JP2005148987A - Object identifying method and device, program and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an object identifying method and device for improving the discriminating capability of in-scenery letters, a program and a recording medium. <P>SOLUTION: This object identifying method and device(identifying device 1) are provided to change the position(position under consideration) and size(size under consideration) in a target image, and to segment the image under consideration(region image under consideration), and to extract features(target feature vectors) from the segmented image, and to compress the feature vectors according to the number of times of repetition (target compressed feature vectors), and to calculate a distance(degree of collation) at the time of projecting the target compressed feature vectors to a partial space by using a preliminarily inputted partial space(dictionary 3), and to decide the spatial connectivity of the degree of collation of each object (1), and to detect the maximum peak in the connection region (2), and to register the detected peak as a candidate (3), and to repeatedly execute the processes (1) to (3) the designated number of times. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to an industrial application system, for example, a character recognition system in a landscape, using an image identification technique for identifying what kind of object is reflected in an image.

  Techniques for recognizing characters in a conventional landscape are roughly classified into two.

  One method is a technique (hereinafter referred to as a character cut-out method) in which an area in which characters are reflected from a landscape is cut out, the area is binarized, and is determined by a conventional character recognition technique. As this method, there is a license plate recognition method disclosed in Japanese Patent Laid-Open No. 6-131492.

The other is a technique of preparing a character deformation template and scanning the entire image (hereinafter referred to as a template matching method). Japanese Patent Laid-Open No. 2001-307021 discloses a “pattern row extraction method and license plate recognition method”.
JP-A-6-131492 Japanese Patent Laid-Open No. 2001-307021

  However, in the character cutout method, it is very difficult to cut out the character area itself when the background has a complicated texture or when the characters are attached to each other. In addition, when there is a shadow in the character area, there is a problem that binarization tends to fail. Therefore, the character cutout method has a large restriction on the landscape image that can be applied.

  In the template matching method, since the matching calculation itself requires a lot of time, there is a problem that the number of categories to be determined is limited in order to construct a practical system. In addition, when the geometric shapes such as fonts are very diverse, such as characters in a landscape, it is difficult to prepare a thorough template in advance, which is not practical. Furthermore, there is a problem that, for example, the end of a signboard is identified as “1”, and a place that is not originally a character is identified as a character.

  The present invention has been made in view of such circumstances, and an object of the present invention is to provide an object identification method, an apparatus, a program, and a recording medium that improve the ability to identify characters in a landscape.

  Therefore, the object identification method and apparatus according to the present invention obtain a feature vector for a certain area of an input image, calculate a matching degree by matching with a dictionary, and narrow down candidates using the distribution tendency of the matching degree. As a result, the speed of identification and the reduction of erroneous extraction are realized.

That is, according to the present invention, in the step of identifying an object,
(1) Extracting an image of interest (attention area image) while changing the position of interest (attention position) and size (attention size) in the target image, and extracting features (target feature vectors) from the extracted image, The distance (matching degree) when the target compressed feature vector is projected onto the partial space using the previously input partial space (recognition dictionary) after compressing the feature vector according to the number of repetitions (target compressed feature vector) To determine the spatial connectivity of the matching degree of each object,
(2) Detect the largest peak in the connected region,
(3) Register this detected peak as a candidate,
Then, the steps (1) to (3) are repeatedly executed a specified number of times.

  In the feature extraction process, the horizontal direction differential and the vertical direction differential component of each image are calculated to calculate the direction and strength of the differential, and then the differential direction of each pixel in the defined region of each image Is quantized to a predetermined stage, and then a differential direction histogram is created by accumulating and adding the strength of the differentiation for each stage, and the differential direction histogram is regarded as a vector and the magnitude thereof is normalized to a predetermined value.

  The matching degree with the dictionary does not change abruptly when the attention position is the same, but changes smoothly. The present invention detects the peak among the similar target positions, and compares it with a detailed dictionary at a later stage, so it is similar to a dictionary, but better candidates are reduced in the vicinity. And speeding up the identification process.

