents the original signal properties (step S32).

In the case of explaining the basis extraction method in step S32 with reference to fig. 9, KL (Karhunen-Loeve) expansion is first performed on each given segment (step S36) in fig. 9 (step S3). Specifically, KL expansion was performed according to the following procedure. Initially, the mean histogram and the co-dispersed rows of the intra-segment histogram are computed. For segment X of j ^(j) ＝[x ₁ ^(j) ，x ₂ ^(j) ，…x _Lj ^(j) ]Co-dispersed matrix S of (j =1,2, \8230M) ^(j) The calculation is as follows.

Where M is the number of segments, lj is the segment length # j,

represents X ^(j) Average histogram of (1) (. C) ^T Representing transposed rows and columns.

Next, the co-dispersed matrix S is obtained ^(j) Eigenvalues and eigenvectors of (j =1,2, … M). The above is a KL expansion procedure.

The value obtained by dividing the eigenvalue corresponding to each eigenvector obtained by the KL expansion by the total value of the eigenvalues of all eigenvectors is referred to as the operation efficiency of the eigenvector. Then, the eigenvectors are rearranged in accordance with the magnitude of the action efficiency, the total value of the action efficiencies is selected in accordance with the procedure before exceeding a predetermined action threshold (steps S38, S39), and the selected eigenvector is used as the basis of the partial space (step S40) to obtain a basis set (step S41).

In step S33 of FIG. 8, the number M of basic pieces extracted from each segment is calculated _j (j =1,2, \ 8230;, M) becomes the dimension of the compressed signal, and thus the average value N 'is normalized by the segment length and calculated as follows' _j (j =1,2, \ 8230;, M-1) (step S33).

Further, the frame average of the basis number is output as the average value N of the dimension number _j ' (step S34).

Next, referring to fig. 10, the partial signal conversion processing shown in fig. 5 (step S4) will be described in detail. Fig. 10 is a flowchart showing the operation of the partial signal reconstruction unit 3.

In fig. 10, first, the partial signal reconstructing unit 3 reads the set of fixed-length segments output from the initial partial signal constructing unit 1 and the set of segment boundary candidates output from the partial signal constructing and selecting unit 2 (step S42). Then, attention is paid to the first one of the divided boundary boundaries (step S43) to move the divided boundary to the head position among the divided boundary candidates of the segment (step S44). Next, for 2 segments sharing the segment boundary, the compressed signal dimension at the current position of the segment boundary is calculated, and an average value normalized by the segment length is calculated (step S45). The average value of the dimensions is calculated in accordance with the programs shown in fig. 8 and 9, as in the partial signal configuration selecting unit 2.

Next, if the minimum value among the average values of the dimensions of the focused division boundaries is calculated (YES in step S46), the average value and the current division boundary position are held (steps S4 and S48). Then, if the calculation is not completed for all candidates (NO in step S49), the division boundary is moved to the next candidate point (step S50), the process returns to step S45, and the operation up to this point of the calculation of the dimension average is repeated. When the candidate point is not present, the segment boundary to be observed is moved to the segment boundary position corresponding to the minimum value of the dimensional average value, and the segment boundary is determined.

If the calculation of the last divided boundary is not completed (NO in step S51), the observed divided boundary is changed to the next divided boundary (step S52), the process returns to step S44, and the operations up to this point are repeated. At a time point after the end of the operation on all the division boundaries (YES in step S51), the partial signal reconstructing unit 3 outputs a variable-length segment set specified by moving the division boundaries (step S53).

Next, the compressed image determination process (step S5) shown in fig. 5 will be described in detail with reference to fig. 11. Fig. 11 is a flowchart showing the operation of the compressed image determination unit 4.

In fig. 11, first, the compressed map determination unit 4 reads the segment set output from the partial signal reconstruction unit 3 (step S55). Next, the basis of each segment is extracted (step S56). The basic extraction is calculated in accordance with the program shown in fig. 9, as in the partial signal configuration selecting unit 2.

Next, the projection onto the partial space is made into a map of the segment (step S58). Then, the compressed image determination unit 4 outputs an image corresponding to each segment (step S59).

Next, the signal compression process shown in fig. 5 (step S6) will be described in detail with reference to fig. 12 to 15. Fig. 12 is a flowchart showing the overall operation of the signal compression unit 5.

In fig. 12, first, the signal compression unit 5 reads the segment set output from the partial signal reconstruction unit 3 and the linear image set output from the compressed image determination unit 4 (step S60). Next, a signal mapping process is performed to project each histogram in the partial signal onto a partial space created from the partial signal, using a predetermined set of the partial signal and the line map (step S61).

Then, projection distance calculation processing is performed using a set of the given partial signal and linear image and a set of compressed histogram series obtained by the signal image processing, and the distance between each histogram and the compressed histogram corresponding thereto is calculated (step S62). Further, a compression feature configuration process of calculating a compression feature series is performed using the set of compressed histogram series obtained by the signal mapping process and the projection distance obtained by the projection distance calculation process (step S63). Then, the signal compression unit 5 outputs the compression feature series obtained by the compression feature configuration processing (step S64).

Fig. 13 is a flowchart showing a procedure of the signal mapping process (step S61) shown in fig. 12.

In fig. 13, first, the signal mapping unit 51 constituting the signal compression unit 5 reads the segment output from the partial signal reconstruction unit 3 and the linear image set output from the compressed image determination unit 4 (step S66). Next, each histogram in the segment is projected onto a partial space created by the segment (step S67).

Specifically, when setting is made by segment X ^(j) The basic set of the obtained partial space is A ^(j) ＝[a ₁ ^(j) ，a ₂ ^(j) ，…a _Nj ^(j) ] ^T (j =1,2, \8230;, M), the histogram series Y is compressed ^(j) ＝[y ₁ ^(j) ，y ₂ ^(j) ，… y _Lj ^(j) ](j =1,2, \8230;, M) is calculated as follows.

Wherein N is _j Is composed of X ^(j) The basis number of the obtained partial space is,

is to vector the column

In rows and columns arranged in Lj strips, i.e.(j＝1，2，…，M)。

Therefore, the signal mapping section 51 outputs the set Y of the compressed histogram series ⁽¹⁾ 、Y ⁽²⁾ 、…、Y ^(M)(step S68).

Fig. 14 is a flowchart showing a procedure of the projection distance calculating process (step S62) shown in fig. 12.

In fig. 14, first, the projection distance calculating unit 52 constituting the signal compressing unit 5 reads the segment output by the partial signal reconstructing unit 3, the set of linear images output by the compressed image determining unit 4, and the set of compressed histogram series output by the signal mapping unit 51 (step S70). Next, the compressed histogram position in the original histogram existence space is obtained by back-projecting the compressed histogram (step S71), and the distance between each histogram and the corresponding compressed histogram is calculated (step S72).

In particular to a method for making a composite material,

wherein the content of the first and second substances,

each one of

Representing a compressed histogram y _i ^(j) The position in space of the original histogram of (a).

And, x _i ^(j) And

defined as the distance of the histogram from the compressed histogram, also called histogram x _i ^(j) The projected distance of (a). That is, the projection distance of x is defined as follows using the euclidean distance.

Where n is the dimension of the histogram, x = (x) ₁ ，x ₂ ，…，x _n )，

Therefore, the projection distance calculation unit 52 outputs the projection distance of each histogram of the corresponding element (step S73).

Fig. 15 is a flowchart showing a procedure of the compression characteristic configuration process (step S63) shown in fig. 12.

In fig. 15, first, the compression characteristic configuration unit 53 constituting the signal compression unit 5 reads the set of compressed histogram series output from the signal mapping unit 51 and the projection distance output from the projection distance calculation unit 52 (step S75). Next, from the compressed histogram y = (y) ₁ ，y ₂ ，…，y _k ) Projection distance calculated corresponding thereto

Forming a compression characteristic y ^＊ As follows (step s 76).

Where N is the dimension of the compressed histogram y.

Therefore, the compression characteristic configuration section 53 outputs the compression characteristic series (step S77).

[ example 2 ]

Fig. 2 is a block diagram showing the structure of embodiment 2. In embodiment 2, a signal search device to which the signal compression device described in embodiment 1 is applied will be described. In fig. 2, the components denoted by the same reference numerals as those of the signal compression device described in embodiment 1 using fig. 1 are components that perform the same operations as those of the signal compression device, and therefore, the description thereof is omitted here.

In fig. 2, reference numeral 6 denotes a reference feature extracting unit which calculates a reference feature series from a reference signal as a target signal. Reference numeral 7 denotes an accumulated feature extraction unit which sets a comment window for an accumulated signal as an original signal registered in advance and calculates a feature series from the signal in the comment window. Reference numeral 8 denotes a reference feature compressing unit which compresses the reference feature series output from the reference feature extracting unit 6 on the basis of the map output from the compressed map determining unit 4. Reference numeral 9 denotes a feature check unit which calculates the distance from the accumulated compressed signal output by the signal compression unit 5 by reusing the reference compressed signal output by the reference feature compression unit 8 and the feature series output by the accumulated feature extraction unit 7. Reference numeral 10 denotes a signal detection determination unit which determines whether or not the reference signal is present at the point of the accumulated signal by comparing the distance output from the feature checker 9 with a search threshold value which is the same value as the corresponding distance.

The signal search device shown in fig. 2 receives a video signal to be searched for as a sample, which is a reference signal, and a video signal to be searched for as an accumulated signal, and outputs a predetermined value of the distance from the reference signal (referred to as a search threshold) at θ ₁ The following accumulate locations in the signal.

Next, the operation of the signal search device shown in fig. 2 will be described with reference to fig. 16 to 24. First, the overall operation flow of the signal search device will be described with reference to a flowchart showing the processing operation of the signal search device according to the present embodiment shown in fig. 16. The detailed operation of each process will be described later.

In fig. 16, first, the reference feature extracting unit 6 performs a reference feature extracting process (step S81). Next, the cumulative feature extracting unit 7 performs a cumulative feature extracting process (step S82). Next, the initial partial signal composing unit 1 performs initial partial signal composing processing (step S83). Next, the partial signal configuration selecting unit 2 performs a partial signal configuration selecting process (step S84). The partial signal reconstruction unit 3 performs a partial signal reconstruction process (step S85). Next, the compressed image determination unit 4 performs a compressed image determination process (step S85). Next, the signal compression section 5 performs signal compression processing (step S87). Further, the reference feature compressing unit 8 performs a reference feature compressing process (step S88).

