JP5353482B2 - Pattern recognition dictionary generation device, pattern recognition device, and pattern recognition dictionary generation method - Google Patents

Description

  The present invention relates to a pattern recognition apparatus, and more particularly to a large classification method for speeding up recognition.

  The pattern recognition device is a device that recognizes an input pattern and determines its category. Examples of the pattern include image data and audio data. For example, in the case of character recognition, the pattern is an image. The category is a classification of patterns to be identified. For example, in character recognition, the category for numeric recognition is 10 character types “0” to “9”, and the category for kanji recognition is several thousand character types.

  When the pattern recognition device is a character recognition device, the character recognition device determines which character type (category) corresponds to a character type (category) among the character types (categories) set in advance in the input image, and determines the determination result. Output.

  The process executed by the pattern recognition apparatus is composed of a learning phase and a recognition phase.

  In the learning phase, the pattern recognition device creates a feature selection dictionary and an identification dictionary using a learning pattern DB (database).

  Specifically, in the learning phase, a large classification function and a detailed classification function are generated, and the generated detailed classification function and large classification function are stored in the identification dictionary. The large classification identification function is a function used to roughly narrow down correct answer candidates. The detailed identification function is a function for calculating the similarity of the narrowed-down correct answer candidates.

  As a detailed discriminant function generation algorithm, there are a nearest neighbor method, a perceptron, an improved projection distance method, a modified quadratic discrimination, a general learning vector quantization, a polynomial network, a support vector machine, etc. Patent Document 2).

  As an algorithm for generating a large classification discriminant function, there are a method using a high-speed discriminant function that is less accurate than a detailed discriminant function, and a method of performing large classification by reducing the number of categories to be identified.

  As a method using a high-speed discriminant function, for example, a discriminant function having a small calculation amount such as a linear discriminant function is used for large classification. The method described above is the same as the algorithm for generating the detailed discriminant function, except that an discriminant function with a small amount of calculation is used.

  A method of performing large classification by reducing the number of categories to be identified is performed by collecting some categories to be recognized or clustering distributions. For example, in alphabet recognition, there is a method of treating character types (“I” and “l” or “K” and “k”, etc.) that are close in the feature space as the same category.

  More specifically, as an algorithm for generating a large classification discriminant function, a method using a linear discriminant function (for example, see Non-Patent Document 1) or a clustering method using general learning vector quantization (for example, Patent Document 1). For example).

  In the recognition phase, the pattern recognition device recognizes an input pattern using the created feature selection dictionary and identification dictionary. In the recognition phase, a large classification identification process and a detailed identification process are executed.

  First, the purpose of executing the large classification identification process will be described.

  The large classification identification process is a process for performing a coarse identification process faster than the detailed identification process using the detailed identification function, and narrowing down candidates for the correct category from all categories.

  When only the detailed identification process is executed, the pattern recognition apparatus needs to execute the process for all categories. However, in the above-described method, for example, the processing time becomes enormous in the recognition of kanji for which several thousand categories or more are to be recognized. Therefore, the pattern recognition apparatus speeds up the processing by executing a two-stage identification process including a large classification identification process and a detailed identification process.

  In the large classification identification process, correct answer candidates are roughly narrowed down. For example, in kanji recognition, the correct answer candidates are reduced to about several tens to several hundreds by executing the large classification identification process. In the detailed classification, the similarity to each of the narrowed correct answer candidate categories is calculated using a detailed identification function.

  In pattern recognition, in order to speed up the recognition process, first, a rough classification process for narrowing down the correct candidate categories is performed using a large classification function that can be calculated at high speed. Thereafter, using the detailed identification function for the correct candidate category, the similarity of the input pattern for each of the correct candidate categories is calculated, and the final recognition result is output.

  As a method of conventional major classification identification processing and detailed identification processing, there is a major classification method using a template method.