  In addition, since the degree of collation with the dictionary is calculated to determine the spatial connection error and the peak is detected in the connection region, it is possible to suppress erroneous detection outside the character region. For example, an error that erroneously detects the end of a sign as “1” appears throughout the end of the sign, but the error candidate position can be narrowed down to one point by this peak detection. In addition, since many candidates can be deleted by performing peak detection in the rough search, the identification process can be speeded up.

  Furthermore, a vector with a high compression rate has a low identification rate, but can obtain a projection distance at a high speed and can exclude many candidates at a high speed. A vector with a low compression rate has a high identification rate but is expensive in calculating the projection distance. According to the present invention, it is possible to combine high-speed processing with poor accuracy and low-speed processing with an emphasis on accuracy by a plurality of compression rates. Thereby, the entire process can be realized at high speed and with high accuracy.

  Moreover, since the normalized differential direction histogram does not require binarization, it is resistant to shadows. Furthermore, since the normalized differential direction histogram does not vary greatly with respect to various shapes of character fonts, it is possible to create a complete recognition dictionary using a typical font in advance.

  The object identification method and apparatus of the present invention can be realized by a program for executing the procedure and means on a computer, and further, the program is recorded on a computer-readable recording medium and provided through a network. It is also possible. Examples of the recording medium include a flexible disk, HDD, MO, ROM, memory card, CD, DVD, and removable disk.

  According to the present invention, since the degree of collation with the dictionary is calculated to determine its spatial connectivity and the peak is detected in the connected region, it is possible to suppress erroneous detection outside the character region. In addition, since many candidates can be deleted by performing peak detection in the rough search, the identification process can be speeded up.

  Furthermore, a vector with a high compression rate has a low identification rate, but a projection distance can be obtained at high speed, so that many candidates can be excluded at high speed. For vectors with a low compression rate, the identification rate is high, but the calculation of the projection distance is expensive. However, according to the present invention, the extraction and identification of a large number of objects existing in the image can be accelerated while maintaining accuracy. it can.

  In addition, the differential direction histogram normalized according to the present invention does not require binarization, and thus is resistant to shadows. Furthermore, since the normalized differential direction histogram does not vary greatly with respect to various shape variations of character fonts, it is possible to create a complete dictionary in advance using typical fonts.

  Embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is a schematic configuration diagram of a character recognition / translation system for landscapes according to the present invention.

  With this system, the user can view translation information of a character based on the photographed character image. However, it is premised on having a character translation dictionary. The system is composed of a recognition device and a translation information storage / retrieval device of the present invention.

  The identification device 1 according to the present invention uses an image input by the user 2 and a recognition dictionary (denoted as the dictionary 3 in FIG. 1) to identify characters photographed in the image.

  The translation device 4 is a means for accumulating character strings and translation information and retrieving translation information from the character strings, and can be constructed with a general database. Therefore, detailed description thereof is omitted in this embodiment.

  In the following embodiment example, a case where three types of characters “den”, “shin”, and “story” are identified will be described. However, the present embodiment is not limited to three types of objects, and can be extended to any number of types.

  FIG. 2 is a schematic configuration diagram showing an embodiment of the invention according to claim 1, and particularly describes the configuration of the identification device 1.

  The identification device 1 includes an input means 10, an all category registration means 11, an iterative control means 12, an image cutout means 13, a feature extraction means 14, a matching degree calculation means 15, a connected region peak detecting means 16, and a connected region peak detecting means. 16 and candidate category updating means 17.

  The input means 10 inputs a target image to be identified.

  All category registration means 11 registers all categories as candidate categories for the determined position and its size.

  The iterative control means 12 repeats the processes of the input number (I) of later-described image cutout means 13, feature extraction means 14, matching degree calculation means 15, connected region peak detection means 16, and candidate category update means 15. Control execution.

  The image cutout unit 13 cuts out a target image (cutout image) while changing a target position (target position) and size (target size) in the target image.

  The feature extraction unit 14 extracts a feature (target feature vector) from the cut-out image.

  The matching level calculation means 15 compares the recognition dictionary (dictionary 3) of each object and the target feature vector according to the number of repetitions (i) created in advance, and calculates a matching level representing the matching level.

  The connected region peak detection means 16 determines the spatial connectivity of the matching degree of each object, and detects the maximum peak of the matching degree in the connected region.

