JP2006313524A - Data storage device, data retrieval device, data reproduction device, data retrieval program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To associate content data with the address where the content data is stored in data storage. <P>SOLUTION: A memory system 1 comprises a convolution arithmetic section 22 for convoluting a first data row inputted through an input part 10 in a first real number value by function operation based on a chaos dynamical system function, and a controller 20 that associates the first real number value with a data storage position of a storage section 30 and stores the associated data associated with the first data row at the data storage position. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a data storage device that records content data such as image data, audio data, music data, or link information thereof, a data search device and a data playback device that use this data storage device, and a data search program.

  2. Description of the Related Art Conventionally, a data storage device or database for storing data has been used as well as a semiconductor memory and various recording media.

  In particular, a data storage device using a semiconductor memory is widely known. Such a semiconductor memory usually has a memory cell as a storage unit designated by a unique memory address, and a memory space as a vast storage area composed of a set of the memory cells.

  The above-described memory cell or memory space will be described by taking a DRAM (Dynamic Random Access Memory) as a kind of semiconductor memory as an example. The DRAM has therein a number of capacitors corresponding to a predetermined number of bits according to the standard. The DRAM stores charge in each capacitor, thereby storing 1-bit data in each capacitor, and periodically amplifies the charge so that the charge accumulated in each capacitor does not decay with time. Holding. In such a DRAM, each capacitor described above corresponds to a memory cell as a storage unit, and each memory cell and its position in the memory space are designated by a unique number (memory address).

  When pattern recognition processing is performed on input image data and audio data using data stored in such a memory, the image data and audio data to be recognized are a plurality of image data already stored in the memory. And the speech data (known pattern data) are compared to determine which known pattern data matches or approximates the recognition target data.

For example, when pattern recognition processing is performed on image data of a two-dimensional image, the following procedure is usually adopted. First, from the image data of the 2D image to be recognized, the feature quantity that characterizes the pattern included in the 2D image, that is, the information indicating the color and shape of the image, the length and inclination of the line segment included in the image Preferably, a plurality of pieces of information representing the thickness and the like are extracted. Similarly, the above-described feature amount is extracted for each image data stored in the memory as known pattern data. Then, by comparing each feature amount or a combination thereof between the recognition target data and the known pattern data, the known pattern data having the closest feature amount to the recognition target data is identified from among a plurality of known pattern data. The identified known pattern data is used as the recognition result of the recognition target data.
JP 5-144276 (released on June 11, 1993)

  However, in such a conventional memory configuration, when pattern recognition processing is performed using the stored data, there is a limit to increase in efficiency and speed.

  This is because, when storing a data string in the conventional memory as described above, the contents of this data string are sequentially stored in the memory cells. Is irrelevant. That is, when a data string such as image data or audio data is stored in a semiconductor memory, there is no correlation between the memory address and the stored contents of each memory cell.

  Therefore, when performing pattern recognition processing using a conventional memory, in order to know the contents of the known pattern data stored in the memory, all the known pattern data is read by sequentially reading the memory contents from the reading start position. It is necessary to read. Therefore, the amount of data to be read out and the memory space are enormous, so that the processing load on the computer is coupled with each processing such as the extraction of the feature amount group and the comparison thereof, and the identification of the data having the optimum similarity as described above. Is heavy, and this causes a long time for pattern recognition processing. In particular, when the object of pattern recognition has a large amount of information such as image data, the amount of data to be read from the memory and the memory space become enormous, and this problem becomes more serious.

  In addition, in order to improve the recognition accuracy of such pattern recognition processing, the number of pattern data stored in the memory is increased, or the feature amount data extracted from the recognition target data and the pattern data is increased. However, such an increase in data spurs an increase in the processing time of the pattern recognition process.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to provide a data storage device for recording content data such as image data, audio data, music data, or link information thereof, and the data storage device. A data search device, a data reproduction device, and a data search program to be used are realized.

  In order to solve the above-described problem, a data storage device according to the present invention uses a storage unit for storing data and a first data string input from the outside by a function calculation based on a chaotic dynamical system function. A first convolution operation unit that convolves with the first storage unit, and storage control that associates the first real value with a data storage position of the storage unit and stores association data associated with the first data string at the data storage position. It is characterized by providing a part.

  In order to solve the above-described problem, the data storage device according to the present invention convolves the first data string input from the outside with the first real value by function calculation based on the chaotic dynamical system function. A calculation unit, and a storage control unit that associates the first real value with a data storage position of the storage unit and stores association data associated with the first data string at the data storage position. It is said.

In the above configuration, the first data string includes 1-bit or multiple-bit data regardless of the type of content such as image data, moving image data (video data), audio data, music data, or the format or content represented by the data. All data strings expressed in digital information are included.
Here, the type of chaotic dynamical system function is not particularly limited, and a triangular mountain function, a logistic function, a quadratic function, etc. described later can be employed. An arbitrary unimodal function can be selected as the chaotic dynamical system function for the convolution operation of the 1-bit data string.

According to said structure, a 1st data sequence is converted based on the function calculation based on a chaotic dynamical system function by the function of a 1st convolution operation part, and is convolved with a 1st real value. Then, by the function of the storage control unit, the first real value is associated with the data storage position of the storage unit, and the association data associated with the first data string is associated with the first real value. It is stored in the data storage position of the storage unit.

  Here, the association data associated with the first data string includes the memorized information and link information associated with the first data string, the converted data of the first data string, and the first data. It contains the sequence itself.

  According to the above configuration, the first real value corresponding to the data content of the first data string is calculated and associated with the first data string based on the value of the first real value. The data storage position of the storage unit in which the association data is to be stored is specified.

  Accordingly, since the storage position of the association data in the storage unit is determined according to the data content of the first data string, the content of the data stored in the storage unit and the storage position thereof are indirectly associated with each other. It will be a thing. Therefore, there is no need to read an irrelevant storage area, and it is possible to quickly access desired stored data, so that high-speed and accurate data reading can be realized.

  In the data storage device according to the present invention, in the above configuration, it is also preferable that the first real value is greater than 0 and less than 1.

According to the above configuration, in addition to the above-described effects, the handling of the first real value or the correspondence between the first real value and the data storage position of the storage unit is facilitated. Data storage and subsequent data access can be made fast and simple.
In the data storage device according to the present invention, it is preferable that the first data is composed of a plurality of bits in order to solve the above problem.

  According to said structure, in addition to said effect, calculation of the 1st real value based on the 1st data sequence comprised by several bits by the function calculation based on a single or several chaotic dynamical system function thru | or It becomes possible to accurately specify the data storage position based on the value of the first real value.

  In the data storage device according to the present invention, in the above configuration, it is also preferable that the first data string is stored in the storage unit.

  According to the above configuration, in addition to the above-described effects, the first data string is stored in the storage device separately from the association data or as an aspect of the association data, thereby improving convenience during access. Can be made.

  In the data storage device according to the present invention, in the above configuration, the storage unit is preferably a semiconductor memory, and the storage control unit preferably associates the first real value with a memory address of the semiconductor memory. .

  According to the above configuration, it is possible to realize a data storage device that associates the storage contents with the memory address while using a semiconductor memory widely used as a memory device.

  In order to solve the above-described problem, the data search device according to the present invention is a data search device using the data storage device, wherein a function operation based on the chaotic dynamical system function is performed on the second data string input from the outside. A second convolution operation unit that convolves the second real value with the second real value, and associates the second real value with the data storage position of the storage unit, and based on the data stored in the data storage position, the second And a search control unit for specifying data corresponding to the data string.

  According to the above configuration, by the function of the second convolution operation unit, the second data string input from the outside is obtained by the function operation based on the same chaotic dynamical system function as that when the first real value is convolved. Folded to the second real value.

  Then, by the function of the search control unit, the second real value is associated with the data storage position of the storage unit, and the data stored in the data storage position is read out. As described above, in the data storage device, the content of data stored in the storage unit and the storage position thereof are indirectly associated with each other. The read data corresponds to the second data string.

  Therefore, the data corresponding to the second data string can be directly specified without reading all the data to be searched in the memory, so that an extremely fast and accurate data search can be realized.

  The search control unit, when associating the second real value with the data storage position of the storage unit, sets the neighborhood value of the second real value as the data storage position of the storage unit, not the second real value itself. You may associate. According to this configuration, not only the same data string as the second data string but also a data string similar to the second data string can be searched.

  In the data search device according to the present invention, in the above configuration, the first data string and the second data string are preferably image data. According to the data search device of the present invention, high-speed and accurate data search can be realized even when searching for data having a large amount of information such as image data.

  In the data search device according to the present invention, in the above configuration, the first data string and the second data string may include a feature amount indicating a feature extracted from at least a part of the image area of the image data. preferable.

  The feature amount extracted from the image data refers to various values indicating the characteristics or characteristics of the image indicated by the image data. For example, the feature amount can be obtained from output values of various image sensors. Specific examples of such feature quantities include line segment extraction sensors, lightness sensors, color sensors, and HSV (“Hue”, “Saturation”, “Value”) sensor outputs. Can be mentioned. These sensors can operate on the entire image data, or can operate on a part of the image data.

  According to said structure, compared with the case where image data is used as said 1st thru | or 2nd data sequence as it is, the data amount of the 1st thru | or 2nd data sequence can be made small. Therefore, it is possible to realize a higher-speed data search while reducing the processing load of the data search.