Then, a comment window for setting the cumulative compression characteristic series output to the signal compression unit 5 is set at the head of the cumulative signal (step S89). Next, the feature check unit 9 performs a feature check process (step S9). Then, the signal detection determination unit 10 performs a signal detection determination process (step S91). When the signal detection determination process is performed, it is determined whether the current position of the gaze window is the accumulated signal end point (step S95), and when the current position of the gaze window is not the accumulated signal end point (NO in step S95), the gaze window is shifted and the process returns to step S90 to repeat the above process. When the current position of the window is the cumulative signal end point (YES in step S95), a signal search result is output (step S96).

The jump width calculation processing in step S92 shown in fig. 16 is not required in embodiment 2, and will be described in embodiment 3. Similarly, the distance recalculation processing at step S93 and the signal redetermination processing at step S94 shown in fig. 16 are not required in embodiment 2 and will be described in embodiment 4 described later.

The initial partial signal configuration process (step S83), the partial signal configuration selection process (step S84), the partial signal reconstruction process (step S5), the compressed map determination process (step S86), and the signal compression process (step S87) are similar to those performed by the signal compression apparatus according to embodiment 1 shown in fig. 6 to 15, and therefore, the description thereof is omitted here. However, the initial partial signal configuration processing (step S83) uses the cumulative feature series output from the cumulative feature extraction processing (step S82) as an input.

Next, the reference feature extraction process (step S81) shown in fig. 16 will be described in detail with reference to fig. 17. Fig. 17 is a flowchart showing the operation of the reference feature extracting unit 6.

In fig. 17, first, the reference feature extracting unit 6 reads the given reference signal (step S98). Next, the read reference signal is subjected to feature extraction (step S99).

Here, the target signal uses the spectral feature as a feature extracted in the case of an acoustic signal. The spectral feature extraction may be performed on the acoustic signal by means of a band-pass filter. For example, when an acoustic signal of about 15 seconds is searched for from acoustic signals reproduced by a television, a radio, or the like, good results are obtained by performing specific setting of the feature extraction as follows. That is, the center frequencies of these filters are set at equal intervals on the logarithmic axis by using 7 bandpass filters, the output square average value of each bandpass filter in the analysis window is calculated while moving the analysis window for a time length of about 60 milliseconds every 10 milliseconds, and the obtained 7 values are grouped as a 7-dimensional feature vector. A feature vector is then obtained every 10 milliseconds.

On the other hand, for the video signal, a color feature is used as the feature. For example, when an image signal of about 15 seconds is searched for from a broadcast image signal of a television or the like, good results are obtained by performing specific setting of feature extraction as follows. That is, each image constituting the map is divided into 2 in the vertical direction and 3 in the horizontal direction, RGB values are calculated for each division, and 3 values of RGB obtained in each division are combined into 18 values in total to form a set, which is used as an 18-dimensional feature vector. Where the map is constructed from 30 frames of images every 1 second, a feature vector is obtained every 30 minutes of a second.

Next, the feature vector is encoded by vector quantization (step S100), and a histogram of the feature vector is created from the time series of the feature vector (step S101). For example, when the number of code words for vector quantization is 512, the number of bins (bin: bin) in the entire histogram becomes 512, and each feature vector is classified into any of the 512 bins. In the following description, a histogram created from a reference signal is referred to as a reference histogram. Then, the reference feature extraction unit 6 outputs the obtained reference histogram (step S102).

Next, the cumulative feature extraction process shown in fig. 16 (step S82) will be described in detail with reference to fig. 18. Fig. 18 is a flowchart showing the operation of the cumulative feature extracting unit 7.

In fig. 18, first, the cumulative feature extracting unit 7 reads the cumulative signal (step S104). Next, a note window is set at the head of the read accumulation signal (step S105). Here, a note window having the same length as the reference signal supplied to the reference feature extracting unit 6 is set.

Next, the feature extraction is performed on the accumulated signal in the annotation window (step S106). Note that the feature extraction is performed in the same manner as the processing performed by the reference feature extraction unit 6. Further, a histogram of the feature vectors is created in accordance with the time series of the feature vectors in the annotation window (steps S107 and S108). The histogram is created by the same method as that performed in the reference feature extraction unit 6. Then, the accumulated feature extraction unit 7 repeatedly executes the processing from step S106 to step S108 until the end of the accumulated signal by shifting the attention window set at the head of the accumulated signal for each feature vector in the program at the start of the processing (step S109 and step S110). In the following description, each histogram created from the accumulated signal is referred to as an accumulated histogram. Finally, the cumulative feature extraction unit 7 outputs the obtained cumulative histogram series (step S111).

Next, the reference feature compression process (step S88) shown in fig. 16 will be described in detail with reference to fig. 19 to 22. Fig. 19 is a flowchart showing the operation of the reference feature compressing unit 8.

In fig. 19, first, the reference feature compressing unit 8 reads the reference histogram output by the reference feature extracting unit 6 and the set of linear images output by the compressed image determining unit 4 (step S113). Next, a reference signal image process of projecting a reference histogram onto the corresponding partial space is performed using each linear image (step S1). The projection is performed in the same manner as the signal compression unit 5 described in embodiment 1. For example, assuming that the number of segments is M =1000, 1000 compressed histograms are made.

Next, a reference projection distance calculation process is performed to calculate the distance between the histogram and each compressed histogram, that is, the projection distance of the histogram (step S115). This calculation is performed in accordance with the same processing as the signal compression section 5. Finally, a reference compressed feature configuration process for configuring the compressed features is performed based on the compressed histogram and the projection distance corresponding thereto (step S116). The compression characteristic is implemented in the same manner as the signal compression unit 5. The reference feature compressing unit 8 outputs a set of reference compression features (step S1).

Fig. 20 is a flowchart showing a procedure of the reference signal mapping process (step S114) shown in fig. 19.

In fig. 20, first, the reference signal mapping unit (not shown) constituting the reference feature compressing unit 8 reads the reference histogram output by the reference feature extracting unit 6 and the linear map set output by the compressed map determining unit 4 (step S119). Next, the reference histogram is projected to a partial space according to the line map corresponding to each segment (step S120).

Therefore, the reference signal mapping unit outputs a set of reference compressed histograms (step S121).

Fig. 21 is a flowchart showing a procedure of the reference projection distance calculating process (step S115) shown in fig. 19.

In fig. 21, first, the reference projection distance calculating unit (not shown) constituting the reference feature compressing unit 8 reads the reference histogram output by the reference feature extracting unit 6, the linear map set output by the compressed map determining unit 4, and the reference compressed histogram set output by the reference signal mapping unit (step S123). Next, the distance between each histogram and the compressed histogram corresponding to the histogram is calculated by back-projecting each compressed histogram (step S124) (step S125).

Therefore, the reference projection distance calculation unit outputs the projection distance corresponding to each histogram (step S6).

Fig. 22 is a flowchart showing a procedure of the reference compression characteristic configuration process (step S116) shown in fig. 19.

In fig. 22, first, the reference compression feature configuration unit (not shown) configuring the reference feature compression unit 8 reads the reference compression histogram set output from the reference signal mapping unit and the projection distance output from the reference projection distance calculation unit (step S128). Next, a compression feature is constructed from the compressed histogram and the projection distance calculated in association with the compressed histogram (step S129).

Therefore, the set of compression features is output with reference to the compression feature configuration section (step S130).

Next, the feature verification process (step S90) shown in fig. 16 will be described in detail with reference to fig. 23. Fig. 23 is a flowchart showing the operation of the feature collating unit 9.

In fig. 23, first, the feature check unit 9 reads the series of accumulated compressed features output from the signal compression unit 5 and the reference compressed feature set output from the reference feature compression unit 8 (step S132). Second, a reference compression characteristic y is calculated ^＊ _R And cumulative compression feature y ^＊ _S Distance (step S133).

Specifically, the distance d (y) ^＊ _R ，y ^＊ _S ) Euclidean distance is used and defined as follows.

Wherein x is _R Is with reference to a histogram, x _S Is a cumulative histogram, y _R And y _S Is corresponding to x _R And x _S The compressed histogram of (a) is calculated,

and

is y _R And y _S Position in histogram space of (a), y ^＊ _Ri And y ^＊ _Si Are each y ^＊ _R And y ^＊ _S The value of the ith dimension of (1).

Here, the following equation holds according to the property of KL expansion (principal component analysis).

And further, d (y) ^＊ _R ，y ^＊ _S ) Has the following properties.

However, the minimum value in equation (4) is given to y _R ，y _S ，

Andall histogram sets of time (x) _R ，x _S ) To take on the value.

In addition, in the principal component analysis, the distance value between the compressed features has a peculiar effect of being the lower limit value of the distance between the histograms according to the property expression (3) in the expression (4). Furthermore, by further using the projection distance, a larger lower limit value d (y) of the inter-histogram distance can be obtained as compared with the case where it is not used ^＊ _R ，y ^＊ _S )。

Therefore, the feature checking section 9 outputs the lower limit value of the obtained distance (step S134).

Next, the signal detection determination process (step S91) shown in fig. 16 will be described in detail with reference to fig. 24. Fig. 24 is a flowchart showing the operation of the signal detection determining unit 10.

In fig. 24, first, the signal detection determination unit 10 reads the distance lower limit value output from the characteristic checking unit 9 (step S139). Next, the distance lower limit value is compared with a search threshold value that is a euclidean distance predetermined value according to the distance scale (step S140). When the distance value is equal to or less than the search threshold (when the annotation window is divided in the time direction, it is found that the distance value is equal to or less than the search threshold for all time divisions) (YES in step S140), it is determined that the reference signal is present at the point of the accumulated signal (step S141), and the current position (flag) of the accumulated signal in the time series is output as a signal detection result (step S142).

[ example 3 ]

Fig. 3 is a block diagram showing the structure of embodiment 3. In the description of embodiment 3, the signal search device described in embodiment 2 is a signal search device in which the attention window jump width set for the cumulative compression feature series output from the signal compression unit 5 is calculated based on the distance output from the feature check unit 9, and the jump width calculation unit 11 for moving the attention window only by the jump width is newly provided. In fig. 3, the same reference numerals are given to the same components as those of the signal compression device described with reference to fig. 1 and the signal search device described with reference to fig. 2 in embodiments 1 and 2, and the components are components that perform the same operations as those of the signal compression device or the signal search device, and therefore, the description thereof is omitted.