  The candidate category update means 17 registers the detected peak as a candidate.

  An example of category registration by all category registration means 11 will be described.

  FIG. 8 is an example of the target image. FIG. 9 is a diagram showing a point of interest (a gray pixel in the image) in the target image, and shows an example of a “predetermined position”. A rectangular image having a predetermined size is defined around the attention point. FIG. 10 is an example of “determined size”, and size A and size B are shown.

  The initial candidate category is registered in the following format.

“Attention position (x) in image, attention position (y) in image, attention size, candidate category 1, candidate category 2... Candidate category Q”
In this example, since there are three categories, the candidate categories are as follows.

  "Attention position x, attention position y, each attention size, telegraph, communication, talk". That is, it means that the candidate character categories of the rectangular image obtained by cutting out each pixel and each attention size are “den”, “shin”, and “story”.

  Next, the recognition dictionary input to the matching degree calculation unit 15 will be described.

  The recognition dictionary for each object corresponding to the number of repetitions (i) is created as follows, for example. A feature vector having a different resolution or a feature vector with a different degree of compression is generated from a plurality of images of each object, and the feature vector group is created by principal component analysis. A recognition dictionary with a low resolution is associated with the number of repetitions i = 1, and a recognition dictionary with a high resolution is associated with the number of repetitions i = I. As an example of this method, there is the following object recognition method.

  In this method, a plurality of objects are registered using a plurality of images including the objects, and one or more registered objects existing in the image are identified. The process of registering the plurality of objects includes a plurality of processes. A feature extraction process for extracting feature vectors from an image, a principal component analysis process for performing principal component analysis on all the extracted feature vectors, and a compressed principal vector for outputting compressed principal component vectors compressed from the calculated feature vectors A component output process, a compression process in which each feature vector is compressed and output at different input compression rates, a subspace generation process in which a subspace of the compressed feature vector of each compressed object is obtained and output, and designation There is a control process in which the compression process and the partial space generation process are repeated a plurality of times by the number of the plurality of compression ratios.

  As the compression process, all the feature input processes for inputting all feature vectors of all objects, and the number of dimensions (N) of the feature vectors to be compressed from the input compression rate, the top N principal component vectors and feature vectors are obtained. And a compressed feature output process for outputting a compressed feature vector in which each inner product value is regarded as a vector.

  As the subspace generation process, each object compression feature input process for inputting the compression feature vector of each object, a principal component analysis process for performing principal component analysis on the compressed feature vector, and the analyzed subspace principal component And a subspace principal component output process for outputting.

  The process of identifying an object includes an input process of inputting a target image to be identified, a target area image cutting process of cutting out a target area image while changing the position and size of interest in the target image, and target features from the area image. A feature extraction process for extracting vectors, an all category registration process for registering all categories as candidate categories, a candidate identification process for narrowing down candidates by a plurality of specified compression ratios, and the number of input multiple compression ratios Some have a control process that repeats the candidate identification process.

  As the candidate identification process, the number of dimensions (N) is obtained from the input compression rate, the inner product of the top N principal components and the target feature vector is calculated, and the target compressed feature vector in which each inner product value is regarded as a vector A target compression feature vector calculation process for calculating the target compression feature vector, and a projection distance calculation process for calculating a projection distance when the target compression feature vector is projected onto the partial space using the main component of the partial space for the compression rate corresponding to each candidate category; Some have a candidate category update process in which the top K projection distances of each candidate below the input threshold are new candidates, and an identification result output process that outputs candidates below the input threshold as an identification result .

  FIG. 3 is a schematic configuration diagram showing an embodiment of the invention according to claims 2 and 3, and particularly shows a schematic configuration of the matching degree calculation means 15.

  The collation degree calculation unit 15 includes a compression unit 151, a projection distance calculation unit 152, a rank calculation unit 153, and a rank cut unit 154.

  The compression unit 151 calculates a target compressed feature vector obtained by compressing the feature vector with an input compression rate (for example, the same compression rate as that of the corresponding dictionary) according to the number of repetitions (i).

  The projection distance calculation unit 152 calculates a distance (projection distance) when the target compression feature vector is projected onto the partial space using the partial space (recognition dictionary) input in advance for the compression rate corresponding to each candidate category.