  In the data search device according to the present invention, in the configuration described above, the first data sequence and / or the second data sequence includes a plurality of the feature quantities, the first convolution operation unit, and / or the second data sequence. The convolution operation unit of the first data string and / or the function data extracted from the image data of a wider image area in the function calculation based on the chaotic dynamical system function is a later calculation target. Alternatively, it is preferable to convolve the second data string with a real value.

  For example, if the first data string (second data string) is predetermined image data, the feature amount extracted from the entire image region, the feature amount extracted from the quarter image region, 16 minutes Consider a case in which a feature amount extracted from one image area is included. In this case, the first convolution operation unit (second convolution operation unit) is a function operation based on the chaotic dynamical system function, and the feature amount extracted from the image data of a wider image area is a later operation object. Thus, the convolution calculation is executed in the order of the feature amount extracted from the 1 / 16th image region, the feature amount extracted from the 1/4 image region, and the feature amount extracted from the entire image region. .

  Various methods are conceivable as specific methods for causing the feature amount extracted from the image data of a wider image area to be a later calculation target. As an example, the first convolution operation unit (second convolution operation unit) performs the above convolution operation in a hierarchical manner (layer) in order from the feature amount extracted from the image data of a narrower image region. A method to do this is conceivable. Further, in the first data string (second data string), a plurality of feature quantities are arranged so that the feature quantities extracted from the image data of a wider image area are located in front, and these data A convolution operation may be performed from the back of the column.

  According to the above configuration, in the function calculation based on the chaotic dynamical system function, the feature amount extracted from the image data of a wider image area is processed later, so that it is strongly reflected by the convolution value. On the other hand, it is considered that the feature amount extracted from the image data of a wider image area more faithfully reflects the characteristics or characteristics of the image as the entire image data.

  Thereby, in addition to the above-described effects, it is possible to improve the accuracy when the image data is convoluted with real values, thereby realizing more accurate image data search.

  In the data search device according to the present invention, in the configuration described above, the first data sequence and / or the second data sequence may be obtained by encoding at least a part of image data as the feature amount. The first convolution operation unit and / or the second convolution operation unit includes a frequency component of the order, and a lower frequency component is a later operation object in the function operation based on the chaotic dynamical system function. Thus, it is preferable to convolve the first data sequence and / or the second data sequence with real values.

  In the above configuration, specific examples of the image data encoding process include affine transformations such as Fourier transformation and cosine transformation, and various orthogonal transformations. In these conversions, for example, the zero-order component is the direct current component, that is, the average brightness of the image, the primary component is the fundamental frequency component, the second-order component is the double period component, and so on. Ingredients are obtained.

  In the above configuration, the first convolution operation unit (second convolution operation unit) is a real value so that a lower frequency component becomes a later operation object in the function operation based on the chaotic dynamical system function. Execute the process of convolution.

  As a specific method for causing the lower frequency component to be a later calculation target, for example, in the first data string (second data string), the lower frequency component is positioned in front. In addition, there is a method of arranging frequency components of respective orders and executing a convolution operation from the back side of these data strings.

  According to said structure, in a function calculation based on the said chaotic dynamical system function, since a lower frequency component is processed later, it is reflected more strongly by a convolution value. In general, it is known that a lower-order frequency component gives a stronger impression to an observer of an image due to human visual characteristics.

  Thereby, in addition to the above-described effects, it is possible to realize a search for accurate image data suitable for human visual characteristics.

  In the data search device according to the present invention, in the above configuration, it is preferable that the same arrangement is repeated in at least one of the first data string and the second data string.

  According to the function calculation based on the chaotic dynamical system function of the present invention, an inherent real value can be made to converge with respect to the input data string. However, the data string has a finite number of digits, and the smaller the number of digits in the data string, the more uneven the convergence occurs. This is because the smaller the number of digits in the input data string, the stronger the output of the chaotic dynamical system function is affected by the initial value.

  According to said structure, the data sequence used as the object of the function calculation based on the said chaotic dynamical system function can increase the number of digits, without changing the information by repetition of the same arrangement | sequence. Therefore, according to the above configuration, it is possible to stabilize the real value obtained by the function calculation based on the chaotic dynamical system function regardless of the initial value. For example, even when the initial values used for the function calculation of the first data string and the function calculation of the second data string are different, the real values obtained by these function calculations can be stabilized.

  As a result, in addition to the above-described effects, it is possible to improve the accuracy of data retrieval performed based on the real value, regardless of the initial value used for the function calculation based on the chaotic dynamical system function.

  In the above configuration, the data search device according to the present invention includes a keyword input unit that receives an input of a keyword, a dictionary image data storage unit that stores a plurality of dictionary image data associated with character string information indicating image contents, A dictionary image for specifying one dictionary image data associated with character string information similar or coincident with the keyword received by the keyword input unit from among a plurality of dictionary image data stored in the dictionary image data storage unit The second convolution operation unit convolves one dictionary image data specified by the dictionary image specifying unit with the second real value as the second data string, and the search The control unit associates the second real value with the data storage position of the storage unit, and based on the data stored in the data storage position, the key. It is preferable to identify the image data corresponding to the word.

  According to the above configuration, one dictionary associated with character string information similar or coincident with the keyword received by the keyword input unit from among the plurality of dictionary image data stored in the dictionary image data storage unit Image data is identified. Then, the specified dictionary image data is convolved with the second real value as the second data string in the second convolution operation unit, and based on the second real value in the search control unit, Image data corresponding to the keyword is specified.

  Here, the mode of the character string information associated with the dictionary image data is not particularly limited. As specific examples thereof, caption information associated with the dictionary image data, file name information, and a header area of dictionary image data (for example, JPEG data). Character string information included in the.

  According to the above configuration, after the first search specifying the dictionary image data based on the keyword is performed, the second search specifying the similar image is performed based on the specified dictionary image data. . Therefore, even when the image data to be searched in the second search is not associated with the character string information indicating the image contents, the image data related to the keyword is specified in the second search. be able to.

  Therefore, even if meta information such as text information is not added to the image data to be searched, it is necessary to prepare a dictionary image with meta information such as text information as a keyword. The search of the image data based on the above can be realized.

  As a result, even if the number of image data to be finally searched is large, it is based on relatively small dictionary image data and associated information without adding meta information such as text information to these image data. Since an image search is possible, an image search using a keyword can be realized at high speed.

  A data search device according to the present invention is a data reproducing device using the data storage device, and performs the inverse function operation based on the inverse function of the chaotic dynamical system function on the first real value, thereby It is characterized by comprising a data reproduction unit for reproducing one data string.

  According to the above configuration, the function of the data reproducing unit performs the inverse function calculation based on the inverse function of the chaotic dynamical system function based on the first real value, and the first data string is obtained by the inverse function calculation. Can be played. That is, by making the most of the characteristic of the function operation based on the chaotic dynamical system function that can be completely converted to the first real value without losing the information amount of the first data string, The first data string can be accurately reproduced based only on the real value.

  A data search program according to the present invention causes a computer to function as a first convolution operation unit, a storage control unit, a second convolution operation unit, and a search control unit in the data search device.

  According to said structure, the same effect as the said data search device can be acquired by making a computer read and execute a data search program.

  According to the data storage device of the present invention, as described above, it is possible to access desired stored data without having to read out an irrelevant storage area, thereby realizing high-speed and accurate data reading. be able to.

  In addition, according to the data search device of the present invention, as described above, the data corresponding to the search data can be directly specified without reading all the search target data in the memory. It is possible to achieve an effect that an accurate data search can be realized.

(Embodiment 1)
An embodiment of the present invention is described below with reference to the drawings.

[1. System configuration〕
Based on FIG. 1, the configuration of a memory system using an embodiment of a data storage device according to the present invention will be described.

  The memory system 1 includes an input unit 10 serving as an interface for inputting a data string (first data string, second data string) input from the outside, and a computer that controls operations of the entire memory system 1 or data input / output operations. The control unit 20 is composed of a storage unit 30 composed of a semiconductor memory.

  The control unit 20 (storage control unit) converts the data string input through the input unit 10 based on a chaotic dynamical system function (described later), and the data based on the conversion table 21. A convolution operation unit 22 (first convolution operation unit and second convolution operation unit) that performs an operation of convolving a column into a real value, and an R / for executing data write control and read control on the storage unit 30 And a W control unit 23.

  The storage unit 30 also includes a storage area that becomes a real address memory 31 (described later).

[2. (Operation flow)
The memory system 1 is a 1-bit or multi-bit digital data regardless of the type of image data, moving image data (video data), audio data, music data, etc. input through the input unit 10 or the type of content or the content represented by the data. All data sequences represented by information can be handled, various operations can be executed, and stored in the storage unit 30.

In the present embodiment, the memory system 1 is assumed to handle image data composed of 1-bit data (0 or 1 information) two-dimensionally arranged in M rows and N columns as shown in FIG. explain. As shown in FIG. 2, image data composed of a two-dimensional array has pixel values p _i corresponding to coordinates on the two-dimensional array, for example, (1, 1) (1, 2) (1, N) · .. (M, 1) (M, 2) ... (M, N) in the order of (p ₁ , p ₂ , ..., p _n ) as a one-dimensional array data string Can be handled (n = N × M).