The signal search device shown in fig. 3 also receives, as input signals, a video signal to be sample searched, which is a reference signal, and a video signal to be searched, which is an accumulated signal, and outputs a preset value (referred to as a search threshold) of the distance from the reference signal at θ ₁ The following accumulate locations in the signal.

Next, the operation of the signal search device shown in fig. 3 will be described with reference to fig. 16 and 25. First, the overall operation flow of the signal search device will be described with reference to fig. 16, and the operation characteristic of the signal search device of the present embodiment is that, in the operation flow described in embodiment 2, between "signal detection determination processing" in step S91 and "determination processing whether the current position of the annotation window is the end point of the accumulated signal" in step S95, the jump width of the attention window set for the series of accumulated compression characteristics output by the signal compression unit 5 is calculated, and jump width calculation processing for moving the annotation window only by the jump width is executed (step S92).

Next, referring to fig. 25, the jump width calculation process shown in fig. 16 (step S92) will be described in detail. Fig. 25 is a flowchart showing the operation of the jump width calculating unit 11.

In fig. 25, first, the jump width calculating unit 11 reads the distance lower limit value outputted from the feature checking unit 9 (step S144). Next, feature collation that guarantees no search omission is calculated as usual, i.e., the jump width of the distance calculation can be saved (step S145).

The following explains a specific principle of determining the jump width.

The histograms classify and accumulate time series of feature vectors, so the distance values between the histograms do not change sharply with the movement of the time window to the accumulated signal feature vectors. The distance of movement of each eigenvector portion of the time window varies from the absolute value by no more than (\58286;). That is, the distance value between histograms in the case where the m 1-th feature vector is at the head of the accumulated signal in the time window is d (x) _R ，x _S (m ₁ ) Time window moves to the m2 th feature vector, and the lower limit d (x) of the distance value _R ，x _S (m ₂ ) When m) is present ₁ ＜m ₂ ＜m ₁ At + D, it is given by the following equation.

d(x _R ，x _S (m ₂ ))＝d(x _R ，x _s (m ₁ ))-(m ₂ -m ₁ )…(5)

Where D represents the time window width, and equation (5) is modified from equation (4) as follows.

The distance value is not below "0", and therefore, when the lower limit value is given by the equation (6)d ^* When the value falls to "0", the value "0" becomes the lower limit value. By searching for the threshold value theta ₁ Replacing the lower limit value by m with the jump width omega ₂ -m ₁ The jump width can be found as follows.

Wherein floor (x) represents the largest integer not exceeding x.

The compressed accumulated features are extracted from the beginning of the compressed feature series at the start of the processing, but the program performs the processing while moving the position where the compressed accumulated features are extracted in the time direction (step S146) during the processing. The amount of temporal movement is given by the jump width calculation section 11.

[ 4 th example ]

Fig. 4 is a block diagram showing the structure of embodiment 4. In embodiment 4, the signal search device according to embodiment 3 will be described in which the signal detection determination unit 10 newly provides the distance recalculation unit 12 for recalculating the distance between the feature series output from the reference feature extraction unit 6 and the feature series output from the accumulated feature extraction unit 7 at the point of the accumulated signal where the reference signal is determined to be present, and the signal search device of the signal detection recalculation unit 13 for recalculating whether or not the reference signal is present at the point of the accumulated signal by comparing the distance output from the distance recalculation unit 12 with the search threshold. In fig. 4, the same reference numerals are given to the same components as those of the signal compression device and the signal search device described in fig. 1 to 3 in the embodiments 1 to 3, and the components that perform the same operations as those of the signal compression device and the signal search device are omitted here.

The signal search device shown in fig. 4 also receives as input a search image signal, which is a reference signal to be a sample, and a searched image signal, which is an accumulated signal, and outputs a predetermined distance value (referred to as a search threshold) from the reference signal at θ ₁ The following accumulate locations in the signal.

Next, the operation of the signal search device shown in fig. 4 will be described with reference to fig. 16, 26, and 27. First, with reference to fig. 16, the overall operation flow of the signal search device will be described, and the operation characteristic of the signal search device of the present embodiment is executed, and with respect to the operation flow described in embodiment 3, between the "jump amplitude calculation process" in step S92 and the "determination process whether or not the current position of the gaze window is the end point of the accumulated signal" in step S95, the distance recalculation process (step S93) of calculating the distance between the feature series output by the reference feature extraction unit 6 and the feature series output by the accumulated feature extraction unit 7 for the location of the accumulated signal for which the signal detection determination unit 10 has determined that the reference signal is present, and the signal search re-determination process (step S94) of re-determining whether or not the reference signal is present at the location of the accumulated signal by comparing the distance recalculated process with the search threshold value.

Next, the distance recalculation processing (step S93) shown in fig. 16 will be described in detail with reference to fig. 26. Fig. 26 is a flowchart showing the operation of the distance recalculating unit 12.

In fig. 26, first, the distance recalculating unit 12 reads the reference histogram output from the reference feature extracting unit 6, the cumulative histogram series output from the cumulative feature extracting unit 7, and the detection result output from the signal detection determining unit 10 (step S148). Next, the distance from the reference histogram is calculated for the cumulative histogram corresponding to the point in the cumulative signal determined to have the reference signal (step S149). The distance between histograms is defined by the euclidean distance as in the above equation (1). Then, the obtained distance value is output (step S150).

Next, the signal detection re-determination process (step S94) shown in fig. 16 will be described in detail with reference to fig. 27. Fig. 27 is a flowchart showing the operation of the signal detection re-determination unit 13.

In fig. 27, first, the signal detection re-determination unit 13 reads the distance value output from the distance re-calculation unit 12 (step S152). Next, the distance value is compared with a search threshold (step S153). As a result of the comparison, when the distance value is equal to or less than the search threshold (YES in step S153), this reference signal is present in the accumulated signal, and therefore the current position in the time series for the accumulated signal is output as a signal detection result (step S154).

In the signal search device shown in fig. 4, the jump width calculating unit 11 may be provided as needed, or may be provided without providing it.

< results of the experiment >

Next, the operation experiment results of the signal search device of the present invention will be described.

In order to confirm the effect of the present invention, first, as the 1 st experiment, assuming that a signal of a histogram created from a 24-hour acoustic signal is an accumulated signal, a change in the average dimension of a compressed histogram normalized by the number of segments and the segment length when an action threshold value changes is examined.

In addition, the compression parameters are set as follows: sampling frequency =29.97[ hz ], number of image segmentations =6 (2 for vertical segmentation and 3 for horizontal segmentation), dimension of histogram =256, and amplitude of time window =15[ seconds ]. Also, the split boundary movable amplitude Δ varies from 1 up to 500, like 1,2, 5, 10, \ 8230A.

Fig. 28 shows the results of experiment 1. The abscissa of the graph represents the dividing boundary movable Width Δ (represented as "Width of movable range" in fig. 28), and the ordinate represents the Ratio of the average dimension based on the average dimension at Δ =0 (represented as "Ratio of dimensions" in fig. 28). For example, when the division boundary movable range Δ =500, the segmentation number M =1000 2 [ segments ], the action threshold (distribution rate) σ =0.75, the average dimension is 2.91, the average dimension when the division boundary movable range Δ =0 is 3.30, and the dimensional reduction ratio is 0.882.

Next, as the 2 nd experiment, it is assumed that a signal in which a histogram is formed from 24-hour acoustic signals is an original signal, and the number of times of calculating the average dimension by the partial signal configuration selecting unit 2 and the partial signal reconstructing unit 3 is examined. In addition, the number of segments M =1000[ segments ], and a threshold value of action (distribution) σ =0.75 are set for the searched parameters, as in experiment 1.

Fig. 29 shows the results of experiment 2. The abscissa of the graph indicates the split boundary movable range Δ (indicated as "Width of movable range" in fig. 29), and the ordinate indicates the Number of calculations (indicated as "Number of calculation" in fig. 29). For example, when the split boundary movable range Δ =500 is set, the method according to the present invention (curve is formed as "speed" in fig. 29) calculates the number of times of 80000 times, the method according to the non-implementation partial signal configuration selector 2 (curve is formed as "non-speed" in fig. 29) calculates the number of times reduction ratio to be about 12.5.

[ 5 th embodiment ]

Next, embodiment 5 of the present invention will be described with reference to the drawings.

In the present embodiment, an acoustic signal is used as an example of the processing target signal. The specific feature extraction method and the histogram creation method are explained by using the processing of the query feature extraction unit 101 described later.

Fig. 30 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 5 is applied. Referring to fig. 30, the signal search device of the present embodiment includes an inquiry feature extracting unit 101 (reference feature extracting unit) for deriving a feature from an inquiry signal (reference signal); a database feature extraction unit (cumulative feature extraction unit) 102 that sets a gazing window for a database signal (cumulative signal) and derives a feature from the signal in the gazing window; a database feature classification unit 103 for classifying a feature series repeatedly output while shifting the observation window in the database feature extraction unit 102; a database feature interval thinning unit 104 for extracting a representative feature from the sorted feature series output from the database feature sorting unit 103 and deriving a representative feature series composed of a smaller number of features; a feature region extracting unit 105 for deriving a region in which a feature included in the classification output by the database feature classification unit 103 exists; a feature matching unit 10 for calculating a distance between the feature series output from the query feature extraction unit 101 and the representative feature series output from the database feature thinning unit 104; a distance correcting unit 107 for correcting the distance output from the characteristic checking unit 106 using the region output from the characteristic checking unit 106; and a signal detection/determination unit 108 for determining whether or not the query signal is present at the position of the database signal by comparing the distance output from the distance correction unit 107 with a search threshold corresponding to the distance threshold.

The signal search device shown in fig. 30 receives as input an acoustic signal desired to be a sample search, which is a query signal, and an acoustic signal to be searched, which is a database signal, and outputs a signal having a distance from the query signal in advanceSet value (called search threshold) θ ₁ The following database of locations in the signal.

Next, an operation flowchart of the signal search device shown in fig. 39 will be described with reference to fig. 30, 38 to 42. The processing of the query feature extraction unit 101, the database feature extraction unit 102, and the signal detection determination unit 108 is the same as that of embodiment 2 described with reference to fig. 16 to 24.