  In the above-described method, a plurality of representative vectors representing one or a plurality of categories are prepared in advance in the feature space by learning using the learning pattern DB for the large classification identification process.

  During the recognition phase, the pattern recognition apparatus calculates the distance between the representative vector and the input pattern, and sets the category included in the category group represented by the representative vector that is close to the input pattern as the correct candidate category. The pattern recognition apparatus performs detailed identification processing for each of the correct candidate categories calculated as described above. In the case described above, the template method is usually used for the detailed identification process. That is, representative vectors representing each category are created in advance by learning, and in the recognition phase, the pattern recognition device calculates the similarity between the input pattern and each category from the distance between the input pattern and the representative vector. .

Japanese Patent No. 3475886

Mohammed Cheriet, Nawwaf Kharma, Cheng lin Liu, and Ching Suen.Character Recognition Systems: A Guide for Students and Practitioners.Wiley-Interscience, 2007. Kenichiro Ishii, Nobuo Ueda, Eisaku Maeda, Hiroshi Murase. Pattern recognition. Ohm Publishing House. Liu, C. L., Sako, H., Fujisawa, H. Performance evaluation of pattern classifier for handwritten character recognition, International Journal of Document Analysis and Recognition, Vol.4, No.3, pp.191-204.

  However, the large classification identification process using the template method has a problem in accuracy. In particular, it is confirmed by experiments that the recognition accuracy using the template method is lower than the recognition accuracy using the neural network or the support vector machine when the large classification recognition process using the template method is executed in handwritten character recognition. (For example, refer nonpatent literature 3).

  Further, when the template method is used only for the large classification identification process and the detailed identification process uses another learning algorithm, the learning algorithm and the recognition algorithm are different between the large classification identification process and the detailed identification process, so that the configuration is complicated. There is an implementation problem of becoming.

  In addition, as a method for conventional large classification identification processing and detailed identification processing, a large classification identification function is used which uses a large classification identification function which is inferior in accuracy such as a linear function but has a small amount of calculation. There is a method to use. Similarly, when the above method is used, there are a problem in accuracy of the large classification identification function and a problem in the construction that the learning algorithm is duplicated in the large classification identification processing and the detailed identification processing. .

  A typical example of the present invention is as follows. A pattern recognition dictionary generation device comprising a processor and a storage medium connected to the processor, wherein the storage medium stores a learning pattern database composed of a plurality of learning patterns, and the pattern The recognition dictionary generation device includes a pattern input unit that acquires each of the learning patterns as one category from the learning pattern database, and a feature extraction unit that extracts an n-dimensional feature for each of the acquired categories. , Using the extracted n-dimensional feature, generating a feature selection function for converting the n-dimensional feature into an m-dimensional feature that is a dimension less than or equal to the n-dimension, and using the generated feature selection function as a feature selection dictionary And using the feature selection function to store the extracted n-dimensional feature as the m-dimensional feature. Using the feature selection unit to convert and the converted m-dimensional feature, a detailed identification function on the m-dimensional feature space for calculating the similarity of the pattern to be recognized for each category is generated, and the generated A discriminant function generation unit that stores a detailed discriminant function as a discriminating dictionary in the storage medium, a dimension of m dimensions or less, a subspace of the n-dimensional feature space, and a subspace of the m-dimensional feature space A large classification feature selection function for converting the m-dimensional feature into an L-dimensional feature on a certain L-dimensional feature space is generated, and the detailed identification function is converted as a function on the L-dimensional feature space to thereby convert the L-dimensional feature space. A large classification identification function for calculating the similarity of the pattern to be authenticated with respect to each category is generated above, and the generated large classification feature selection function is used as the feature selection dictionary. Stored in 憶媒 body, characterized in that it comprises an identification function main portion extraction unit for storing in the storage medium a large classification identification function said generated as a dictionary for the identification.