  Note that the data strings to be handled as image data and audio data are determined by various file formats such as BMP, TIFF, JPEG, MPEG, and the like. May be handled.

The memory system 1 transmits the data string input through the input unit 10 to the control unit 20 as the state of such one-dimensional array of digital data. The data string (p ₁ , p ₂ ,..., P _n ) is convolved with a real value based on the conversion table 21 by the operation of the convolution operation unit 22.

[2-1. (Convolution operation)
Here, the details of the calculation operation by the convolution calculation unit 22 will be described.

  In the conversion table 21, a convolution function (chaos dynamic system function) that expresses a chaotic dynamic system is registered in advance.

  A representative example of this convolution function is the triangular mountain function shown in FIG. 3 can be mathematically expressed as y = 0.5x (0 ≦ y ≦ 0.5) and y = −0.5x + 1 (0.5 <y ≦ 1). As shown in FIG. 3, this triangular mountain function is a bivalent function in which two y coordinates correspond to a predetermined x coordinate.

  However, the chaotic dynamical system function to be registered in the conversion table 21 is not limited to the above triangular mountain function, and follows a chaotic dynamical system such as a logistic function or a quadratic function y = 4x (1-x). Any unimodal function can be used. Alternatively, a plurality of chaotic dynamical system functions may be registered in advance in the conversion table 21, and one of the chaotic dynamical system functions may be selected and used for the calculation at the time of convolution calculation described later.

The convolution operation unit 22 first determines an arbitrary initial value x ₀ that satisfies 0 <x ₀ <1 (x ₀ ≠ 0.5) by a random number generation operation or the like. As the value of the y coordinate corresponding to the x coordinate x = x ₀ on the above triangular mountain function (FIG. 3) with respect to the initial value x ₀ , two different values y _large and y _small sandwiching y = 0.5 Exists (y _large > y _small ).

Then, the convolution operation unit 22 determines whether each value of the data string (p ₁ , p ₂ ,..., P _n ) sequentially input through the input unit 10 is 0 or 1 as described above. One of the two y coordinates y _large and y _small (y _large > y _small ) corresponding to the x coordinate x = x ₀ of the triangular mountain function is selected.

Specifically, the convolution operation unit 22 first focuses on the first data value p ₁ of the data string (first data string, second data string), and if the p ₁ is 0, the two y Among the values, y _small smaller than y = 0.5 is selected as the initial value y ₀ , while if p ₁ is 1, y _large larger than y = 0.5 among the two y values is the initial value y _0. Select as.

Next, the convolution unit 22 sets the selected initial value y ₀ as a new x coordinate x = x _{1 and} sets two y coordinates y _large , y _small (y _large > y _small ) corresponding to the x coordinate x _1. Ask. Then, paying attention to the _second data value p ₂ of the data string, if p ₂ is 0, y _small smaller than y = 0.5 is selected as y ₁ out of the two y values, while p ₂ If y is 1, y _large larger than y = 0.5 is selected as y ₁ out of the two y values.

Then, the convolution calculator 22, similarly select _{y 2} a _{y 1} selected as a new x-coordinate x = _{x 2,} such sequential operations, the data sequence _{(p 1,} p _2, ··· , P _n ) to obtain a convolution value x _n (first real value) by repeating n times while paying attention to all data values sequentially.

The convolution value x _n consisting of a single real value obtained in this way completely reflects the data string (p ₁ , p ₂ ,..., P _n ), which will be described later. Thus, based on this convolution value x _n , the data string (p ₁ , p ₂ ,..., P _n ) can be completely reproduced.

Next, the control unit 20 treats the convolution value _xn consisting of the above-described real value as a memory address of the storage unit 30, whereby data is transferred to the real number address memory 31 of the storage unit 30 via the R / W control unit 23. The operation for storing the will be described.

The control unit 20 treats a convolution value x _n composed of a single real value that completely reproduces the data string (p ₁ , p ₂ ,..., P _n ) as memory address information of the storage unit 30. Thus, the data is stored in the real address memory 31 of the storage unit 30 via the R / W control unit 23. That is, the control unit 20 associates the convolution value _xn composed of the above real value with the memory address (data storage position) of the storage unit 30 and associates the association data associated with the data string (first data string). Store in the memory address.

First, a method in which the control unit 20 generates an address cell on the memory will be described. Hereinafter, a convolution value x _n made up of a real value is referred to as “real number address x _n ” in the sense that it is handled as address information in the storage unit 30.

As shown in FIG. 4, the table of the storage unit 30 corresponds to inscriptions (symbols), data strings (p ₁ , p ₂ ,..., _Pn ) associated with the inscriptions as patterns, and content data. The real number address _{xn to} be stored is stored. All the information in FIG. 4 may be stored in the table format. However, as will be described later with reference to FIGS. 5 and 6, if the real address _xn is stored, the pattern data is not necessarily stored in the table format. Therefore, at least the real number address _xn and the memorized data linked to the real number address _xn may be stored in the storage unit 30.

The control unit 20 constructs a storage structure of the real number address memory 31 of the storage unit 30 based on the value of the real number address _xn .

  As shown in FIG. 5, a real address memory can be constructed by expressing real values in units of cells. By writing the in-cell information of FIG. 5, a cell tree structure is generated according to the real digit position as shown in FIG. 6, and a memory structure is constructed. In such a memory structure, pattern data are arranged on real linear addresses from 0 to 1, so that, as will be described later, in data retrieval, data that matches or approximates data to be retrieved is retrieved at high speed. In some cases, it is effective.

As shown in FIG. 5, the real number address memory 31 includes the cell number, the cell number of the upper digit (digit position m-1), the value of the real number address _xn in the upper digit (digit position m-1), and the real number indicated by this cell. The cell structure includes a digit position m of the address _xn (10- ^m digit corresponds to the digit position m), numeric values 0 to 9 in the cell, and cell numbers of the following digits (digit position m + 1).

  An example of a configuration for searching real values at high speed will be shown. This configuration is a storage configuration that is effective not only when searching for data that matches the search target data but also when searching for data that approximates the search target data in the data search described later.

  In FIG. 5, cell number 0 is called a top cell, and a real number line from 0 to 1 is divided into 10 equal parts and indicates the first digit after the decimal point. Below each value of 0.0, 0.1, 0.2,..., 0.9, a cell having a second decimal point is generated and linked. In this example, the link is made by cell number. One number line in FIG. 6 corresponds to one cell. FIG. 6 shows an example of 0.32857, but one cell is sequentially generated for each digit and linked.

The real number address memory 31 is constructed by repeatedly generating the cell structure sequentially in accordance with the number of digits of the real number address _xn .

Based on FIG. 6, the generation of the cell structure will be described. In the figure, cells corresponding to ^{10 -2} The value of the digit 0-9 for each value of the ^{10 -1} digit is one generation. A maximum of 10 cells are generated. Then, one cell corresponding to the numeric value 0-9 of the 10 ⁻³ digit is generated for one numeric value of the 10 ⁻² digit. In this way, a maximum of 10 cells are generated from one cell for the lower digits. In this manner, it is shown that address cells are generated one after another according to the number of digits of the real number address _xn .

The cell generation will be described in detail as follows. For example, values of the real address _{x n} is assumed to be 0.32857. In this case, the value of the first digit after the decimal point in the real number address _xn is 3.

  First, a top cell of the first decimal place with cell number 0 is generated. As shown in FIG. 5, a Continue mark is written at the value 3 in the cell. Then, the cell number A in which the lower digit number 2 is remembered is written. The written cell number A is a cell number A generated by storing the numerical value 2 of the second decimal place of the lower digit. Next, the cell number B for storing the numerical value 8 of the third decimal place is written at the same time as the continue mark is written at the value 2 in the cell of the cell number A. The cell number B stores the numerical value 8 of the third decimal place.

By repeating this procedure, it is possible to generate as many cells as necessary regardless of how many digits the real address _xn consists of. When the number of digits (accuracy) of the real number address _xn is increased by one digit, the number of cells to be generated is about ten times. In this specification, a cell generation unit corresponding to each digit of the real address _xn is referred to as an address cell. By constructing the real address memory 31 in this way, it becomes clear which address cell stores the numerical value corresponding to which digit of the real address _xn .

  In this way, the control unit 20 secures an area of the real address memory 31 on the storage area of the storage unit 30 via the R / W control unit 23. Note that the address pointer when the area of the real address memory 31 is secured on the storage area of the storage unit 30 is obtained by multiplying the latest generated cell number by the size [M (word)] of each cell.

  Then, the control unit 20 writes an end mark (End mark) in the cell corresponding to the last digit of the address cell in the real address memory 31, while the continuation mark ( (Cont mark) is written. In each cell, link information to a table storing content data associated with each pattern data is entered. However, when link information has already been generated in the memory cell, additional link information is added only when the link information of the destination is different. Moreover, you may fill in the link information of several memorization with respect to the same address cell. In addition, when reading the real address memory 31 with rough accuracy, there is a case where the link destination of the memorization does not exist for the obtained real number address. It is preferable to refer to the address cell corresponding to the digit position.