Fig. 38 is a flowchart showing the processing of the database feature distinguishing unit.

Next, as shown in fig. 38, the database feature classification unit 103 reads a series (feature series) of DB histograms (cumulative histograms) output in the processing of the database feature extraction unit 102 while shifting the attention window (step S164). Then, the database feature distinguishing unit 103 distinguishes the histogram series (step S165). Various methods of this distinction can be considered, but two methods are described here. In the first method, a histogram series is equally divided in a predetermined division width. For example, when the segmentation amplitude is 50, a plurality of partial histogram series (partial feature series) of 50 frames (12501125242452. In the method of the 2 nd method, the length of the partial histogram series is adjusted so that each histogram of the partial histogram series is within a certain distance from the histogram representing the partial histogram series derived by the feature region extraction unit 105. Specifically, the following is performed. First, a partial histogram series is set to a certain length, and a representative histogram is extracted from the series. The representative histogram is selected according to the processing of the database inter-feature thinning section 104. Next, the distance from each histogram of the partial histogram series to the representative histogram is calculated, and the maximum value thereof is obtained. The above operation is repeated for each length of the partial histogram series from 1 until the maximum value is equal to or less than the predetermined threshold value. The above is the 2 nd method of the differential histogram. The database feature distinguishing unit 103 outputs a set of the divided partial histogram series (step S166).

Fig. 39 is a flowchart showing a process of the database inter-feature thinning section. Next, as shown in fig. 39, the inter-database-feature thinning unit 104 reads the partial histogram series set output from the database-feature distinguishing unit 103 (step S167). Next, the database inter-feature thinning unit 104 extracts representative features from the partial histogram series (step S168). Although various representative feature extraction methods are conceivable, in the first method, any one of the partial histogram series is used as a representative feature as it is. For example, the first histogram of the partial histogram series is represented. In the 2 nd method, the centroid (of euclidean distance space) of the histogram is calculated among the partial histogram series, and this is taken as a representative. Finally, the representative features are arranged in the original partial histogram series program, and a new series is formed (step S168). Hereinafter, the feature series derived by the interval program is referred to as a representative feature series. The database inter-feature thinning unit 104 outputs a representative feature series (step S170).

Fig. 40 is a flowchart showing the processing of the feature region extracting unit.

Next, as shown in fig. 40, the feature region extraction unit 105 reads the partial histogram series set output by the database feature classification unit 103 and the representative feature series output by the database feature thinning unit 104 (step S1 71). Next, for each histogram in the partial histogram series, the distance from the representative feature is calculated, and the maximum value dmax thereof is calculated (step S172). In accordance with the above, the range within which the histogram within the partial histogram series exists can be determined, which is within a distance dmax from the representative feature. Then, the characteristic region extraction unit 105 outputs a set of dmax (step S173).

Fig. 41 is a flow chart showing the process of the feature collating part 2.

Next, as shown in fig. 41, the characteristic checking section 106 reads the representative characteristic series output from the database characteristic thinning section 104 and the query characteristic output from the database characteristic extracting section 102 (step S174). The feature matcher 106 then calculates a query feature x _Q And represents feature x _D Is detected (step S175). Although various distance metrics may be used, for example, a manhattan distance or a euclidean distance is used.

The manhattan distance is defined as:

the euclidean distance is defined as:

the feature check part 106 outputs a distance d (x) _Q ，x _D ) (step S176).

Fig. 42 is a flowchart showing a process of the distance correcting section.

Next, as shown in fig. 42, the distance correction unit 107 reads the distance value dmax output from the characteristic region extraction unit 105 and the distance value d (x) output from the characteristic check unit 106 _Q ，x _D ) (step S177). The distance correcting section 107 then calculates a corrected distance value. I.e., calculating d (x) _Q ，x _D ) Dmax (step S178). For features corresponding to the representative featuresThe partial histogram series, the minimum value of the distance to one of the histograms is obtained. The distance correcting unit 107 then outputs the corrected distance value (corrected distance value) (step S179).

[ 6 th example ]

Next, embodiment 6 of the present invention will be described with reference to the drawings.

Fig. 31 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 6 is applied. The signal search device according to the present embodiment is such that the signal search device according to embodiment 5 further includes: a segment extraction unit 10; a compressed image determining unit (110); a database feature compression section (cumulative compression section) 111; and a search feature compression unit (reference feature compression unit) 112 for receiving the acoustic signal to be searched for as a sample, which is the query signal, and the acoustic signal to be searched for as a database signal, and outputting a predetermined value of the distance from the query signal (referred to as a search threshold) at θ ₁ The following database locations in the signal.

The segment extraction unit 109 extracts segments (partial signals) as a partial sequence by dividing the feature sequence repeatedly derived by the database feature extraction unit 102 while shifting the annotation window. The compressed map determination unit 110 determines a feature map for calculating a feature number lower than the feature number, based on each segment output by the segment extraction unit 109. The database feature compressing unit 111 calculates features having a lower feature dimension than the feature dimension corresponding to the segment extracted by the segment extracting unit 109, based on the map output by the compressed map determining unit 110. The query feature compressing unit 112 calculates features having a dimension lower than the feature dimension corresponding to the features output by the query feature extracting unit 101, based on the map output by the compressed map determining unit 11 0.

Next, the processing of the signal search device according to the present embodiment will be described with reference to fig. 31 and 43.

First, the processes of the query feature extraction unit 101, the database feature extraction unit 102, the compressed map determination unit 110, the database feature compression unit 111, and the query feature compression unit 112 are the same as those of embodiment 2, and therefore, the description thereof will be omitted.

FIG. 43 is a flowchart showing the processing of the segment extracting section.

The segment extraction unit 109 can be considered to have 2 configurations. The first configuration is a configuration in which only the initial partial signal forming section 1 is used. The structure of type 2 is a structure in which the initial partial signal configuring section 1, the partial signal configuring selecting section 2, and the partial signal reconstructing section 3 are used. Since any of the processes is similar to that of embodiment 2, it is omitted (step S183 to step S184). The segment extraction unit 109 outputs a set of segments as a segmented histogram series (step S185).

Then, the database feature distinguishing unit 103 to the signal detection determination unit 108 are executed. Note that the processing from the database feature distinguishing unit 103 to the signal detection judging unit 108 is the same as that in embodiment 5, and therefore is omitted.

[ 7 th example ]

Next, embodiment 7 will be described with reference to the drawings.

Fig. 32 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 7 of the present invention is applied. The signal search device according to the present embodiment is configured such that the signal search device according to embodiment 6 is further added with a configuration of a distance recalculating unit 113 and a signal detection and re-determination unit 114, and an acoustic signal to be searched for, which is a query signal and is desired to be a sample search, and an acoustic signal to be searched for, which is a database signal, are input, and a preset value (referred to as a search threshold) of a distance from the query signal is output at θ ₁ The following database of locations in the signal.

Here, the distance recalculating unit 113 calculates the distance between the feature series output by the query feature extracting unit 101 and the feature series output by the database feature extracting unit 102 for the location of the database signal determined to be the query signal by the signal detection determining unit 108. The signal detection re-determination unit 114 compares the distance output from the distance re-calculation unit 113 with the search threshold, and then determines whether or not the query signal is present at the corresponding point of the database signal.

The other processes are the same as those in embodiment 4, and thus omitted.

[ 8 th example ]

Next, embodiment 8 will be described with reference to the drawings.

Fig. 33 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 8 of the present invention is applied. The signal search device of the present embodiment is configured such that the jump width calculation unit 118 is added to the signal search device of embodiment 7, and that a search signal, i.e., an acoustic signal desired to be subjected to sample search and a searched acoustic signal, i.e., a database signal, are input, and a preset value (referred to as a search threshold) of the distance from the search signal is output at θ ₁ The following database locations in the signal.

Here, the jump width calculation unit 118 calculates the set attention window jump width for the compression feature series output from the database feature compression unit 111 based on the distance output from the feature check unit 106, and moves the attention window only by the jump width. The processing is the same as in example 3. The other processes are the same as those in embodiment 7, and thus omitted. In the signal search device shown in fig. 33, the distance recalculating unit 113 and the signal detection re-determining unit 114 may be provided as needed, but may not be provided if necessary.

< results of the experiment >

Next, the operation experiment results of the signal search device according to embodiment 8 will be described.

Assuming that the sound signal of the 200-hour section is an accumulated signal, this is to search 200 reference signals of 15-second different signals arbitrarily selected from separately prepared signals, and to study the time required for the search after the provision of the reference signals and the number of times of execution of the processing by the feature check section 106, that is, the number of times of checking the features. In addition, the parameters at this time are: dimension of the feature vector =7, time amplitude of the feature vector =60msec, time scale of the feature vector =10msec, dimension of the histogram =128, and the compressed feature is made every a =50 frames. And assume action threshold =0.9 and search threshold =85.

Fig. 48 and 49 show the experimental results of this experiment. FIG. 48 is a 3 rd explanatory view showing an experimental result. FIG. 49 is a 4 th explanatory view showing an experimental result.

The horizontal axis represents the number of segments, and the vertical axis represents the time required for retrieval (fig. 48) and the number of checks (fig. 4). When the number of segments is 10000, the time required for the search is 0.364 seconds and the number of checks is 772784 for the method of the present invention (deployed method), the time required for the search is 1.491 seconds and the number of checks is 1036493 for the method (project distance unavailable) in which the search is not performed using the projection distance, and the time required for the search is 4.218 seconds and the number of checks is 633047 for the conventional method (japanese patent No. 3065314: contextual method).

[ 9 th embodiment ]

Next, embodiment 9 will be described with reference to the drawings.

Fig. 34 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 9 of the present invention is applied. The signal search device according to this embodiment is configured such that the signal search device according to embodiment 7 is further added with a database feature classification unit 115, a selection threshold setting unit 116, and a database feature selection unit 117, and that an acoustic signal to be a sample search, which is a query signal, and a searched acoustic signal, which is a database signal, are input, and a preset value (referred to as a search threshold) of a distance from the query signal is output at θ ₁ The following database of locations in the signal.