  By using the large classification discriminant function calculated as a function on the L-dimensional feature space as the detailed discriminant function, the accuracy of the pattern recognition process can be maintained and the speed can be increased. Further, since the major classification function is generated from the detailed classification function, the configuration of the learning algorithm for the major classification process and the detailed classification process can be facilitated.

It is a block diagram which shows an example of a structure of the pattern recognition apparatus of the 1st Embodiment of this invention. It is a flowchart explaining the structure of the module and DB (database) which perform the process of the learning phase in the pattern recognition apparatus of the 1st Embodiment of this invention. It is a flowchart explaining the structure of the module and DB (database) which perform the process of the recognition phase in the pattern recognition apparatus of the 1st Embodiment of this invention. It is a flowchart explaining the process which the discriminant function main part extraction part of the 1st Embodiment of this invention performs. It is a flowchart explaining an example of the process which the feature extraction part in the character recognition of the 1st Embodiment of this invention performs. It is a flowchart explaining the process of the learning phase in the conventional pattern recognition apparatus. It is a flowchart explaining the process of the recognition phase in the conventional pattern recognition apparatus. It is a flowchart which shows the flow of a series of processes of the conventional pattern recognition apparatus.

  First, the prior art will be described.

  FIG. 6 shows a configuration diagram of modules and DBs (databases) that execute processing in the conventional learning phase, and FIG. 7 shows a configuration diagram of modules and DBs that execute processing in the conventional recognition phase. The learning phase and the recognition phase are summarized as shown in FIG.

  FIG. 6 is a flowchart for explaining the learning phase processing in the conventional pattern recognition apparatus.

  The pattern input unit 201 acquires a pattern from the learning pattern DB 207 and outputs the acquired pattern to the feature extraction unit 202.

  The feature extraction unit 202 extracts an n-dimensional vector from the input pattern. Hereinafter, the n-dimensional vector extracted by the feature extraction unit 202 is referred to as an n-dimensional feature.

  The extracted n-dimensional features are output to the feature selection dictionary generation unit 203 and the feature selection unit 204.

  By the process executed by the feature extraction unit 202, the input pattern is expressed as an n-dimensional vector even if the data for pattern recognition is audio or image. Therefore, the pattern recognition apparatus can apply the same processing regardless of the type of pattern.

  The subsequent feature selection dictionary generation unit 203 may require n-dimensional features of a plurality of patterns in order to generate the feature selection dictionary 208. In this case, every time an n-dimensional feature is required, the pattern input unit 201 and the feature extraction unit 202 may execute processing to extract the necessary n-dimensional feature.

  In addition, the pattern input unit 201 and the feature extraction unit 202 execute processing for all learning patterns in advance to convert them into n-dimensional features, and store the n-dimensional features in the external storage device 107 (see FIG. 1) or the like. Alternatively, the n-dimensional feature may be acquired from the external storage device 107 whenever the feature selection dictionary generation unit 203 is required.

  The feature selection dictionary generation unit 203 generates a conversion function for converting n-dimensional features into m-dimensional features (m ≦ n), and stores the generated conversion functions in the feature selection dictionary 208.

  For example, when an n-dimensional feature is represented by x and an m-dimensional feature is represented by y, the conversion function f is represented by y = f (x). When the conversion is limited to linear conversion, y = Yx can be expressed using an m × n matrix Y. In some cases, n-dimensional features are converted to m-dimensional features using different conversion functions for each category. In this case, the m-dimensional feature yk of the category k is expressed as yk = fk (x) using the conversion function fk of the category k. In this case, the feature selection dictionary generation unit 203 generates a conversion function f or a conversion function fk for each category, and stores the generated conversion function f or fk in the feature selection dictionary 208.

  As a method for generating the conversion function, a method using a principal component analysis method, a linear discrimination method, or the like can be considered. The purpose of feature selection is to realize high-speed and high-precision recognition processing by extracting effective components in identification processing from n-dimensional features and reducing the number of dimensions of n-dimensional features.