In the memory system 1, the real number address x _n obtained by the conversion process of the data string (p ₁ , p ₂ ,..., P _n ) that is the content data is associated with the data string as the address information of the storage unit 30. Since the associated data is stored, a content addressable memory in which the storage contents and the addresses in the storage unit 30 correspond to each other is configured. The memory mechanism of the human cranial nervous system is also considered to have such a content addressable memory structure.

FIG. 7 is an explanatory diagram showing an operation when the memory system 1 reads the real address memory 31 based on the pattern data input through the input unit 10. In this case, the memory system 1 generates a real number address x _n of x _n = 0 to 1 based on the pattern data by the function of the control unit 20, and sets the address of the real number address memory 31 based on the real number address x _n. Point. The control unit 20 can know where the writing trace and the link destination of the recording code stored in advance are stored from the storage area of the pointing destination.

[2-2. Pattern recognition method)
Next, a pattern recognition method for searching for data stored in the memory system 1 (data search device) will be described with reference to FIG.

First, with respect to the pattern data to be searched, feature quantities are extracted by a plurality of feature extraction sensors, and the real addresses Y _n are obtained by treating the feature quantities as pattern data (multidimensional data) input through the input unit 10. generated in advance to build a real address memory 31 on the basis of the real address Y _n.

Thereafter, feature data is extracted from the pattern data to be recognized by the same feature extraction sensor as described above, and this feature value is treated as pattern data (multidimensional data) input through the input unit 10 to obtain a real address Y _n 'is generated. Then, _'to access real address memory 31 on the basis of said real address Y _n' The real address Y _n if they match the real address Y _n which is already stored in the real address memory, real address memory 31 by reading the memorization based on the link information stored in the region corresponding to the real address Y _n in, is employed as the recognition result of the recognition target pattern. On the other hand, if the real address Y _n ′ does not match the real address Y _n already stored in the real address memory, the stored contents of the real address memory 31 corresponding to the real address closest to the real address Y ′ are stored. By reading, the link information or inscription is specified and adopted as the recognition result of the recognition target pattern.

  That is, the memory system 1 convolves a data string input from the outside with a second real value by a function operation based on a chaotic dynamical system function, and associates the second real value with a data storage position in the storage unit 30. Based on the data stored in the data storage position, the data corresponding to the second data string is specified.

  As described above, the storage unit 30 and the real number address memory 31 in the memory system 1 are content addressable memories in which content data and addresses correspond to each other, so that known pattern data having an optimum similarity is specified for pattern recognition. Even in this case, there is no need to search in a huge memory space. Therefore, according to the memory system 1, in order to improve recognition accuracy, even when a large amount of data is recorded as known pattern data, the known pattern data having a high similarity to the recognition target pattern is directly used. Can be identified. As a result, high-speed and accurate search or pattern recognition is possible. In particular, the memory system 1 can be suitably used for recognizing a large amount of data such as image data.

  When extracting feature amounts from a two-dimensional image, as shown in FIG. 9, the target image data is divided into a plurality of field regions (fields 1 to 4 in FIG. 9), and the feature amounts in each region are extracted. If the configuration for extraction is adopted, more accurate feature amount extraction or pattern recognition becomes possible. Further, as shown in the figure, the visual field 1 (the upper left region in FIG. 9) is further divided into the visual fields 1-1 to 1-4, and first, the feature values of the respective visual fields 1 to 4 are extracted. Furthermore, if the feature amount is extracted for each of the visual fields 1-1 to 1-4, feature amount extraction or pattern recognition that more accurately reflects the local feature amount in the target image can be performed.

  In the above description, the memory system 1 has exemplified the two-dimensional image data including 1-bit data (0 or 1) of M rows and N columns as a data column to be written and read. The target digital data is not limited to two-dimensional image data.

  In addition, the memory system 1 can write and read continuous data along a time series such as audio data instead of discrete data. When writing and reading data continuously in chronological order, an address corresponding to this real address is obtained sequentially for each predetermined amount of data (number of samples) in the order of data input / output. A cell will be generated.

[2-3. Application to video data archive system)
Next, the configuration of a video data archive system to which the above-described pattern recognition method (data search method) is applied will be described with reference to FIG. The configuration of the video data archiving system shown in FIG. 10 is a recording medium such as a CD-ROM or DVD-ROM for supplying a large amount of video data to the system as content data and its reading unit, and video as search data. Extraction sensor group for extracting image feature quantities of input data for inputting thumbnail data (or cut data) such as images and images to the system, video data supplied by the reading unit, and search data input from the input unit (Cut detection sensor, voice sensor, color detection sensor, position detection sensor, shape detection sensor, motion detection sensor, texture sensor, face detection sensor for detecting a face image based on skin color information), each of these feature extraction sensors Converts the feature quantity extracted by using a function of chaotic dynamical system and real number Arithmetic unit convolution seek dresses, and a storage unit.

  In addition, the storage unit includes a video data archive area for storing the video data supplied by the reading unit as a database together with time stamp information in a WM (Windows Media) format (registered trademark), MPEG2 format, or the like, and a convolution operation unit. Each storage area of the real address memory constructed based on the obtained real number address is provided. As shown in FIG. 10, the storage unit may store the feature amount extracted by the feature extraction sensor.

  Next, the operation of the above video archive system will be described.

  The video or image data to be searched in advance is read from the recording medium and its reading unit, and the read video or image data is stored in the video data archive with time stamp information added thereto, while each feature extraction sensor Enter the group as well. Each feature extraction sensor group extracts feature amounts of input video or image data based on the respective detection characteristics, and generates the above-described real address based on these feature amounts. At this time, the real number address may be generated independently based on the feature amount of each feature extraction sensor group. However, in order to more accurately associate the generated real number address with video or image data, the above feature extraction sensor is used. It is preferable to obtain a feature amount in a state where the obtained feature amounts are appropriately combined, and generate a real number address based on the combined feature amount. Then, based on the generated real address value, an address cell of the real address memory is generated.

At the time of data search, first, thumbnail data (or cut data) such as video and images is input from the input unit as search data, and the input search data is based on the same combination as at the time of writing described above. Input to the extraction sensor group. Each feature extraction sensor group extracts feature amounts of input search data based on the respective detection characteristics, and generates a real number address based on these feature amounts. Then, based on the value of the generated real number address, the storage contents of the real number address memory are read, and the pattern data to be read from the video data archive is specified. Here, the content data stored in the video data archive and the link information stored in the real address memory are linked via the time stamp information.

  According to the above video archive system, the pattern data can be directly specified based on the feature amount of the search data, and the original video or image data stored in the pattern data and the video data archive can be specified. Are linked through the time stamp information, so that an extremely fast and accurate search can be performed without searching the entire storage area.

[3. (Data reproduction method based on real address)
A further embodiment of the present invention will be described below with reference to the drawings.

In this embodiment, the memory system 1 does not refer to the memorization based on the link information stored in the real address memory 31, but reproduces the data stored directly based on the real address _xn. explain.

The configuration of the memory system 1 according to the present embodiment shown in FIG. 11 is the same as that of FIG. 1 of the first embodiment. However, the playback unit is further configured to play back the data stored directly based on the real address _xn. Forty configurations are added. Similar to the functions registered in the conversion table 21 of the convolution operation unit 22, a triangular mountain function is registered as table data in the reproduction unit 40 as table data. However, the reproduction function is an inverse function in which the x axis and the y axis are exchanged in the function registered in the convolution operation unit 22. The triangular mountain function shown in FIG. 12 is mathematically expressed as y = 2x (0 ≦ x ≦ 0.5) and y = −2x + 2 (0.5 <x ≦ 1).

First, the control unit 20 of the memory system 1 inputs the real number address _xn output from the convolution operation unit 22 to the reproduction unit 40. The reproduction unit 40 obtains the y coordinate of the point (X = x _n ) corresponding to the real number address x _n as the y value y _{0 in the} reproduction function. Then, the y value y ₀ obtained as corresponding to the real address x _n is set as a new X value x _n−1, and the y coordinate of the point (X = x _n−1 ) corresponding to the X coordinate x _n−1. Is determined as the next X value _xn-2 .

By repeating such a recursive operation, the reproducing unit 40, based on a single real address X _n , all of the above-mentioned chaotic dynamical system sequence x _i (i = 1, 2,..., N). Can be generated. The reproducing unit 40 then calculates the values x _i (i = 1, 2,..., N) of the sequence x _i (i = 1, 2,..., N) of the chaotic dynamical system. If x _i ≦ 0.5, a value of 0 is assigned to p _i , while if x _i > 0.5, a value of 1 is assigned to p _i , so that the data sequence before conversion to the above-described number sequence x _i ( p ₁ , p ₂ ,..., p _n ) can be completely reproduced.

However, as described above, the reproducing unit 40 obtains the numerical sequence x _i of the chaotic dynamical system in the _order of x _n , x _n−1 , x _n−2 ,..., X _{1. p i (i = 1,2, ···} , n) when obtaining the data sequence _p 1 before _conversion, p 2, · · ·, order than _{p n} is reversed _{p n, p n-1,} Note that p _n−2 ,..., p ₁ are obtained in this order.

Therefore, when continuous data in time series such as music data is reproduced as content data, the real address x _n is obtained or the real address memory is generated and the pattern data based on the real address memory is reproduced. It is preferable that the data is reproduced in the same order at the time of reproduction, that is, at the time of data writing and at the time of data reproduction.