The database feature classification unit 115 determines representative features of the classification by repeating the features derived by classification while shifting the attention windows in the processing of the database feature extraction unit 102 based on a predetermined distance. The selection threshold setting unit 116 calculates a selection threshold for the distance defined by the database feature classification unit 115 based on a predefined search threshold. The database feature selecting unit 117 selects features included in the classification in which the distance from the feature output by the query feature extracting unit 101 satisfies the condition derived from the selection threshold output by the selection threshold setting unit 116, among the classifications output by the database feature classifying unit 115.

Next, the processing of the signal search device according to the present embodiment will be described with reference to fig. 34 and 44 to 47. First, the processes of the query feature extraction unit 101 and the database feature extraction unit 102 are the same as those in embodiment 7, and therefore, the description thereof is omitted.

Fig. 44 is a flowchart showing the processing of the database feature classification unit. Next, as shown in fig. 44, the database feature classification unit 115 reads a histogram series output by the database feature extraction unit 102 while repeating the output with the attention windows shifted (step S210). Next, each histogram of the histogram series is classified by, for example, euclidean distance (step S211). The histogram classification is considered to be performed by encoding a vector having a dimension equal to the number of bins (bins) of each histogram by vector quantization. For example, if the number of code words for vector quantization is 1024, the histogram should be classified as one of 1024 sets (called clusters). Then, a representative cluster is determined by a histogram (called a centroid histogram) that is the centroid of the histogram to which each histogram belongs. In this case, the cluster is configured such that the sum of the distances between the histogram to which the cluster belongs and the histogram of the center of gravity is minimized, and the distance between the histogram to which the cluster belongs and the histogram of the center of gravity of the cluster to which the cluster belongs is reduced to be smaller than the distances between the histograms of the center of gravity of the other clusters. Then, the database feature classification unit 115 outputs a set of clusters classified as histograms (step S212).

Fig. 45 is a flowchart showing a process of selecting the threshold setting unit.

Next, as shown in FIG. 45, the selection threshold setting section 116 reads the search threshold θ ₁ (step S213). According to the search threshold theta ₁ Calculating a selection threshold value theta ₂ (step S214). The selection threshold here means that the selection is possibleWhen a cluster including a DB histogram corresponding to a signal to be retrieved is found, the histogram is queried to be at an upper distance limit from the cluster. When the distance scale during the search is the euclidean distance, the selection threshold is the same value as the search threshold. And is common with the scale used in classification and selection, so that search omission does not occur. When the distance scale during time checking is the manhattan distance, the parameter p is used, and the selection threshold is set as follows.

As p increases, the selection area of the cluster narrows, and p =0 is the maximum value theoretically ensuring that no search omission occurs. In practice, search omission hardly occurs even if p =1 or so, and thus, for example, a selection threshold is calculated as p =1. Then, the selection threshold setting unit 116 outputs the selection threshold θ ₂ (step S215).

Fig. 46 is a flowchart showing the processing of the database feature selecting unit.

Next, as shown in fig. 46, the database feature selection unit 117 reads the query histogram output by the query feature extraction unit 101, the classification (clustering) of the histogram output by the database feature classification unit 115, and the selection threshold output by the selection threshold setting unit 116 (step S216). Next, the (euclidean) distance between the read query histogram and the histogram of the center of gravity of each histogram on the database signal side is calculated (step S217). Next, based on the calculated distances, a cluster is selected that may include a histogram corresponding to the signal to be retrieved (step S218). The principle thereof is explained below.

Here, FIG. 47 shows the results at Q and C ₁ 、C ₂ This plane of the three-point configuration truncates the sample subgraph in histogram space (128 dimensions in the above example). Here, Q denotes a query histogram, C ₁ Center of gravity histogram, C, representing the cluster to which histogram Q belongs ₂ Histogram of centers of gravity representing other clusters, d _Q1 、d _Q2 And d _Q3 Respectively represent Q and C ₁ Distance of (A), Q and C ₂ A distance of (C), and ₁ and C ₂ The distance of (c). Here, if it is necessary to detect a database signal corresponding to a histogram within a distance d from the histogram Q, the location corresponds to a database signal corresponding to a histogram located inside a hyper-sphere (circle in fig. 47) having a radius d around Q as a center. When the radius of the hyper-sphere centered on Q is increased to be larger than d _θ When it belongs to C ₂ The histogram of the representative cluster may include a histogram corresponding to the location of the database signal to be retrieved. Therefore, the threshold θ is selected ₂ Increase to greater than d _θ When, select C ₂ A cluster of representatives.

D is obtained as follows _θ . According to fig. 47, the following equation holds.

Is obtained by the formula (10),

therefore, from the formula (11),

when this is true, all histograms belonging to the cluster represented by C2 are selected (step S219). This procedure is performed for all clusters except the cluster to which the histogram Q belongs, and the location in the database signal corresponding to the selected histogram is output (step S220). The following processing is executed only for the spot output in the database feature selection unit 117.

The processes of the segment extraction unit 109 to the database feature compression unit 111, and the processes of the database feature segmentation unit 103 to the signal detection determination unit 108, the distance recalculation unit 113, and the signal detection re-determination unit 114 are performed. These processes are the same as those in embodiment 7. The distance recalculating unit 113 and the signal detection re-determining unit 114 may be provided as necessary, or may not be provided if necessary.

Next, the query feature compressing unit 112 reads the location in the database signal output by the database feature selecting unit 117, the query histogram output by the query feature extracting unit 101, and the linear map set output by the compressed map determining unit 110. Next, each line image is mapped to a partial space corresponding to the histogram. The map is used at the location output by the database characteristics selecting unit 117. Moreover, it is sufficient to perform line mapping, and it is not necessary to perform all line mapping. Thus exhibiting a peculiar effect of reducing the processing time.

[ 10 th embodiment ]

Next, the 10 th embodiment will be described with reference to the drawings.

Fig. 35 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 10 of the present invention is applied. The signal search device of the present embodiment is configured such that the jump amplitude calculation unit 118 is added to the signal search device of embodiment 9, and that a search signal, i.e., an acoustic signal desired to be subjected to a sample search, and a searched acoustic signal, i.e., a database signal, are input, and a preset value of the distance from the search signal (referred to as a search threshold) is output at θ ₁ The following database of locations in the signal.

The jump width calculation unit 118 calculates the jump width of the gaze window based on the distance output from the distance correction unit 107, and moves the gaze window by only the jump width. The processing of the jump width calculating unit 118 is omitted as in embodiment 3.

< results of the experiment >

Next, the operation experiment results of the signal search device of the present embodiment will be described.

Here, the experimental conditions are the same as those in example 7. As another parameter, the number of clusters is set to 1024.

Fig. 50 and 51 show the results of this experiment. FIG. 50 is a 5 th explanatory view showing the experimental results. FIG. 51 is a 6 th explanatory view showing the experimental results. Here, the horizontal axis represents the number of segments, and the vertical axis represents the time required for retrieval (fig. 50) and the number of checks (fig. 51). When the number of segments is 10000, the time required for retrieval is 0.234 seconds and the number of check times is 305351 times for the method of the embodiment; for the method (using feature compression) which only implements feature compression, the time required for retrieval is 0.364 seconds, and the number of times of check is 772784 times; for the existing method (Time-series Active Search), the Time required for searching is 4.218 seconds, and the number of checks is 633047 times.

[ 11 th embodiment ]

Next, embodiment 11 will be described with reference to the drawings.

Fig. 36 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 11 of the present invention is applied. The signal search device of the present embodiment is configured such that the distance recalculating unit 113, the signal detection re-determining unit 114, the database feature classifying unit 115, the selection threshold setting unit 116, and the database feature selecting unit 117 are added to the signal search device of embodiment 5, and the acoustic signal desired to be a sample search, which is the query signal, and the acoustic signal to be searched, which is the database signal, are input, and a preset value of the distance from the query signal (referred to as a search threshold) is output at θ ₁ The following database of locations in the signal.

In the processing of the present embodiment, after the processing in the query feature extracting unit 101 and the processing in the database feature extracting unit 102 described in the above-described 5 th embodiment, the processing in steps S210 to S212 in the database feature classifying unit 115, the processing in steps S213 to S215 in the selection threshold setting unit 116, and the processing in steps S216 to S220 in the database feature selecting unit 117 described in the above-described 9 th embodiment are performed.

Next, the processing of steps S1 to S166 in the database feature distinguishing unit 103, steps S167 to S17 in the database feature thinning unit 104, and steps S171 to S173 in the feature area extracting unit 105 described in embodiment 5 is performed. Then, the feature check unit 106 reads the representative feature series output from the database feature thinning unit 104 and the query feature output from the database feature selecting unit 117, and performs the processing of steps S174 to S176. Then, the processing of steps S177 to S17 of the distance correction unit 107, the processing of the signal detection determination unit 108, and the processing of the distance recalculation unit 113 and the signal detection re-determination unit 114 described in embodiment 5 are performed. The distance recalculating unit 113 and the signal detection and re-determination unit 114 may be provided as needed, or may not be provided if not needed.

[ 12 th embodiment ]

Next, embodiment 12 will be described with reference to the drawings.

Fig. 37 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 12 of the present invention is applied. The signal search device of the present embodiment is configured such that the jump width calculation unit 118 is added to the signal search device of embodiment 11, and an acoustic signal to be searched for as a sample search, which is a query signal, and an acoustic signal to be searched for as a database signal are input, and a preset value (referred to as a search threshold) of a distance from the query signal is output at θ ₁ The following database of locations in the signal.

Here, the jump width calculation unit 118 calculates the attention window jump width set for the DB histogram series output from the database feature extraction unit 102 based on the distance output from the distance correction unit 107, and moves the attention window only by the jump width. The other processes are the same as those in embodiment 11.

[ 13 th example ]

Next, embodiment 13 of the present invention will be described with reference to the drawings.

Although the present embodiment can use various processing target signals, here, an acoustic signal is used as an example of the processing target signal.

Fig. 52 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 13 is applied. Referring to fig. 5, the signal search device of the present embodiment includes a query feature extraction unit 101, a database feature extraction unit 102, a database feature classification unit 115, a selection threshold setting unit 116, a database feature selection unit 117, a segment extraction unit 109, a compressed map determination unit 110, a database feature compression unit 111, a query feature compression unit 112, a feature matching unit 106, and a signal detection determination unit 108.