  The feature selection unit 204 uses the conversion function stored in the feature selection dictionary 208 to convert the n-dimensional feature into an m-dimensional feature. When the conversion function is f, the m-dimensional feature y is expressed as y = f (x) with respect to the n-dimensional feature x. When different conversion is performed for each category, the m-dimensional feature yk for each category is expressed as yk = fk (x). The converted m-dimensional feature is output to the identification dictionary generation unit 205.

  The subsequent identification dictionary generation unit 205 may require m-dimensional features of a plurality of patterns in order to generate the identification dictionary 209. In this case, each time an m-dimensional feature is required, the pattern input unit 201, the feature extraction unit 202, and the feature selection unit 204 may perform processing, and the identification dictionary generation unit 205 may acquire the required m-dimensional feature. .

  Further, the pattern input unit 201, the feature extraction unit 202, and the feature selection unit 204 execute processing for all the learning patterns in advance to convert them into m-dimensional features, and the m-dimensional features are converted into the external storage device 107 (see FIG. 1). The identification dictionary generating unit 205 may acquire the m-dimensional feature from the external storage device 107 whenever necessary.

  The identification dictionary generation unit 205 includes two processes: a process for generating a detailed identification function and a process for generating a large classification identification function.

  In the process of generating the detailed identification function, the detailed identification function is generated using the m-dimensional feature generated from the pattern stored in the learning pattern DB 207 and the set of labels indicating the category to which the m-dimensional feature belongs. The

  Similarly, in the process of generating the large classification identification function, using the m-dimensional feature generated from the pattern stored in the learning pattern DB 207 and a set of labels indicating the category to which the m-dimensional feature belongs, A detailed discriminant function is generated. The generated detailed identification function and large classification identification function are stored in the identification dictionary 209.

  First, processing for generating a detailed identification function will be described. In this process, a detailed identification function is generated using the learning pattern DB 207, and the generated detailed identification function is stored in the identification dictionary 209.

  A detailed identification function exists for each category. The detailed identification function uk of the category k is a function for calculating the similarity of the pattern to the category k. The similarity of the pattern to category k is calculated by uk (y). When different m-dimensional features yk are extracted for each category, the similarity of the pattern to category k is calculated as uk (yk).

  The generation algorithm of the detailed discrimination function includes a nearest neighbor method, a perceptron, an improved projection distance method, a modified secondary discrimination, a general learning vector quantization, a polynomial network, and a support vector machine (for example, Non-Patent Document 1, Non-Patent Document 1). Reference 2).

  Next, processing for generating a large classification identification function will be described. In this processing, a large classification identification function is generated using the learning pattern DB 207, and the generated large classification identification function is stored in the identification dictionary 209.

  First, the purpose of executing the large classification identification process will be described.

  The large classification identification function is for performing coarse identification processing faster than the detailed identification processing using the detailed identification function, and narrowing down candidates for the correct category from all categories.

  When only the detailed identification function is executed, the pattern recognition apparatus must calculate the similarity for each category of the pattern with respect to all categories using the detailed identification function uk (y) or uk (yk). .

  However, in the above-described method, for example, in kanji recognition for which several thousand categories or more are to be recognized, the processing time is enormous. For this reason, the speed of the identification process is increased by two stages of the large classification identification process and the detailed identification process. In the large classification identification process, correct answer candidates are roughly narrowed down.

  As an algorithm for generating a large classification discriminant function, there are a method using a high-speed discriminant function that is less accurate than a detailed discriminant function, and a method of performing large classification by reducing the number of categories to be identified.

  As a method using a high-speed discriminant function, for example, a discriminant function having a small calculation amount such as a linear discriminant function is used for large classification. The method described above is the same as the generation algorithm of the detailed discriminant function except that an discriminant function with a small amount of calculation is used.