  In such a case, if the configuration of FIG. 13 is adopted, even if continuous data along the time series is reproduced, the data can be reproduced in the same order as when data is written. In the configuration shown in the figure, the reproduction output from the memory system 1 described above is input to a further memory system 1 ′ and the input is taken out to reproduce the output data from the feature extraction sensor in the same order. Yes.

  That is, in the configuration of FIG. 13, while the reference time (Sensor time) T increases to 1, 2,..., N, a real number address is generated based on the feature amount of the content data output from the feature extraction sensor. The input signal is continuously convolved. When the reference time (Sensor time) T reaches a predetermined time n, the real number address at the time T = n is fetched, and T = n,..., 2, 1 based on the real number address. The content data is reproduced.

  Then, the reproduction data is input to a further memory system, and a real number address based on the input data is generated. When the content data is reproduced based on this real number address, the data is reproduced in the reverse order of the previous reproduction, that is, in the same order as T = 1, 2,..., N when the first data is written. It becomes possible.

  FIG. 14 illustrates the configuration of FIG. 13 as a circuit block. Data of continuous feature values input from the feature extraction sensor (external sensor) is binarized by the binarization circuit and then transmitted to the convolution circuit V1. The feature amount data convolved as a real number address by the convolution circuit V1 is transmitted to the first reproduction circuit every predetermined amount of data. The reproduction data output from the first reproduction circuit is transmitted again to the binarization circuit, and is output from the binarization circuit and then transmitted to the convolution circuit V2. The reproduction data convolved as a real number address by the convolution circuit V2 is transmitted to the second reproduction circuit every predetermined amount of data. Since the reproduction data output from the second reproduction circuit is reproduced in the same order as that inputted from the feature extraction sensor (external sensor), continuous data output should be performed through the last binarization circuit. Is possible. Note that the operation of the first reproduction circuit and the operation of the second reproduction circuit are synchronized by the switching signal SW1 output from the SW signal generator and the switching signal SW2 obtained by delaying the same by a delay circuit. .

15 operates the convolution circuit V1, the real address memory, the first reproduction circuit, the convolution circuit V2, and the second reproduction circuit based on the output signal (S _n ) from the feature extraction sensor in the circuit configuration of FIG. It is a timing chart which shows the timing when making it. The symbolization circuit in FIG. 15 corresponds to the binarization circuit in the circuit configuration of FIG.

  FIG. 16 is a timing chart showing a state of synchronization by the switching signal SW1 and the switching signal SW2 obtained by delaying the switching signal SW1 by a delay circuit. When the switching signal SW1 is generated at the timing when the reference time of the sensor circuit output is increased by a predetermined period and the output of the convolution circuit V1 becomes 100, the first reproduction circuit 100, 99,. Reproduction output is performed in the reverse order, but the switching signal SW2 delayed from the switching signal SW1 is generated by the time considering the operation time of the convolution circuit V1, the reproduction circuit, and the first symbolizing circuit, and convolution is performed. By taking in the output of the circuit V2, the re-inversion of the data order by the second convolution reproduction is performed every correct period.

  Note that the memory system 1 of the present embodiment can record or reproduce content data in association with a single and unique real value, and can control the accuracy of the reproduced data by changing the accuracy of the real value. It can also be applied to the development.

[4. Supplement)
As described above, the embodiment of the present invention has been described by taking the case where the data string as the content data is 1-bit information as an example. However, the data string as the content data is not 1-bit information but multi-bit such as 2 bits or 3 bits. When it is composed of information, the memory system 1 according to the present invention is applied even when the data string that is the content data is composed of multi-bit information by changing the convolution function and the playback function in the memory system 1 as follows. can do.

That is, if the bimodal function shown in FIG. 17 is adopted as a convolution function, convolution of a real number address can be performed based on 2-bit (4-value) content data. The bimodal function in the figure is a tetravalent function in which two triangular mountain functions are arranged. The convolution operation unit 22 determines an arbitrary initial value x ₀ that satisfies 0 <x ₀ <1 (x ₀ ≠ 0.5) by a random number generation operation or the like, as in the case of 1 bit described above. Since the data string sequentially input through the input unit 10 takes the value of the 2-bit data string p _i = 00, 01, 10, 11, the initial value x ₀ depends on the value of the data string p _i. On the other hand, among the y values y ₁ , y ₂ , y ₃ , y ₄ of the above tetravalent function, the initial value y ₀ corresponding to the smallest one is selected. Thereafter, the convolution operation is executed by repeating the same sequential operation as in the case where the data string is composed of 1-bit information.

On the other hand, when data reproduction is performed based on a real number address, if the inverse function of the bimodal function shown in FIG. 18 is adopted as a convolution function, the 2-bit (4-value) content data can be completely reproduced. Can do. The reproduction function shown in the figure is a monovalent function in which the X axis and the Y axis of the above tetravalent function are interchanged. Then, the control unit 20 inputs the real number address _xn output from the convolution operation unit 22 to the reproduction unit 40. The reproduction unit 40 obtains the y coordinate of the point (X = x _n ) corresponding to the real number address x _n as the y value y _{0 in the} reproduction function. Then, the y value y ₀ obtained as corresponding to the real address x _n is set as a new X value x _n−1, and the y coordinate of the point (X = x _n−1 ) corresponding to the X coordinate x _n−1. Is determined as the next X value _xn-2 .

By repeating such a recursive operation, the reproducing unit 40, based on a single real address X _n , all of the above-mentioned chaotic dynamical system sequence x _i (i = 1, 2,..., N). Can be generated. Then, the reproducing unit 40 performs P for each value x _i (i = 1, 2,..., N) of the sequence x _i (i = 1, 2,..., N) of the chaotic dynamical system. _i = 00 (0 <x _i ≦ 0.25), P _i = 01 (0.25 <x _i ≦ 0.5), P _i = 10 (0.5 <x _i ≦ 0.75), P _i By assigning = 11 (0.75 <x _i <1), it is possible to completely reproduce the data string before being converted into the number sequence x _i .

In addition, when convolution or reproduction of content data of k bits (k is an integer of 3 or more), a k-peak function (which becomes a function of 2 ^k valence) in which k triangular mountain functions are arranged may be used. In the simplest case, a trigonometric mountain function arranged by the number of peaks can be used as a convolution function. And as a reproduction | regeneration function, what is necessary is just to use the inverse function which replaced x coordinate and y coordinate of the said convolution function.

  In the above description, the data search device, the pattern recognition process, and the data reproduction device according to the present invention have been described based on the memory system 1, but the scope of the present invention is not limited to this.

  In each processing step of the data search method and the data reproduction method in the memory system 1, a control unit configured by a CPU or the like executes a program stored in a storage means such as a ROM (Read Only Memory) or a RAM. It can be realized by controlling input means such as a keyboard, output means such as a display, or communication means such as an interface circuit. Therefore, the computer having these means can realize various functions and various processes of the apparatus of the present embodiment simply by reading the storage medium storing the program and executing the program. Further, by storing the above program in a removable storage medium, the above various functions and various processes can be realized on an arbitrary computer.

  The storage medium may be a program medium such as a memory (not shown) such as a ROM for processing by a microcomputer, or a program reader may be provided as an external storage device (not shown). It may be a program medium that can be read by inserting a storage medium therein.

  In any case, the stored program is preferably configured to be accessed and executed by the microprocessor. Furthermore, it is preferable that the program is read out, and the read program is downloaded to the program storage area of the microcomputer and the program is executed. It is assumed that the download program is stored in the main device in advance.

  The program medium is a storage medium configured to be separable from the main body, a tape system such as a magnetic tape or a cassette tape, a magnetic disk such as a flexible disk or a hard disk, or a disk such as a CD / MO / MD / DVD. Fixed disk, IC card (including memory card), etc., or semiconductor ROM such as mask ROM, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), flash ROM, etc. In particular, there are storage media that carry programs.

  In addition, if the system configuration is capable of connecting to a communication network including the Internet, the storage medium is preferably a storage medium that fluidly carries the program so as to download the program from the communication network.

  Further, when the program is downloaded from the communication network as described above, it is preferable that the download program is stored in the main device in advance or installed from another storage medium.

(Embodiment 3)
A further embodiment of the present invention will be described below with reference to the drawings.

  With the advent of the multimedia era, opportunities to store large amounts of image and video materials as archives have increased, and it has become important to search for desired data from the vast amount of archived image data. Yes. For example, handwriting recognition, fingerprint recognition, or presenting a picture of a flower to look up the name of a flower compared to a plant pictorial database (image identification problem), surveillance camera data, video taken at home, etc. For example, an image in which a specific image (for example, an image of Mt. Fuji or a specific person image) is taken out from image data for each frame to which no image content information is attached, such as recorded video data (similar image search) ) Effective technology is eagerly desired.

  In such a case, a large amount of video archive data is collated with individual image data, and it is impossible to process in a realistic time when comparing image data as they are. Therefore, a method of extracting and comparing feature amounts of image data is generally used, but feature amounts of all image data are calculated, and these are compared to obtain the image data with the shortest distance (most similar). The requested processing is not appropriate for large-scale data that requires a large amount of processing.

  As a typical search method for solving this problem, a method for specifying target image data based on meta information indicating the image content of the image data has been proposed.