The signal search device shown in fig. 52 receives as input an acoustic signal desired to be a sample search, which is a query signal, and an acoustic signal to be searched, which is a database signal, and outputs a value θ where the distance from the query signal is set in advance (referred to as a search threshold value) ₁ The following database locations in the signal.

The query feature extraction unit 101 derives a feature from the query signal. The database feature extraction unit 102 sets a gaze window for the database signal, and derives the features while shifting the gaze window. The database feature classification unit 115 classifies the features derived by repeating the processing performed by the database feature extraction unit 102 while shifting the observation window according to a predetermined distance, and determines representative features of the classification. The selection threshold setting unit 116 calculates a search threshold for a distance defined in the database feature classification unit 11 based on a predefined search threshold. The database feature selection unit 117 selects a feature included in the classification in which the distance between the classification output from the database feature classification unit 115 and the feature output from the query feature extraction unit 101 satisfies a condition derived from the selection threshold output from the selection threshold setting unit 116.

The segment extracting unit 109 extracts segments as partial sequences by dividing the feature sequence derived by repeating the above-described database feature extracting unit 102 while shifting the annotation window. The compressed map determination unit 110 determines a map for calculating a feature lower than the feature dimension, based on each segment output from the segment extraction unit 109. The database feature compression unit 111 calculates features having a lower number of feature dimensions than the number of feature dimensions corresponding to the segment output by the segment extraction unit 109, based on the map output by the compressed map determination unit 1 10. The query feature compressing unit 112 calculates features having a dimension lower than the feature dimension corresponding to the features output by the query feature extracting unit 101, from the map output by the compressed map determining unit 110.

The feature check unit 106 calculates the distance between the compressed feature series output from the database feature compression unit 111 and the compressed feature series output from the query feature compression unit 112 for a point in the database signal output from the database feature selection unit 117. The signal detection determination unit 108 determines whether or not the query signal is present at the location of the database signal by comparing the distance output by the feature verification unit 106 with a search threshold value, which is a threshold value corresponding to the distance.

Next, the processing of the signal search device of the present embodiment will be described with reference to fig. 52 and 55.

The processing of the query feature extraction unit 101 and the database feature extraction unit 102 is the same as in embodiment 2. The processing of the database feature classification section 115, the selection threshold setting section 116, and the database feature selection section 117 is the same as in embodiment 9. The process of the segment extracting unit 109 is also the same as in embodiment 6. The processes of the compressed map determination unit 110, the database feature compression unit 111, the feature verification unit 106, and the signal detection determination unit 108 are the same as those of embodiment 2. The processing of the query feature compressing section 112 is the same as in embodiment 9.

[ 14 th embodiment ]

Next, embodiment 14 will be described with reference to the drawings.

Fig. 53 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 14 of the present invention is applied. The signal search device of the present embodiment is configured such that the distance recalculating unit 113 and the signal detection re-determining unit 114 are added to the signal search device of embodiment 13, and an acoustic signal to be sample searched, which is a query signal, and an acoustic signal to be searched, which is a database signal, are input, and an output and query signal are outputDistance of the hornThe predetermined value of the distance (called the search threshold) is θ ₁ The following database of locations in the signal.

Here, the distance recalculating unit 113 calculates the distance between the feature series output by the query feature extracting unit 101 and the feature series output by the database feature extracting unit 102 for the location of the database signal determined by the signal detection determining unit 108 to be the presence of the query signal. The signal detection re-determination unit 114 compares the distance output from the distance re-calculation unit 113 with the search threshold, and then determines whether or not the query signal is present at the location of the database signal.

The distance recalculating unit 113 and the signal detection and re-determination unit 114 perform the same processing as in embodiment 4.

The processes of the query feature extracting unit 101, the database feature extracting unit 102, the database feature classifying unit 115, the selection threshold setting unit 116, the database feature selecting unit 117, the segment extracting unit 109, the compression map determining unit 110, the database feature compressing unit 111, the query feature compressing unit 112, the feature checking unit 106, and the signal detection determining unit 108 are the same as those of embodiment 13, and therefore, the description thereof will be omitted.

[ 15 th embodiment ]

Next, embodiment 15 will be described with reference to the drawings.

Fig. 54 is a functional block diagram showing a signal retrieval apparatus to which the method of embodiment 15 of the present invention is applied. The signal search device of the present embodiment is configured such that the jump width calculation unit 118 is added to the signal search device of embodiment 14, and an acoustic signal to be searched for as a sample search, which is a query signal, and an acoustic signal to be searched for as a database signal are input, and a preset value (referred to as a search threshold) of a distance from the query signal is output at θ ₁ The following database of locations in the signal.

The jump width calculation unit 118 calculates the jump width of the note window based on the distance output from the characteristic checking unit 106, and moves the note window only by the jump width.

Next, the processing of the signal search device of the present embodiment is described.

First, the processes of the query feature extracting unit 101, the database feature extracting unit 102, the database feature classifying unit 115, the selection threshold setting unit 116, the database feature selecting unit 117, the segment extracting unit 109, the compressed map determining unit 110, the database feature compressing unit 111, and the query feature compressing unit 112, and the processes of the feature checking unit 106, the signal detection judging unit 108, the distance recalculating unit 113, and the signal detection re-judging unit 114; since both are the same as those in embodiments 13 and 14, the description thereof will be omitted. The processing of the jump width calculating unit 118 is the same as in embodiment 3.

< results of the experiment >

Next, an operation experiment example to which the apparatus of example 5 is applied is shown.

In order to confirm the effect of the present invention, first, a signal in which a histogram is created from an acoustic signal of 10 hours is set as a database signal, and the file size, search time, and the number of times of feature checks when a compressed feature is written in a file are examined. The parameters of the search are: acoustic signal sampling frequency =33kHz, feature vector dimension =7, feature vector temporal amplitude =60ms, feature vector time scale =10ms, histogram dimension =128, time window amplitude =15[ sec ], action threshold =0.9, number of segments =20, and search threshold =85. In feature thinning, the series is distinguished by equal segmentation, representing that the beginning of the partial histogram series is utilized as is. The distance scale at the time of the lookup is set as the euclidean distance.

FIG. 56 is a 7 th explanatory view showing an experimental result. FIG. 57 is a 9 th explanatory view showing an experimental result. Fig. 56 and 57 show the results of the experiment in which the amplitude of the partial histogram series was varied every 10 intervals from 10 to 50. The horizontal axis of the graph represents the partial histogram series width a, and the vertical axis represents the file size (fig. 5), the number of search points (left side of fig. 57), and the search time (right side of fig. 57). a =0, which is the case without the method of the invention.

According to fig. 56 and 57, the file size and the number of check points are monotonically decreased by increasing the magnitude of the partial histogram series. Further, according to fig. 57, as the number of search sites decreases, the search time is also reduced. The search time increases with a =40 as a boundary because points requiring histogram recheck increase by enlarging the range of existence of the feature.

From the above, it can be considered that a =40 is the optimum amplitude of the partial histogram series, and the file size at this time is 5.8 megabytes (about 1/30 when not using the present invention), the search time is 23 nanoseconds (about 60% when not using the present invention), and the number of check sites is 41855 (about 85% when not using the present invention).

As described above, according to the present invention, the time direction is compressed by the characteristics of the pre-thinning database signal, thereby significantly reducing the number of indexes without causing search omission, and providing an advantage of performing a higher-speed signal search as compared with the known method.

Further, the signal compression processing or the signal search processing can be performed by recording a program for realizing the functions of the processing units shown in fig. 1 to 4 on a computer-readable recording medium, and reading and executing the program recorded on the recording medium on a computer system. The term "computer system" as used herein is generally considered to include hardware such as an OS and an external device. Further, the term "computer system" is used in the case of using the WW W system, and is considered to include a homepage providing environment (or display environment). The "computer-readable recording medium" refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or the like, or a storage device such as a hard disk built in a computer system. Further, the term "computer-readable recording medium" is generally considered to include a medium that holds a program for a certain period of time, such as a server that transmits the program via a network such as the internet or a telephone line, and a volatile memory (RAM) that is an internal part of a client computer system.

The program may be transferred from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or via a transmission wave in the transmission medium. Here, the "transmission medium" for transmitting the program is a medium having a function of transmitting information, such as a network (communication network) such as the internet and a communication line (communication line) such as a telephone line. The program may be a part for realizing the above function. Further, the functions may be realized by combining the program in which all of the functions are recorded in the computer system, a so-called differential file (differential program).

Claims (34)

1. A signal search method for calculating a distance from an arbitrary point of an accumulated signal which is a previously registered original signal to a reference signal which is a set target signal, and finding a point which is lower than a threshold value preset by the distance from the accumulated signal to the reference signal, the method comprising:

an initial partial signal forming process for forming a partial signal having a length shorter than that of the original signal from the original signal;

a selection step of forming candidate partial signals, each of which is composed of partial signals having a smaller data amount than the original signals, into the partial signals derived in the initial partial signal forming step;

determining a partial signal reconstruction process for constructing a partial signal to be actually used, using the candidate composed of the partial signal derived by the partial signal construction selection process;

a compressed image determining step of determining an image for calculating a compressed signal from each partial signal obtained by the partial signal reconstructing step;

a signal compression step of calculating a compressed signal corresponding to each partial signal obtained in the partial signal reconstruction step, based on the map obtained in the compressed map determination step;

a reference feature extraction process for deriving features from the reference signal;

setting a gaze window for the accumulated signal, and deriving an accumulated feature extraction process for the features from the signal within the gaze window;

a reference feature compression step of compressing the reference feature derived in the reference feature extraction step, based on the map derived in the compressed map determination step;

a feature check step of calculating a distance between a reference compressed signal derived in the reference feature compression step and an accumulated compressed signal derived in the signal compression step by reusing a feature series derived by repeating the accumulated feature extraction step while shifting a gaze window; and

a signal detection determination step of determining whether or not a reference signal is present at the position of the accumulated signal by comparing the distance derived in the feature check step with a search threshold value which is a threshold value corresponding to the distance,

the processing according to the feature checking process and the signal detection determining process is repeated while shifting the viewing window.