  The rank calculation means 153 obtains a rank according to the projection distance of each candidate category.

  The rank cutting means 154 increases the projection distance of candidate categories below the specified rank.

  The collation degree represents the similarity between the rectangular image and the category, and is calculated for each position of interest of each size.

  FIG. 11 shows an example of a calculation result by the matching degree calculation means 15.

  An example of the matching degree of “den” at size A and the matching degree at size B is shown. In the figure, the darker the pixel, the higher the matching degree (the shorter the projection distance). At this time, if the collation degree at a certain position is equal to or lower than the threshold value or equal to or lower than the determined rank, the position is deleted as a candidate.

  The compression is performed by performing feature conversion using the top K principal components obtained by performing principal component analysis on the feature vector group in advance. The compression rate can be changed by changing K. For the principal component analysis, a known one may be employed, for example, the one disclosed in “Pattern Recognition” (published by Asakura Shoten, pp. 35) by Nobuyuki Otsu et al. In the partial space, the principal component analysis is performed again on the pre-compressed feature vector group, and the partial space is configured by the top L of the principal components. L is a parameter that varies depending on the number of repetitions i.

  The projection distance may be calculated by the following method. In this method, the projection distance when the target compression feature vector is projected onto the partial space using the principal component of the partial space corresponding to the compression rate corresponding to each candidate category is calculated.

  First, the number of dimensions (N) is obtained from the input compression rate, the inner product of the top N principal components and the target feature vector (IF) is calculated, and the target compressed feature vector (AIF) in which each inner product value is regarded as a vector. Is calculated. Next, a distance (projection distance) when the target compression feature vector is projected onto the partial space using the partial space principal component corresponding to the compression rate corresponding to each candidate category is calculated. The projection distance L (c) of each category is calculated by the equation (1). In this equation, the r-th subspace principal component is represented as BAF (c, r), and BAF (c, r) is represented by an N-dimensional vector (where r = 1 to N). R corresponding to each compression rate is a constant input in advance.

  FIG. 4 is a schematic configuration diagram of the feature extraction means in the embodiment of the invention according to claim 4.

  The feature extracting unit 14 includes an input unit 141, a differential intensity direction calculating unit 142, a differential direction histogram forming unit 143, and a differential histogram normalizing unit 144.

  The input unit 141 inputs a plurality of images.

  The differential intensity direction calculating unit 142 calculates the differential and vertical differential components of each image to calculate the differential direction and strength.

  The differential direction histogram forming means 143 quantizes the differential direction of each pixel in a predetermined area for each image at a predetermined stage, and creates a differential direction histogram in which the strength of the differential is cumulatively added for each stage. .

  The differential histogram normalizing unit 144 regards the created differential direction histogram as a vector and normalizes the magnitude to a predetermined value.

  FIG. 5 illustrates an example of the operation of the differential intensity direction calculation unit 142.

  The horizontal direction of the original image I is considered to be the x axis and the vertical direction is considered to be the y axis. The image has horizontal X pixels × vertical Y pixels, and the image size is X × Y. First, a Sobel operator is applied to the original image to generate an x-direction differential image Dx in which the x-direction differential is calculated and a y-direction differential image Dy in which the y-direction differential is calculated. However, using the Sobel operator is merely an example, and other methods may be used. Next, each pixel of the differential intensity image Di and the differential direction image Dd is calculated by the following equation (2).

  An example of the operation of the differential direction histogram forming means is shown in FIG.

  In each pixel, the angle is first quantized. Next, the intensity of the pixel is added to the histogram of the corresponding angle in the region to which the pixel belongs. Thus, a differential direction histogram is created. In the example, a histogram is created by quantizing the direction into 5 (m) steps within an area obtained by dividing the image into 4 (n), and the feature is expressed by a vector G of 20 (n × m) dimensions.

  The differential direction histogram is normalized by the following equation 3 to obtain a feature vector (F). However, the scalar value of the kth dimension of the vector G is represented as G (k).

  FIG. 7 is a schematic configuration diagram of the connected region peak detecting means in the embodiment of the invention according to claim 5.