  However, it is known that the search accuracy is generally poor in a method of simply specifying target image data based on meta information. In order to improve the search accuracy in this method, it is conceivable to increase the meta information. However, if the meta information is increased, the time for meta information matching or the search time increases, or the evaluation or accuracy of matching. If it is bad, there is a problem that the search accuracy deteriorates due to local matching.

  Therefore, in the present embodiment, the feature amount of image data (multidimensional data output from the shape sensor or the color sensor) is represented by a bit string, and this bit string is converted into a real value (0) by conversion using the function of the chaotic dynamical system described above. The memory system 1 (data search device) that can search even a large amount of image data at high speed is proposed by convolution to -1) and further using this real value as index data.

[1. System configuration〕
The basic configuration of the memory system 1 (data search apparatus) according to the present embodiment is the same as that shown in FIG. However, in the present embodiment, the real address memory 31 is referred to as a real index memory 31 in order to emphasize the meaning of storing index information. In addition to the real number index memory 31, the storage unit 30 includes a video data archive area 32 that stores the supplied video data as a database together with time stamp information in the WM format or various MPEG formats. The memory system 1 can be mounted on various memory devices such as an IC card.

[2. (Operation flow)
FIG. 19 is a conceptual diagram for explaining the operation when the memory system 1 of the present embodiment stores the video data and its index data in advance in the storage unit 30 prior to the search of the image data.

  The memory system 1 of the present embodiment captures a frame image from still image or moving image data (WM, MPEG, JPEG, etc .; hereinafter collectively referred to as “video data”) input through the input unit 10 (video capture). After that, thumbnail image data having a size (for example, 64 × 64 pixel size) obtained by standardizing the captured image data is created. At this time, it is preferable to create thumbnail image data by cutting out the image data without changing the vertical / horizontal scale ratio of the original image data so that the image is not deformed.

  From this thumbnail image data, a feature extraction sensor group (cut detection sensor, audio sensor, color detection sensor, position detection sensor, shape detection sensor, motion detection sensor, texture sensor, skin color information, etc. for detecting face images) A feature amount (feature vector) of the image is extracted by a face detection sensor. The types and combinations of these feature extraction sensors are arbitrary. For example, in the case of image pattern recognition, the shape detection sensor can be emphasized or the color detection sensor can be emphasized.

  Then, the feature quantity (feature vector) extracted by each of these feature extraction sensors is converted using the above-described chaotic dynamical system function by the calculation operation of the conversion table 21 or the convolution calculation unit 22, and convolved into a real value. Using this real number (real number index) as index data, an address cell of the real number index memory 31 is generated based on this value, and the real number index memory 31 in the storage unit 30 stores the original data stored in the video data archive area 32. Stores link information to video data. When generating address cells of the real number index memory 31, it is preferable to adopt a decimal tree structure so that high-speed search using index data is possible. In addition, when adopting the decimal tree structure, the number of digits operation for assigning one decimal digit to the binary real number 3 bits is appropriately performed.

The memory system 1 transmits the data string input through the input unit 10 to the control unit 20 as the state of such one-dimensional array of digital data. The data string (p ₁ , p ₂ ,..., P _n ) is convolved with a real value based on the conversion table 21 by the operation of the convolution operation unit 22.

  The memory system 1 may independently generate a real number index based on the feature amount of each feature extraction sensor group, or in order to more accurately associate the generated real number index with video or image data, the feature extraction sensor. It is also possible to obtain a feature amount in a state in which the obtained feature amounts are appropriately combined and generate a feature vector or a real number index based on the combined feature amount. The feature vector will be described later in detail.

  FIG. 20 is a block diagram showing the flow of the above processing. That is, in the figure, three feature amounts (S1 to S3) are extracted from image data (ID = 1 to nnnn) to be registered in a memory as a search target, and these feature amounts are real numbers (real number indexes). Has been converted. Based on the real value (real number index), the link information of the image data (ID = 1 to nnnn) is stored in the memory area of the real number index memory 31 for each feature amount (S1 to S3) in a decimal tree structure. It is written in. In the figure, for example, the feature quantity (S1) is the position of the line segment extracted by the feature extraction sensor, the feature quantity (S2) is the brightness extracted by the brightness sensor, and the feature quantity (S3) is the hue extracted by the hue sensor. .

  FIG. 21 is a conceptual diagram illustrating an operation when the memory system 1 of the present embodiment searches for image data based on video data stored in the storage unit 30 and index data thereof.

  First, when a query image or an unknown image is input as search image data from the input unit, the input image data is processed into thumbnail image data of a standardized size (for example, 64 × 64 pixel size). At this time, it is preferable to cut out the image data of the query image and the unknown image without changing the vertical / horizontal scale ratio so that the image is not deformed.

  An image feature amount (feature vector) is extracted from the thumbnail image data by the above-described feature extraction sensor group. The types and combinations of these feature extraction sensors are the same as those used when index data is created (when index data is written).

  Then, the feature quantity (feature vector) extracted by each of these feature extraction sensors is converted using the above-described chaotic dynamical system function by the calculation operation of the conversion table 21 or the convolution calculation unit 22, and convolved into a real value. The address cell closest to the convolved real value (real number index) is read from the real number index memory 31. Then, based on the link information read from the address cell, the original video data stored in the video data archive area 32 is accessed. Here, the content data stored in the video data archive area 32 and the link information stored in the real number index memory 31 are linked via time stamp information.

  FIG. 22 is a block diagram showing the process flow. That is, in the figure, when image data having a predetermined ID is presented as a question image, the three feature amounts (S1 to S3) are extracted, and these feature amounts are converted into real values (real number indexes). ing. Then, based on the real value (real number index), the address cell of the real number index memory 31 is read for each feature amount (S1 to S3). Then, the original video data stored in the video data archive area 32 is accessed based on the read link information.

  Note that the memory system 1 of the present embodiment, when the original video data to be accessed differs depending on the feature quantities (S1 to S3), the search results (candidate data) for all of the accessed original video data for the query image. ).

  According to the memory system 1 of the present embodiment, the pattern data can be directly specified based on the feature amount of the search data, and the original video or the video stored in the pattern data and the video data archive can be specified. Since the image data is linked through the time stamp information, an extremely fast and accurate search is possible without searching the entire storage area. In particular, as illustrated in FIG. 23, the memory system 1 generates (or reads out) address cells of the real number index memory 31 in a decimal tree structure, so that a real value (real number) obtained based on the feature amount is generated. Fast search by index) is possible. This figure shows how address cells are generated or stored when the real value (real number index) obtained based on the feature amount is 0.6238.

  If a real value (real number index) is stored by such a decimal tree structure, even if the user interactively follows the tree structure as compared with the case where the feature value is stored as a bit string by the binary tree structure. There is an advantage that it is easier to understand more intuitively. This is considered to suggest that the storage structure of real values (real number indexes) by the decimal tree structure is similar to the memory regeneration model of biological systems.

[3. (Feature vector)
As described above, the memory system 1 of the present embodiment includes a feature amount indicating a feature extracted from at least a part of an image area of image data, and based on a bit string (feature vector) of the feature amount, A numerical value (real number index) is calculated.

  That is, as illustrated in FIG. 24, a line segment extraction sensor is applied to image data of a character pattern, or HSV (“Hue”, “Saturation”, “Value” ) ") By applying a sensor, a bit string (feature vector) of a feature amount of image data is output, and a real value R is calculated based on the feature vector. As for the HSV sensor, it is preferable to allow a multi-value expression by combining a plurality of sensors (1 bit output). In the HSV sensor, it is also preferable not to detect the hue when the saturation value is detected to be low. This is because when the saturation value is low, the hue detection or determination accuracy tends to be insufficient.

  These sensors can operate on the entire image data, or can operate on a part of the image data.

  According to the memory system 1 of the present embodiment, the amount of data to be processed by the convolution calculation by the function of the chaotic dynamical system is small as compared with the case where the real value is directly calculated from the image data itself. Therefore, it is possible to realize a higher-speed data search while reducing the processing load of the data search.

[3-1. Feature vector array 1]
In the memory system 1 of the present embodiment, the feature vector of the image data to be processed by the convolution operation by the function of the chaotic dynamical system includes a plurality of feature amounts, and in the function calculation based on the chaotic dynamical system function, The feature vector is convolved with a real value so that the feature amount extracted from the image data of a wide image area becomes a later calculation target.

  Specifically, as shown in FIG. 25, the memory system 1 according to the present embodiment divides the image data in three stages of layer 1, layer 2, and layer 3, while stepwise dividing the image data. Is extracted hierarchically. Here, layer 1 is an operation for extracting a feature amount from the entire image data, layer 2 is an operation for extracting a feature amount from a quarter of the image area of the image data, and layer 3 is an operation of 1/16 of the image data. This refers to an operation for extracting a feature amount from an image area. The technique for extracting feature quantities hierarchically is not particularly limited, and various feature extraction sensors such as a line segment extraction sensor and a hue extraction sensor can be used.

  Then, the memory system 1 of the present embodiment uses the feature amount extracted in the layer 3, the feature amount extracted in the layer 2, and the feature amount extracted in the layer 1 in the function calculation based on the chaotic dynamical system function. Convolution operations are executed in the order of. In order to execute the convolution operation in this order, the feature quantity extracted in layer 1, the feature quantity extracted in layer 2, and the feature quantity extracted in layer 3 in a single feature vector May be arranged in this order, and the convolution operation may be executed from the back side of the feature vector.