2. The signal retrieval method according to claim 1, comprising:

a distance recalculation step of calculating a distance between the feature series derived in the reference feature extraction step and the feature series derived in the cumulative feature extraction step for the location of the database signal determined to be the presence of the query signal in the signal detection determination step; and

a signal detection re-decision process of re-deciding whether the query signal is present at the location of the database signal by comparing the distance derived in said distance re-calculation process with a search threshold,

while the watching window is staggered, the processing of the characteristic checking process, the signal detection judging process, the distance recalculating process and the signal detection judging process is repeated, the distance between the inquiry signal and a plurality of places of the database signal is calculated, and whether the inquiry signal exists in the place of the database signal is determined.

3. The signal retrieval method according to claim 1,

comprising: a jump amplitude calculation step of calculating a jump amplitude of the gaze window based on the distance calculated in the feature check step and moving the gaze window by the jump amplitude,

while the fixation window is staggered, the processing of the characteristic checking process, the signal detection judging process and the jumping amplitude calculating process is repeated, the distance between the data base signal and the inquiry signal is calculated for several positions of the data base signal, and whether the inquiry signal exists in the position of the data base signal is determined.

4. A signal search device for calculating a distance from a reference signal as a set target signal to an arbitrary point of an accumulated signal as a previously registered original signal and finding out a point lower than a threshold value preset by the distance from the reference signal from the accumulated signal, the signal search device comprising:

an initial partial signal forming unit for forming a partial signal having a length shorter than that of the original signal from the original signal;

a partial signal configuration selecting unit configured to squeeze a candidate partial signal configuration having a smaller data amount than the original signal into each partial signal derived by the initial partial signal configuring unit;

a partial signal reconstruction unit for determining a partial signal configuration to be actually used, using the partial signal configuration candidates derived by the partial signal configuration selection unit;

a compressed image determining unit that determines an image for calculating a compressed signal from each partial signal obtained by the partial signal reconstructing unit;

a signal compression unit for calculating a compressed signal corresponding to each partial signal obtained by the partial signal reconstruction unit, based on the map obtained by the compressed map determination unit;

a reference feature extracting unit that derives a feature from the reference signal;

a cumulative feature extraction unit for setting a gaze window for the cumulative signal and deriving a feature from the signal in the gaze window;

a reference feature compressing unit that compresses the reference feature derived by the reference feature extracting unit, based on the map derived by the compressed map determining unit;

a feature checking unit that calculates a distance between the reference compressed signal derived by the reference feature compressing unit and an accumulated compressed signal derived from the signal compressing unit by reusing the feature series derived by repeating the accumulated feature extracting unit while shifting the attention window; and

a signal detection determining unit that determines whether or not the reference signal is present at the point of the accumulated signal by comparing the distance derived by the feature checking unit with a search threshold value that is a threshold value corresponding to the distance,

the feature checking means and the signal detection judging means are repeatedly operated while shifting the note window.

5. The signal retrieval device according to claim 4, comprising:

a distance recalculating unit that calculates a distance between the feature series derived by the reference feature extracting unit and the feature series derived by the cumulative feature extracting unit for the location of the database signal determined by the signal detection determining unit to be the query signal; and

a signal detection re-determination section for re-determining whether or not the query signal is present at the position of the database signal by comparing the distance derived in the distance re-calculation section with a search threshold,

the processing of the characteristic checking means, the signal detection judging means, the distance recalculating means, and the signal detection re-judging means is repeated while shifting the attention window, and the distance to the query signal is calculated for several locations of the database signal, and it is determined whether the query signal is present at the location of the database signal.

6. The signal retrieval device of claim 4,

the disclosed device is provided with: a jump amplitude calculating section for calculating a jump amplitude of the gaze window based on the distance calculated in the feature checking section and moving the gaze window by the jump amplitude,

the processing by the characteristic checking means, the signal detection judging means and the jump width calculating means is repeated while the watching window is shifted, and the distance to the query signal is calculated for several places of the database signal, and it is determined whether the query signal exists at the place of the database signal.

7. A signal search method for searching out a portion lower than a threshold preset for a distance from a query signal as a target from a database signal registered in advance, the method comprising:

a query feature extraction process for deriving features from the query signal;

setting a gazing window for a database signal, and deriving a database characteristic extraction process of characteristics from the signal in the gazing window;

a database feature distinguishing step of distinguishing the derived feature series by repeating the database feature extraction step while staggering the gaze window;

extracting representative features from the distinguished feature series obtained in the database feature distinguishing process, and deriving a database feature thinning process of the representative feature series consisting of fewer features;

a feature region extraction step of deriving a region in which a feature included in the discrimination derived in the database feature discrimination step exists;

a feature matching process of calculating a distance between a feature series derived in the query feature extraction process and a representative feature series derived in the database feature thinning process;

a distance correction step of correcting the distance calculated in the feature verification step using the region derived in the feature region extraction step; and

a signal detection determination process of determining whether or not the query signal exists at the location of the database signal by comparing the corrected distance derived in the distance correction process with a search threshold value as a threshold value corresponding to the distance,

the distance between the inquiry signal and a plurality of locations of the database signal is calculated by repeating the processing from the characteristic checking process to the signal detection judging process while shifting the observation window, and whether or not the inquiry signal exists at the location of the database signal is determined.

8. The signal retrieval method as claimed in claim 7, wherein in the database feature thinning-out process, any one feature within the specified section is a representative feature.

9. A signal retrieval method according to claim 7 wherein the feature barycenters within the defined region are representative features in the thinning-out of the database features.

10. The signal retrieval method according to claim 7, wherein in the database feature classification step, the feature series derived by repeating the database feature extraction step while shifting the fixation window is equally divided by a predetermined length.

11. The signal retrieval method according to claim 7, wherein in the database feature classification step, the feature series derived by repeating the database feature extraction step while shifting a window of interest is divided so that a feature existing region derived in the feature region extraction step is smaller than a predetermined maximum region.

12. The signal retrieval method according to claim 7, comprising:

a segment extraction step of dividing a feature series derived by repeating the database feature extraction step while shifting the focus window, and extracting segments as a partial series;

a compressed image determining step of determining an image for calculating a feature having a dimension lower than the feature dimension from each segment obtained in the segment extracting step;

a database feature compression step of calculating features lower than the feature dimension corresponding to the segments obtained in the segment extraction step, based on the map obtained in the compressed map determination step; and

a query feature compression step of calculating a feature having a dimension lower than the feature dimension corresponding to the feature obtained in the query feature extraction step, based on the map obtained in the compressed map determination step,

in the database feature thinning process, the compressed feature series derived in the database feature compression process is used as a new feature series derived representative feature series, in the feature checking process, the compressed feature derived in the query feature compression process is checked as a new feature, and further, the processing from the feature checking process to the signal detection judging process is repeated while the watching window is staggered, and the distance from the query signal is calculated for several places of the database signal, so as to determine whether the query signal exists at the place of the database signal.

13. The signal retrieval method according to claim 12, wherein,

the database feature compression process has:

a database feature mapping process of mapping the segments obtained in the segment extraction process according to the images obtained in the compressed image determination process;

a database projection distance calculation step of calculating a distance between the compressed feature series derived in the database feature mapping step and the feature series derived in the database feature extraction step; and

a database compressed feature construction process of reconstructing a compressed feature series based on the compressed feature series derived in the database feature mapping process and the projection distances derived in the database projection distance calculation process,

the query feature compression process has:

a query feature mapping process of mapping the features obtained in the query feature extraction process according to the map obtained in the compressed map determination process;

a query projection distance calculation process of calculating a distance to a feature derived in the query feature extraction process for the compressed feature derived in the query feature mapping process; and

and a query compression feature construction process for reconstructing a compression feature based on the compression feature derived in the query feature mapping process and the projection distance derived in the query projection distance calculation process.

14. A signal retrieval method according to claim 12 wherein the compressed map determination process extracts representative features by Karhunen-Loeve unfolding.

15. The signal retrieval method according to claim 7, comprising:

a distance recalculation step of calculating a distance between the feature derived in the query feature extraction step and the feature series derived in the database feature extraction step for the location of the database signal determined to have the query signal in the signal detection determination step; and

a signal detection re-decision process for re-deciding whether the query signal is present at the location of the database signal by comparing the distance derived in the distance re-calculation process with a search threshold,

while the watching window is staggered, the processing of the characteristic checking process, the distance correcting process, the signal detection judging process, the distance recalculating process and the signal detection judging process is repeated, the distance between the inquiry signal and a plurality of places of the database signal is calculated, and whether the inquiry signal exists in the place of the database signal is determined.

16. The signal retrieval method according to claim 7, comprising:

a database feature classification step of classifying, based on a predetermined distance, each feature derived by repeating the database feature extraction step while shifting the observation window, and determining a representative feature of the classification;

a selection threshold setting process of calculating a selection threshold for a distance defined in the database feature classification process from a predetermined search threshold; and

and a database feature selection step of selecting a feature included in a class of a representative feature having a condition that satisfies a condition derived from the selection threshold calculated in the selection threshold setting step, from among the classes derived in the database feature classification step, a distance from the feature derived in the query feature extraction step.

17. The signal retrieval method according to claim 12, comprising:

a database feature classification step of classifying, based on a predetermined distance, each feature derived by repeating the database feature extraction step while shifting the observation window, and determining a representative feature of the classification;

a selection threshold setting process of calculating a selection threshold for a distance defined in the database feature classification process from a predetermined search threshold; and

and a database feature selection step of selecting a feature included in a classification of a representative feature having a condition that satisfies a condition derived from the selection threshold calculated in the selection threshold setting step, from among the classifications derived in the database feature classification step, a distance from the feature derived in the query feature extraction step.

18. The signal retrieval method of claim 16, wherein the database feature classification process classifies features according to a vector quantization algorithm and uses euclidean distance as a distance scale.

19. The signal retrieval method of claim 7, wherein the feature verification process calculates the distance based on any one of a manhattan distance or a euclidean distance.

20. A signal retrieval method according to claim 13 wherein the database projection distance calculation procedure calculates the distance based on either one of manhattan distance or euclidean distance.

21. The signal retrieval method of claim 15, wherein the distance recalculation process calculates the distance based on either one of a manhattan distance or a euclidean distance.

22. The signal retrieval method according to claim 7, wherein the query feature extraction step and the database feature extraction step classify features in accordance with a predetermined method, create a histogram as a frequency distribution table for each classification, and output the histogram as a new feature.