  The connected region peak detection means 16 includes a repetition control means A 161, a repetition control means B 162, a right lateral connectivity determination means 163, a lower connectivity determination means 164, a region extraction means 165, and a peak point output means 166.

  The repetition control unit A161 executes and controls repetition of processing of the later-described repetition control unit B162 at all attention positions and attention sizes.

  The repetition control unit B162 executes and controls the repetition of processing by the later-described right lateral connectivity determination unit 163 and the lower connectivity determination unit 164 for the categories registered as candidates for the target position and target size.

  The right lateral connectivity determining unit 163 determines that the target position and the right lateral position are connected when the matching degree between the target position and the right lateral position is equal to or less than the threshold value TH.

  The lower connectivity determination unit 164 determines that the target position and the lower position are connected when the matching degree between the target position and the position below the target position is equal to or less than a predetermined threshold TH.

  The area extraction unit 165 extracts the connected attention position group as one area.

  The peak point output means 166 obtains and outputs the position and size at which the matching degree in each region is maximum.

  As an embodiment of the invention according to claim 6, the in-connection region peak detection means 16 may include a pre-connectivity determination means.

  In the connected region peak detection means 16 shown in FIG. 7, the repetition control means A161 repeats the processing of the repetition control means B162 at all the attention positions and attention sizes. Next, the repeat control unit B162 repeats the right lateral connectivity determination unit and the lower connectivity determination manual process for the categories registered as candidates for the target position and target size. Next, the right lateral connectivity determining unit 163 determines that the target position and the right lateral position are connected when the matching degree between the target position and the right lateral position is equal to or less than the threshold value TH. Next, the lower connectivity determination unit 164 determines that the target position and the lower position are connected when the collation degree between the target position and the position below the target position is equal to or less than a predetermined threshold value TH.

  Here, the previous connectivity determination means (not shown) determines that the target position and the target position and the target size are connected to each other when the degree of matching at the same position of the target size is equal to or less than a predetermined threshold.

  Next, the region extraction unit 165 extracts the connected attention position group as one region. Then, the peak point output means 166 obtains and outputs the position and size at which the matching degree in each region is maximum.

  FIG. 12 shows an example of the connection determination at the point P. In this example, the same position of the size B on the right side and the lower side of the P point and one level above is determined as the connection.

  FIG. 13 is an example of determination of connection at all points. In this example, the left side, right side, top and bottom, and the same position (Q point) of size B one level above are determined to be connected. . In addition, at the Q point, the same position (P point) of size A on the left, right side, up and down, and one step below is determined to be connected.

  FIG. 14 shows an example of the result of region extraction, and shows the result of extracting connected regions as one region and labeling each region. The pixel in the region including the P point and the Q point is labeled with the region 1, and the pixel in the region including the R point and the S point is labeled with the region 2.

  FIG. 15 shows an example of the peak point output result. The P point having the minimum projection distance in the region 1 and the R point having the minimum projection distance in the region 2 are used as the peak points. It is output.

  FIG. 16 is a schematic configuration diagram of candidate category update means in the object recognition apparatus according to claim 7.

  The candidate category update unit 17 shown in the figure includes a peripheral region calculation unit 171 and a peripheral region registration unit 172. The peripheral area calculation unit 171 calculates the peripheral area of the obtained peak point. The peripheral area registration unit 172 registers the peak point and its peripheral area as candidates.

  FIG. 17 shows an example of the result of the peripheral area calculation means 171. In this example, the upper, lower, left, and right sides of the peak point and one step above the size are defined as the peripheral area. However, this example is merely an example, and other polyhedron definitions may be used.

  The operation example of the component of the object recognition apparatus demonstrated above is shown.

  FIG. 18 is a flowchart illustrating an operation example of the identification device 1.

Step 1) Input of target image Step 2) Registration of all categories Step 3) Repeat for the number of points of interest V × the number of assumed sizes W [1]
Step 4) Extraction of region of interest Step 5) Feature extraction Step 6) Repeat [1] End Step 7) Repeat for input number I [2]
Step 8) Calculation of matching degree Step 9) Peak detection Step 10) Update of candidate category Step 11) Repeat [2] End Step 12) Output of identification result FIG. 19 is a flowchart showing an operation example of the feature extraction means 14 It is.