  Thereby, in the function calculation based on the chaotic dynamical system function, the memory system 1 of the present embodiment is more strongly reflected by the convolution value because the feature amount extracted from the image data of a wider image area is processed later. It will be. The accuracy of convolution of image data with real values can be improved, and more accurate image data search can be realized.

  FIG. 26 is a diagram illustrating an application example of the line segment extraction sensor. In the figure, a vertical sensor H1 is a sensor that detects whether or not the pixels of the image data are arranged in the vertical direction, that is, in a state where a predetermined number of pixels or more are aligned in the vertical direction on the paper. The figure shows a state in which the vertical sensor H1 detects six pixels in the vertical direction. The vertical direction sensor H1 detects whether or not a vertical line segment area exists in the entire area of the image data by sequentially scanning the image data. The vertical sensor H1 turns on the output if the presence ratio of the vertical line segment area is equal to or greater than a predetermined threshold as a result of scanning the image data, while the vertical sensor H1 scans the image data. If the existing ratio of the line segment region is less than a predetermined threshold value, the output is turned off.

  The operations of the horizontal sensor H2 and the oblique sensors H3 and H4 are the same as those of the vertical sensor H1, and whether or not the pixels of the image data are arranged in a state where a predetermined number of pixels or more are arranged in each direction. It is to detect. The horizontal sensor H2 and the oblique sensors H3 and H4 also detect whether or not a line segment area in a predetermined direction exists in the entire area of the image data by sequentially scanning the image data.

  FIG. 27 shows the output of the vertical sensor H1, the horizontal sensor H2, and the oblique sensors H3 and H4 when the image data has a predetermined line segment image. In the same figure, the output of each sensor is represented as “0” while the output is represented as “1”. According to the figure, when the image data has a line segment image in the vertical direction, the output of the vertical direction sensor H1 is turned on, while the image data has a line segment image in the vertical direction. If not, the output of the vertical direction sensor H1, each of the vertical direction sensor H1, the horizontal direction sensor H2, and the oblique sensors H3 and H4, depending on the direction of the line segment image in the image data, such as the off state “0”. It shows how the output changes.

The 4-bit output of these line segment extraction sensors H1 to H4 can be used as feature quantity extraction of image data in each of the aforementioned layers (layer 1, layer 2, layer 3). Here, since the image division is gradually made finer for each hierarchy, each hierarchy is called a layer. Layer 2 is divided into 4 and layer 3 is divided into 16. As a result, the image can be detected more precisely [3-2. Feature vector array 2]
In the memory system 1 of the present embodiment, a feature vector of image data to be processed by a convolution operation using a function of a chaotic dynamical system has a plurality of orders obtained by encoding processing of at least a part of the image data as a feature amount. When the frequency component is included, the feature vector is convolved with the real value so that the lower frequency component becomes a later calculation target in the function calculation based on the chaotic dynamical system function.

  Here, specific examples of the image data encoding process include affine transformation such as Fourier transformation and cosine transformation, and various orthogonal transformations. In these conversions, for example, the zero-order component is the direct current component, that is, the average brightness of the image, the primary component is the fundamental frequency component, the second-order component is the double period component, and so on. Ingredients are obtained.

  The memory system 1 according to the present embodiment performs a real value convolution process so that a lower frequency component becomes a later calculation target in the function calculation based on the chaotic dynamical system function. As a specific method for causing the lower frequency component to be a later calculation target, for example, in the first data string (second data string), the lower frequency component is positioned in front. In addition, there is a method of arranging frequency components of respective orders and executing a convolution operation from the back side of these data strings.

  As described above, according to the memory system 1 of the present embodiment, in the function calculation based on the chaotic dynamical system function, the lower frequency component is processed later, so that it is strongly reflected by the convolution value. In general, it is known that a lower-order frequency component gives a stronger impression to an observer of an image due to human visual characteristics. In other words, the low-order frequency component shows the characteristics of the image when the image is viewed roughly as if the image was passed through a low-pass filter, and the high-order frequency component is fine as if it was passed through a high-pass filter. It shows the characteristics of the image when observed.

  Therefore, according to the memory system 1 of the present embodiment, it is possible to realize an accurate image data search suitable for human visual characteristics.

  Such an arrangement order of frequency components can be adopted in combination with the feature quantity arrangement in the hierarchical order described in the column 3-1. For example, the coefficient (frequency component) obtained by cosine transform of the image data of layer 1 is A (C0, C1, C2, C3), and the coefficient (cosine transform of the image data of layer 2 ( If the frequency component is B1 (C0, C1, C2, C3), B2 (C0, C1, C2, C3), B3 (C0, C1, C2, C3), B4 (C0, C1, C2, C3) , Feature vectors [A (C0, C1, C2, C3), B1 (C0, C1, C2, C3), B2 (C0, C1, C2, C3), B3 (C0, C1, C2, C3), B4 (C0, C1, C2, C3) ...] can be arranged.

  In the above, the method of determining the calculation order of the frequency components based on giving a stronger impression to the observer of the image as the lower frequency components are given in terms of human visual characteristics. The order of function calculations and the arrangement of feature vectors may be determined so that feature values that are more important in terms of visual characteristics are later calculated in the function calculation based on the chaotic dynamical system function.

  For example, [C0 (A), C1 (A), C2 (A), C3 (A), C0 (B1, B2, B3, B4), C1 (B1, B2, B3, B4), C2 (B1, B2 , B3, B4), C3 (B1, B2, B3, B4)], it is also possible to arrange the frequency components from the lower order to the higher order in each layer.

[3-3. Feature vector array 3]
The memory system 1 of the present embodiment repeatedly arranges the same arrangement in the feature vector. For example, if the bit string of the feature vector is V, the array of the same feature quantity [abcdefg... Z] is repeated as V = [abcdefg... Z abcdefg... Z abcdefg. For example, a long bit string of about 32 bits to 64 bits is used.

  The reason why such a periodic repeating structure is adopted for the feature vector array is as follows.

  According to the function calculation based on the chaotic dynamical system function described above, it is possible to make the corresponding real value converge with the input data string. However, the data string that is actually input has a finite number of digits, and the smaller the number of digits in the data string, the more uneven the convergence. This is because the smaller the number of digits in the input data string, the stronger the output of the chaotic dynamical system function is affected by the initial value.

  Therefore, the memory system 1 of the present embodiment increases the number of digits without changing the feature vector information by repeating the same arrangement. As a result, it is possible to stabilize the real value obtained by the function calculation based on the chaotic dynamical system function regardless of the initial value, and to improve the accuracy of the data search performed based on the real value.

  Note that such an arrangement order of frequency components can be adopted in combination with the feature quantity arrangement in the hierarchical order described in the column 3-1, and the frequency component arrangement in the order described in the column 3-2. For example, the feature vectors obtained based on these arrangement orders may be further arranged repeatedly to make a long bit string of about 64 bits, for example.

[4. Keyword search〕
If the configuration of the memory system 1 of the present embodiment is developed, image data related to the keyword can be quickly retrieved from a huge amount of image data to which no meta information is added by inputting text information as a keyword. It becomes possible to search accurately and accurately.

  That is, by adding a configuration to the memory system 1 of the present embodiment, an input unit 10 that receives input of search keywords, and a storage unit 30 that stores a plurality of dictionary image data associated with character string information indicating image contents. The dictionary image database (dictionary image data storage unit) and a plurality of dictionary image data stored in the dictionary image database are associated with character string information similar or coincident with the keyword received by the input unit 10 A dictionary image specifying unit of the control unit 20 for specifying one dictionary image data, and converting the one dictionary image data specified by the dictionary image specifying unit of the control unit 20 into the conversion table 21 or the convolution calculating unit 22. , The control unit 20 convolves the real number (real number index) with the real number (real number in). Correspondence to the data storage position in the storage unit 30 the box), the data storage on the basis of the data stored in the position can be configured to identify image data corresponding to the search keywords.

  According to such a configuration, one dictionary image data associated with character string information similar or coincident with the keyword received by the input unit 10 from among a plurality of dictionary image data stored in the dictionary image database. Is identified. Then, the specified dictionary image data is convolved with a real value (real number index), and the control unit 20 specifies image data corresponding to the search keyword.

  Here, the mode of the character string information associated with the dictionary image data is not particularly limited. As specific examples thereof, caption information associated with the dictionary image data, file name information, and a header area of dictionary image data (for example, JPEG data). Character string information included in the.

  According to the above configuration, after the first search specifying the dictionary image data based on the keyword is performed, the second search specifying the similar image is performed based on the specified dictionary image data. . Therefore, even when the image data to be searched in the second search is not associated with the character string information indicating the image contents, the image data related to the keyword is specified in the second search. be able to.

  Therefore, if you want to search video recording data that has not been given meta information such as text information, just prepare a predetermined amount of dictionary image database and search for keywords such as “Sun” and “Mt. Fuji”. The search of the image data based on it is realizable.

  As a result, even if the number of image data to be finally searched is large, it is based on relatively small dictionary image data and associated information without adding meta information such as text information to these image data. Since an image search is possible, an image search using a keyword can be realized at high speed.