23. The signal retrieval method according to claim 7, wherein,

comprises the following components: a jump amplitude calculation step of calculating a jump amplitude of the gaze window based on the distance calculated in the distance correction step and moving the gaze window by the jump amplitude,

while the watching window is staggered, the processing of the characteristic checking process, the distance correcting process, the signal detection judging process and the jump amplitude calculating process is repeated, the distance between the inquiry signal and a plurality of places of the database signal is calculated, and whether the inquiry signal exists in the place of the database signal is determined.

24. A signal search method for searching out a portion lower than a threshold value preset for a distance from a query signal as a target from a database signal registered in advance, comprising:

a query feature extraction process for deriving features from the query signal;

setting a watching window for a database signal, and deriving a database characteristic extraction process of the characteristic from the signal in the watching window;

a database feature classification step of classifying, based on a predetermined distance, each feature derived by repeating the database feature extraction step while shifting the gaze window, and determining a representative feature of the classification;

a selection threshold setting process of calculating a selection threshold for a distance defined in the database feature classification process from a predetermined search threshold;

a database feature selection step of selecting a feature included in a classification of a representative feature having a condition satisfying a condition derived from the selection threshold calculated in the selection threshold setting step, from among the classifications derived in the database feature classification step, a distance from the feature derived in the query feature extraction step;

a segment extraction step of dividing a feature series derived by repeating the database feature extraction step while shifting the focus window, and extracting segments as a partial series;

a compressed image determining step of determining an image for calculating a feature having a dimension lower than the feature dimension from each of the segments obtained in the segment extracting step;

a database feature compression step of calculating features lower than the feature dimension corresponding to the segments obtained in the segment extraction step, based on the map obtained in the compressed map determination step;

a query feature compression process of calculating features lower than the feature dimension corresponding to the features obtained in the query feature extraction process, based on the map obtained in the compressed map determination process;

a feature check process of calculating a distance between a series of compressed features derived in the database feature compression process and compressed features derived in the query feature extraction process; and

a signal detection determination process of determining whether or not the query signal exists at the location of the database signal by comparing the distance calculated in the feature checking process and a search threshold value as a threshold value corresponding to the distance,

the distance between the inquiry signal and a plurality of locations of the database signal is calculated by repeating the processing from the characteristic checking process to the signal detection judging process while shifting the observation window, and whether or not the inquiry signal exists at the location of the database signal is determined.

25. The signal retrieval method as set forth in claim 24, wherein:

a distance recalculation step of calculating a distance between the feature derived in the query feature extraction step and the feature series derived in the database feature extraction step for the location of the database signal determined to be the query signal in the signal detection determination step; and

a signal detection re-decision process for re-deciding whether the query signal is present at the location of the database signal by comparing the distance derived in the distance re-calculation process with a search threshold,

while the watching window is staggered, the processing of the characteristic checking process, the signal detection judging process, the distance recalculating process and the signal detection judging process is repeated, the distance between the inquiry signal and a plurality of places of the database signal is calculated, and whether the inquiry signal exists in the place of the database signal is determined.

26. The signal retrieval method according to claim 24,

comprises the following components: calculating a jump amplitude of the gaze window based on the distance calculated in the feature checking and moving the gaze window by the jump amplitude,

while the fixation window is staggered, the processing of the characteristic checking process, the signal detection judging process and the jumping amplitude calculating process is repeated, the distance between the data base signal and the inquiry signal is calculated for several positions of the data base signal, and whether the inquiry signal exists in the position of the data base signal is determined.

27. A signal search device for searching out a portion lower than a threshold value preset for a distance from a query signal as a target from a database signal registered in advance, the signal search device comprising:

a query feature extraction section that derives a feature from the query signal;

a database feature extraction part for setting a gazing window for a database signal and deriving features from the signal in the gazing window;

a database feature extracting unit that extracts a feature series derived by repeating the processing of the database feature extracting unit while shifting a focus window;

a database inter-feature thinning unit that extracts a representative feature from the classified feature series obtained by the database feature classification unit and derives a representative feature series composed of a smaller number of features;

a feature region extracting unit that derives a region in which a feature included in the classification derived by the database feature classifying unit exists;

a feature matching unit configured to calculate a distance between the feature series derived by the query feature extraction unit and a representative feature series derived by the database feature thinning unit;

a distance correcting unit for correcting the distance calculated by the feature checking unit using the region derived by the feature region extracting unit; and

a signal detection determination section that determines whether or not the query signal is present at the location of the database signal by comparing the corrected distance derived in the distance correction section with a search threshold value that is a threshold value corresponding to the distance,

the processing from the feature check means to the signal detection/determination means is repeated while shifting the observation window, and the distance to the query signal is calculated for several locations of the database signal, thereby determining whether or not the query signal is present at the location of the database signal.

28. The signal retrieval device according to claim 27, comprising:

a segment extraction unit that divides a feature series derived by repeating the processing of the database feature extraction unit while shifting a focus window, and extracts segments as a partial series;

a compressed image determining unit that determines an image for calculating a feature having a dimension lower than the feature dimension from each of the segments obtained by the segment extracting unit;

a database feature compression unit that calculates features lower than the feature dimension corresponding to the segments obtained by the segment extraction unit, based on the map obtained by the compressed map determination unit; and

query feature compressing means for calculating a feature having a dimension lower than the feature dimension corresponding to the feature obtained by the query feature extracting means, based on the map obtained by the compressed map determining means,

in the database feature thinning means, the compressed feature series derived by the database feature compressing means is used as a new feature series derived representative feature series, and in the feature checking means, the compressed feature derived by the query feature compressing means is checked as a new feature, and further, the processing from the feature checking means to the signal detection judging means is repeated while shifting the attention window, and the distance to the query signal is calculated for several locations of the database signal, and whether or not the query signal is present at the location of the database signal is determined.

29. The signal retrieval device according to claim 27, comprising:

a distance recalculating unit that calculates a distance between the feature series derived by the query feature extracting unit and the feature series derived by the database feature extracting unit for the location of the database signal determined to be the query signal by the signal detection determining unit; and

a signal detection re-determination section for re-determining whether or not the inquiry signal is present at the position of the data signal by comparing the distance derived in the distance re-calculation section with a search threshold,

the processing of the characteristic checking means, the distance correcting means, the signal detection judging means, the distance recalculating means, and the signal detection re-judging means is repeated while shifting the observation window, and the distance to the query signal is calculated for several points of the database signal, and whether or not the query signal is present at the point of the database signal is determined.

30. The signal retrieval device according to claim 27, comprising:

a database feature classification unit that classifies, based on a predetermined distance, each feature derived by repeating processing by the database feature extraction unit while shifting a gaze window, and determines a representative feature of the classification;

a selection threshold setting section that calculates a selection threshold for a distance defined in the database feature classification section from a predetermined search threshold; and

and a database feature selection unit configured to select a feature included in a class of a representative feature satisfying a condition derived from the selection threshold calculated by the selection threshold setting unit, from distances between the class derived by the database feature classification unit and the feature derived by the query feature extraction unit.

31. The signal retrieval device of claim 27,

the disclosed device is provided with: a jump amplitude calculation section for calculating a jump amplitude of the gaze window based on the distance calculated by the distance correction section and moving the gaze window by the jump amplitude,

the processing of the characteristic checking means, the distance correcting means, the signal detection judging means, and the jump width calculating means is repeated while shifting the observation window, and the distance to the query signal is calculated for several points of the database signal, and it is determined whether or not the query signal is present at the point of the database signal.

32. A signal search device for searching out a portion lower than a threshold preset for a distance from a query signal as a target from a preregistered database signal, comprising:

a query feature extraction section that derives a feature from the query signal;

setting a gazing window for a database signal, and deriving a database characteristic extraction component of the characteristic from the signal in the gazing window;

a database feature classification unit that classifies, based on a predetermined distance, each feature derived by repeating the process of the database feature extraction unit while skipping a gaze window, and determines a representative feature of the classification;

a selection threshold setting section that calculates a selection threshold for a distance defined in the database feature classification section from a predetermined search threshold;

a database feature selection unit that selects features included in a classification of representative features that satisfy a condition derived from the selection threshold calculated by the selection threshold setting unit, from among the classifications derived by the database feature classification unit, a distance from the feature derived by the query feature extraction unit;

a segment extraction unit that divides a feature series derived by repeating the processing of the database feature extraction unit while shifting a focus window, and extracts segments that are partial series;

a compressed image determining unit that determines an image for calculating a feature having a dimension lower than the feature dimension from each segment obtained by the segment extracting unit;

a database feature compression unit that calculates features lower than the feature dimension corresponding to the segments obtained by the segment extraction unit, based on the map obtained by the compressed map determination unit;

a query feature compression unit that calculates a feature having a dimension lower than the feature dimension, which corresponds to the feature obtained by the query feature extraction unit, from the map obtained by the compressed map determination unit;

a feature matching unit that calculates a distance between the series of compressed features derived in the database feature compression unit and the compressed features derived in the query feature extraction unit; and

a signal detection judging means for judging whether or not the query signal is present at the position of the database signal by comparing the distance calculated by the feature checking means with a search threshold value which is a threshold value corresponding to the distance,

the processing from the feature checking means to the signal detection and determination means is repeated while shifting the observation window, and the distance to the query signal is calculated for several points of the database signal, and it is determined whether the query signal is present at the point of the database signal.

33. The signal retrieval device according to claim 32, comprising:

a distance recalculating unit that calculates a distance between the feature series derived by the query feature extracting unit and the feature series derived by the database feature extracting unit for the location of the database signal determined to be the query signal by the signal detection determining unit; and

a signal detection re-determination section for re-determining whether or not the inquiry signal is present at the location of the data signal by comparing the distance derived in the distance re-calculation section with a search threshold,

the processing of the characteristic checking means, the signal detection judging means, the distance recalculating means, and the signal detection re-judging means is repeated while shifting the observation window, and the distance to the inquiry signal is calculated for several locations of the database signal, and it is determined whether or not the inquiry signal is present at the location of the database signal.

34. The signal retrieval device of claim 32,

the disclosed device is provided with: a jump amplitude calculating section for calculating a jump amplitude of the gaze window based on the distance calculated in the feature checking section and moving the gaze window by the jump amplitude,

the processing by the characteristic checking means, the signal detection judging means, and the jump width calculating means is repeated while shifting the observation window, and the distance to the query signal is calculated for several points of the database signal, and it is determined whether or not the query signal exists at the point of the database signal.

Images