Step 1) Repeat for P number of pixels of image [1]
Step 2) Calculation of differential intensity Step 3) Repeat [1] End Step 4) Repeat for n divisions [2]
Step 5) Calculation of Differential Histogram Step 6) Repeat [2] End Step 7) Normalization of Differential Histogram FIG. 20 is a flowchart showing an operation example of the matching degree calculation means 15.

Step 1) Compression Step 2) Repeat for C categories [1]
Step 3) Projection distance calculation Step 4) Repeat [1] end Step 5) Repeat for C categories [1]
Step 6) Rank calculation Step 7) Repeat [1] End Step 8) Rank cut FIG. 21 is a flowchart showing an operation example of the connected region peak detecting means 16.

Step 1) Repeat for J candidates [1]
Step 2) Right lateral connectivity determination Step 3) Lower connectivity determination Step 4) Previous connectivity determination Step 5) Repeat [1] End Step 6) Region extraction (labeling)
Step 7) Repeat for the number of areas [2]
Step 8) Peak Point Output Step 9) Repeat [2] End FIG. 22 is a flowchart showing an operation example of the candidate category update means 17.

Step 1) Whether i is I or not. If yes, the process ends. Step 2) Peripheral Area Calculation Step 3) Peripheral Area Registration According to the above flowchart, an output example of the following identification result can be obtained for the target image in the example of FIG.

[5, 5, 10, electricity, 1000. ]
[10, 5, 10, Shin, 900]
[15, 5, 10, electricity, 820]
[20, 5, 10, story, 1200]
It should be noted that the object identification methods described in the above embodiments can be configured by configuring the processing steps shown in FIGS. 1 to 22 with a computer program and causing the computer to execute the program. A recording medium that can be read by the computer, such as a flexible disk, MO, ROM, memory card, CD, DVD, a program for realizing the function, or a program for causing the computer to execute the process. It can be recorded on a removable disk, HDD, etc., and stored or distributed. It is also possible to provide this program via a network such as the Internet or electronic mail.

  Then, the present invention can be implemented by installing the program from these recording media into a computer, or by downloading the program from a network and installing the program into the computer. However, installation on a computer is a computer unit, and when there are multiple computers to be installed due to multiple devices and systems, it is natural that the program is installed for each necessary processing part. is there. In this case, the program may be recorded on a recording medium corresponding to a computer, or downloaded via a network.

The schematic block diagram of the character recognition translation system in a landscape for this invention. The schematic block diagram which showed the embodiment of the invention which concerns on Claim 1. FIG. 4 is a schematic configuration diagram of the collation degree calculating means 15 in particular in the embodiment of the invention according to claims 2 and 3. FIG. The schematic block diagram of the characteristic extraction means in the embodiment of the invention concerning Claim 4. Explanatory drawing of the operation example of a differential intensity | strength direction calculation means. Explanatory drawing of the operation example of a differential direction histogram-izing means. The schematic block diagram of the connection area | region peak detection means in the example of embodiment of the invention which concerns on Claim 5. FIG. An example of a target image. The figure which showed the attention point in a target image. An example of “specified size”. An example of the calculation result by a collation degree calculation means. An example of the connection determination in P point. An example of the connection determination in P point. An example of the result of area extraction. An example of the peak point output result. The schematic block diagram of the candidate category update means in the object recognition apparatus which concerns on Claim 7. An example of the result of a surrounding area calculation means. The flowchart which showed the operation example of the identification device. The flowchart which showed the operation example of the feature extraction means. The flowchart which showed the operation example of the collation degree calculation means. The flowchart which showed the operation example of the peak detection means in a connection area | region. The flowchart which showed the operation example of the candidate category update means.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Identification apparatus, 2 ... User, 3 ... Dictionary, 4 ... Translation apparatus 10 ... Input means, 11 ... All category registration means, 12 ... Repetition control means, 13 ... Image extraction means, 14 ... Feature extraction means, 15 ... Collation Degree calculation means, 16 ... connected region peak detection means, 17 ... candidate category update means 151 ... compression means, 152 ... projection distance calculation means, 153 ... rank calculation means, 154 ... rank cut means 141 ... input means, 142 ... differentiation Intensity direction calculating means, 143... Differential direction histogram forming means, 144... Differential histogram normalizing means 161 .. Repeat control means A, 162 .. Repeat control means B, 163. 165 ... Area extraction means, 166 ... Peak point output means