  A dictionary image database that stores a large number of dictionary image data associated with character string information indicating image contents corresponds to a kind of knowledge image collection. Image data that is preferably stored in such a dictionary image database includes a character recognition database (for example, “ETL character database” of the Electronic Technology Research Institute), and video image capture data captured from a television broadcast. , Various pictorial book data (encyclopedia, animal pictorial book, plant pictorial book, mineral picture book, etc.), photographic data that includes famous images such as World Heritage, various commemorative photo data, personal face photo data and biometrics data (fingerprint) , Vein pattern). It is also preferable to adopt a configuration in which such a dictionary image database is automatically generated using a known image search system.

  Referring to FIG. 28, a keyword search procedure by the memory system 1 based on an example of searching Mt. Fuji image data from video recording data is as follows.

  The dictionary image database includes dictionary image data in which Mt. Fuji images are associated with character string information in advance. In addition, it is assumed that the real number index R is obtained beforehand from the feature amount of the dictionary image data in the dictionary image database. On the other hand, with respect to video recording data in the search target database, a real number index R is obtained from the feature amount in advance and registered.

  When the user inputs “Mt. Fuji” as a search keyword, the memory system 1 identifies a dictionary image having character string information similar or identical to “Mt. Fuji” from the dictionary image database. For this specification, a known text search method can be used. If the memory system 1 cannot identify a dictionary image having character string information similar or identical to “Mt. Fuji”, the memory system 1 replies to the user that there is no image to be searched.

  When the memory system 1 specifies a dictionary image having character string information similar or coincident with “Mount Fuji”, the memory system 1 reads a real number index R corresponding to the dictionary image data. Subsequently, based on the real number index R, the index memory of the database to be searched is accessed, and the recording data having the real number index R similar or coincident with the real number index R is specified from the dictionary image database.

  That is, the memory system 1 has the dictionary image data (character string information similar or coincident with “Mt. Fuji”) identified from the dictionary image database by a similar image search method using a convolution operation by a chaotic dynamical system function. And the similar image data is specified from the stored video recording data.

  As described above, by associating the dictionary image database with the search target image database based on the real number index R, it is possible to realize an image search using a keyword at high speed.

  Such a search procedure may be automatically performed by the memory system 1 or may be partly performed interactively according to the user's instruction selection based on the interactive method with the user. Also good.

  The present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the claims, and embodiments obtained by appropriately combining technical means disclosed in different embodiments. Is also included in the technical scope of the present invention.

  The data storage device, data search device, data playback device, and data search program according to the present invention are effective for storing, searching, and recognizing various content data, and can also be used for solving data mining problems. In particular, the data search can be used widely and suitably for image recognition of biometric data (fingerprints, vein patterns), etc. in addition to searching for character images and face images.

  In addition, the data storage device, data search device, data reproduction device, and data search program according to the present invention can be implemented in various memories or database devices, and can be applied to, for example, an IC card or an IC chip thereof.

1 is a block diagram showing a configuration of a memory system using an embodiment of a data storage device according to the present invention. It is explanatory drawing which shows 1 bit data (0 or 1) two-dimensionally arranged by the M row N column which the said memory system handles. It is a graph which shows an example of a convolution function. It is a table which shows the data memorize | stored in a memory | storage part. It is a table which shows the cell structure of real number address memory. It is explanatory drawing regarding the production | generation of the cell structure of real number memory. It is explanatory drawing which shows operation | movement in case a memory system reads the real address memory based on the pattern data input through the input part. It is explanatory drawing regarding the pattern recognition method (data search method) which searches the data memorize | stored in the memory system. It is explanatory drawing which shows the visual field division | segmentation of the image data used as recognition object. It is a figure which shows the structure of the video data archive system to which the data search method based on this invention is applied. FIG. 6 is a block diagram showing a configuration of a memory system using a further embodiment of the data storage device according to the present invention. It is a graph which shows an example of a reproduction | regeneration function. It is a figure which shows the structure of the reproduction | regeneration system which reproduces | regenerates data so that it may become the same order as the time of data writing, even when continuous data along a time series is reproduced | regenerated. 14 is a circuit block showing a circuit configuration of the reproduction system of FIG. 15 is a timing chart showing timing when the circuit configuration of FIG. 14 is operated based on an output signal from a feature extraction sensor in the circuit configuration of FIG. It is a timing chart which shows the mode of the synchronization by each switching signal. It is a graph which shows the further example of a convolution function. It is a graph which shows the further example of a reproduction | regeneration function. It is a conceptual diagram explaining operation | movement when a memory system stores video data and its index data in a memory | storage part. It is a block diagram which shows the flow of a process when a memory system stores video data and its index data in a memory | storage part. It is a conceptual diagram explaining operation | movement when a memory system searches image data based on the video data stored beforehand, and its index data. It is a block diagram which shows the flow of a process when a memory system searches image data based on the video data stored beforehand, and its index data. It is a figure which shows the mode of the production | generation or storage of the address cell of the decimal tree structure in a real number index memory. It is a figure explaining application of the line segment extraction sensor to image data of a character pattern, and application of an HSV sensor to image data of a photograph. It is a figure explaining a mode that a memory system extracts the feature-value hierarchically, dividing | segmenting image data in steps. It is a figure which shows the example of application of a line segment extraction sensor. When image data has a predetermined line segment image, it is a figure which shows each output of a vertical direction sensor, a horizontal direction sensor, and a diagonal sensor. It is a figure which shows the flow of the keyword search by a memory system.

Explanation of symbols

1 Memory system (data storage device, data retrieval device)
10 Input unit 20 Control unit (storage control unit, computer)
21 Conversion Table 22 Convolution Operation Unit 23 R / W Control Unit 30 Storage Unit 31 Real Address Memory 40 Playback Unit

Claims (16)

A storage unit for storing data;
A first convolution operation unit that convolves a first data string input from the outside into a first real value by a function operation based on a chaotic dynamical system function;
A storage control unit that associates the first real value with a data storage position of the storage unit and stores association data associated with the first data string at the data storage position. Storage device.   The data storage device according to claim 1, wherein the chaotic dynamical system function is a unimodal function.   3. The data storage device according to claim 1, wherein the first real value is in a range greater than 0 and less than 1.   4. The data storage device according to claim 1, wherein the first data is composed of a plurality of bits.   5. The data storage device according to claim 1, wherein the first data string is stored in the storage unit. 6. The storage unit is composed of a semiconductor memory,
6. The data storage device according to claim 1, wherein the control unit associates the first real value with a memory address of the semiconductor memory. A data search device using the data storage device according to any one of claims 1 to 6,
A second convolution operation unit that convolutes a second data string input from the outside into a second real value by a function operation based on the chaotic dynamical system function;
A search control unit that associates the second real value with a data storage position of the storage unit and specifies data corresponding to the second data string based on the data stored in the data storage position; A data search apparatus characterized by that.   8. The data search apparatus according to claim 7, wherein the first data string and the second data string are image data.   8. The data search apparatus according to claim 7, wherein the first data string and the second data string include a feature amount indicating a feature extracted from at least a part of an image area of the image data. The first data sequence and / or the second data sequence includes a plurality of the feature quantities,
In the function calculation based on the chaotic dynamical system function, the first convolution calculation unit and / or the second convolution calculation unit is configured such that a feature amount extracted from image data of a wider image region is a later calculation target. The data search apparatus according to claim 9, wherein the first data string and / or the second data string are convolved with real values. The first data sequence and / or the second data sequence includes, as the feature amount, frequency components of a plurality of orders obtained by encoding processing of at least a part of image data,
In the function calculation based on the chaotic dynamical system function, the first convolution calculation unit and / or the second convolution calculation unit is configured so that a lower frequency component is a later calculation target. The data search apparatus according to claim 9 or 10, wherein the column and / or the second data column is convolved with a real value.   12. The data search device according to claim 7, wherein at least one of the first data string and the second data string has the same arrangement repeated. A keyword input unit that accepts keyword input;
A dictionary image data storage unit for storing a plurality of dictionary image data associated with character string information indicating image contents;
A dictionary for specifying one dictionary image data associated with character string information similar or coincident with the keyword received by the keyword input unit from among a plurality of dictionary image data stored in the dictionary image data storage unit An image specifying unit,
The second convolution operation unit convolves one dictionary image data specified by the dictionary image specifying unit into a second real value as the second data string,
The search control unit associates the second real value with a data storage position of the storage unit, and specifies image data corresponding to the keyword based on the data stored in the data storage position. The data search device according to any one of claims 7 to 12. A data reproducing device using the data storage device according to any one of claims 1 to 6,
A data reproduction apparatus comprising: a data reproduction unit that reproduces the first data string by performing an inverse function operation based on an inverse function of the chaotic dynamical system function on the first real value. Computer
In the data search device according to any one of claims 7 to 12,
A data search program that functions as the first convolution operation unit, the storage control unit, the second convolution operation unit, and the search control unit. A first convolution operation unit that convolves a first data string input from the outside into a first real value by a function operation based on a chaotic dynamical system function;
A data storage comprising: a storage control unit that associates the first real value with a data storage position of a storage unit and stores association data associated with the first data string at the data storage position. apparatus.

Images