Claims (16)

An object identification device for identifying a plurality of objects in an image,
An all-category registration means for inputting an identification target image and registering all categories as candidate categories for a predetermined position in the image and its size;
Control means for repeatedly executing processing by the following means;
Image cutout means for cutting out an image of interest while changing the position of interest and the size thereof in the target image;
Feature extraction means for extracting a target feature vector from the cut-out image;
A degree-of-matching calculating unit that compares a recognition dictionary of each object according to the number of repetitions created in advance and a target feature vector, and calculates a degree of matching that represents the degree of matching;
Determining the spatial connectivity of the matching degree of each object, and detecting a peak in the connected area for detecting the maximum peak of the matching degree in the connected area;
An object identification device comprising candidate category update means for registering detected peaks as candidates. The collation degree calculating means includes
Compression means for calculating a target compressed feature vector obtained by compressing a feature vector with an input compression rate corresponding to the number of repetitions;
2. A projection distance calculating means for calculating a distance when a target compression feature vector is projected onto a partial space using a partial space input in advance for a compression rate corresponding to each candidate category. The object identification device described. The collation degree calculating means includes
Compression means for calculating a target compressed feature vector obtained by compressing a feature vector with an input compression rate corresponding to the number of repetitions;
A projection distance calculating means for calculating a projection distance when the target compression feature vector is projected onto the partial space using a partial space previously input with respect to a compression rate corresponding to each candidate category;
A rank calculating means for calculating a rank according to the projection distance of each candidate category;
The object discriminating apparatus according to claim 1, further comprising rank cutting means for increasing a projection distance of a candidate category having a specified rank or less. The feature extraction means includes
Differential intensity direction calculation means for calculating the differential and vertical differential components of each image and calculating the differential direction and strength;
For each image, a differential direction histogram forming means for quantizing the differential direction of each pixel in a predetermined region to a predetermined stage and creating a differential direction histogram in which the strength of the differential is cumulatively added for each stage;
The object identification according to any one of claims 1 to 3, further comprising normalizing means that regards the created differential direction histogram as a vector and normalizes the magnitude to a predetermined value. apparatus. The connected region peak detection means,
Right-side connectivity determination means for determining connection when the degree of matching between the noted position and its right-side position is equal to or less than a predetermined threshold;
Lower connectivity determination means for determining that the connection is established when the degree of matching between the noted position and the position below it is equal to or less than a predetermined threshold;
Region extraction means for extracting a group of connected positions of interest as one region;
Peak point output means for obtaining and outputting the position and size where the matching degree in each region is maximum,
A step of repeatedly executing the processing by the right lateral connectivity determining means and the lower connectivity determining means for the category registered as a candidate for the noted position and its size, and at all the noted positions and their sizes 5. The object identification device according to claim 1, further comprising a control unit that controls a step of repeatedly executing the processing of the step. 6. The connected region peak detection means,
Right-side connectivity determination means for determining connection when the degree of matching between the noted position and its right-side position is equal to or less than a predetermined threshold;
Lower connectivity determination means for determining that the connection is established when the degree of matching between the noted position and the position below it is equal to or less than a predetermined threshold;
Pre-connectivity determining means for determining connection when the degree of matching at the same position of the noted position and the size of the position one step above is not more than a predetermined threshold;
A region extracting means for extracting the connected attention position group as one region;
Candidate point output means for obtaining and outputting the position and size where the matching degree in each region is maximum,
The step of repeatedly executing the processing by the right lateral connectivity determining means, the lower connectivity determining means, the front connectivity determining means, and the back connectivity determining means for the category registered as a candidate for the noted position and its size. The object identification according to claim 1, further comprising: a control unit that controls a process of repeatedly executing the process of the process at all focused positions and sizes thereof. apparatus. The candidate category update means includes:
A peripheral area calculation means for calculating a peripheral area of the obtained peak point;
7. The object identification device according to claim 1, further comprising peripheral region registration means for registering the peak point and its peripheral region as candidates. An object identification method for identifying a plurality of objects in an image,
All category registration means, image cutout means, feature extraction means, matching degree calculation means, connected region peak detection means, and control means for controlling these means,
A step of registering all categories as candidate categories with respect to a predetermined position and its size in the input identification target image;
A step in which the control means repeatedly executes processing by the following means;
A step of cutting out an image of interest while changing the position and size of the image of interest in the target image;
A feature extracting means for extracting a target feature vector from the cut-out image;
A step of comparing the object feature vector with the recognition dictionary of each object according to the number of repetitions created in advance, and calculating a matching degree indicating the matching degree;
A step in which a peak detection unit in the connected region determines the spatial connectivity of the matching degree of each object, and detects the maximum peak of the matching degree in the connected region;
And a candidate category updating means for registering the detected peak as a candidate. In the step of calculating the matching degree by the matching degree calculating means,
Calculating a target compressed feature vector obtained by compressing a feature vector with an input compression rate corresponding to the number of repetitions;
9. The object according to claim 8, further comprising a step of calculating a distance when the target compression feature vector is projected onto the partial space using a partial space previously input with respect to a compression rate corresponding to each candidate category. Identification method. In the step of calculating the matching degree by the matching degree calculating means,
Calculating a target compressed feature vector obtained by compressing a feature vector with an input compression rate corresponding to the number of repetitions;
Calculating a projection distance when the target compression feature vector is projected onto the partial space using the partial space input in advance for the compression ratio corresponding to each candidate category;
Calculating a rank according to the projection distance of each candidate category;
The object identification method according to claim 8, further comprising a step of increasing a projection distance of candidate categories equal to or lower than a specified rank. In the step of extracting the target feature vector by the feature extraction means,
Calculating a differential direction and strength of each image by calculating a horizontal differential and a vertical differential component of each image;
For each image, a step of quantizing the differential direction of each pixel in a predetermined region to a predetermined stage and creating a differential direction histogram in which the strength of the differential is cumulatively added for each stage;
The object identification method according to claim 8, further comprising a step of regarding the generated differential direction histogram as a vector and normalizing the magnitude to a predetermined value. In the step of detecting the maximum peak of the matching degree by the connected region peak detection means,
Determining the connection when the degree of matching between the noted position and the position on the right side thereof is not more than a predetermined threshold;
Determining the connection when the degree of matching between the noted position and the position below it is equal to or less than a predetermined threshold;
Extracting a group of linked positions of interest as one region;
A step of obtaining and outputting a position and size having the maximum matching degree in each region; and a right-side connectivity determination means and a lower connectivity with respect to a category registered as a candidate for the noted position and its size. A step of repeatedly executing processing by the determination means;
The object identifying method according to claim 8, further comprising a step of repeatedly executing the processing of the step at all focused positions and sizes thereof. In the step of detecting the maximum peak of the matching degree by the connected region peak detection means,
Determining the connection when the degree of matching between the noted position and the position on the right side thereof is not more than a predetermined threshold;
Determining the connection when the degree of matching between the noted position and the position below it is equal to or less than a predetermined threshold;
A step of determining connection when the degree of matching at the same position of the noted position and its size one step above is equal to or less than a predetermined threshold;
Extracting the connected attention position group as one area;
A step of obtaining and outputting a position and size having the maximum matching degree in each region; and a right-side connectivity determination means and a lower connectivity with respect to a category registered as a candidate for the noted position and its size. A step of repeatedly executing processing by the determination means, the front connectivity determination means and the back connectivity determination means;
13. The object identification method according to claim 8, further comprising a step of repeatedly executing the processing of the step at all focused positions and sizes thereof. In the step performed by the candidate category update means,
Calculating the area around the obtained peak point;
The object identification method according to claim 8, further comprising a step of registering the peak point and its peripheral region as candidates. The program for making a computer perform the object identification method of any one of Claims 8-14. The computer-readable recording medium which recorded the program for making a computer perform the object identification method of any one of Claims 8-14.

Images