  A method of performing large classification by reducing the number of categories to be identified is performed by collecting a plurality of categories as one category, reducing the number of categories to be identified, or clustering the distribution. For example, in alphabet recognition, there is a method of treating character types (“I” and “l” or “K” and “k”, etc.) that are close in the feature space as the same category.

  A feature space different from the detailed discriminant function may be used for the major discriminant function. In the above case, the feature selection function for the large classification identification process is generated, and the generated feature selection function for the large classification identification process is stored in the feature selection dictionary 208. When the feature used for the large classification identification process is expressed as z and the conversion function is expressed as g, the characteristic used for the large classification identification process is calculated as z = g (x). When the large classification identification function of the large classification category c is represented as vc, the similarity to the category c is calculated by the large classification identification function vc (z).

  FIG. 7 is a flowchart for explaining the process of the recognition phase in the conventional pattern recognition apparatus.

  The pattern input unit 201 acquires a pattern from the recognition target pattern DB 304 and outputs the acquired pattern to the feature extraction unit 202.

  The feature extraction unit 202 extracts n-dimensional features from the input pattern, and outputs the extracted n-dimensional features to the feature selection unit 204.

  The feature selection unit 204 uses the conversion function stored in the feature selection dictionary 208 to convert the n-dimensional feature into an m-dimensional feature. The m-dimensional feature y is expressed as y = f (x) using the conversion function f. When different conversion is performed for each category, the m-dimensional feature yk of the category k is expressed as yk = fk (x) using the conversion function fk. The converted m-dimensional feature is output to the large classification identifying unit 301.

  In addition, when a feature z different from the detailed identification process is used in the large classification identification process, z = g (x) is calculated using the large classification feature selection function g and is output to the large classification identification unit 301. The When a different feature is used for each category, the feature zc = gc (x) is calculated for each major classification category c, and the calculated feature zc is output to the major classification identification unit 301.

  The large classification identifying unit 301 uses the large classification identification function stored in the identification dictionary 209 to calculate a correct category candidate to which the input pattern belongs, and performs detailed identification of the m-dimensional feature and the correct category candidate. The data is output to the unit 302.

  The detailed identification unit 302 uses the detailed identification function stored in the identification dictionary 209 to calculate the similarity of the input pattern with respect to the correct answer candidate category, and the calculated similarity to the recognition result output unit 303. Output.

  The recognition result output unit 303 outputs the final recognition result using the similarity to the correct candidate category.

  Usually, the category with the highest similarity is output as the recognition result. It should be noted that categories whose similarity is second or later may also be output as second candidates, third candidates, and the like.

  If the similarity is smaller than the specified threshold value, it may be rejected as not corresponding to any category. For example, in the case of digit recognition, this applies to the case where a kanji is input. In addition, when the difference between the first and second similarities is smaller than the specified threshold, it may be rejected because it is difficult to determine which category. For example, the letters “I” and “l” and the number “1” are rejected because they are difficult to identify depending on the font.

  Here, “reject” may include outputting information indicating that there is no corresponding category for the input pattern.

  The purpose of executing the two-stage identification process of the large classification identification process and the detailed identification process in the large classification identification unit 301 and the detailed identification unit 302 is to speed up the identification process. First, the correct answer candidates are roughly narrowed down by rough identification processing, and detailed identification processing is executed on the narrowed correct answer candidates.

  The large classification identification function used in the large classification identification process is required to be able to be calculated at high speed, and that the correct answer category is included in the correct candidate category calculated by the large classification identification function with high accuracy.

  In order to create a major classification function used in the major classification processing, a method using a linear discriminant function (see, for example, Non-Patent Document 1) or a clustering method using general learning vector quantization (for example, Patent Document 1). For example).

  In pattern recognition, in order to speed up recognition processing, first, rough identification processing for narrowing down correct candidate categories is performed by a large classification identification function that can be calculated at high speed. Thereafter, using the detailed identification function for the correct candidate category, the similarity of the input pattern for each of the correct candidate categories is calculated, and the final recognition result is output.

  As a method of conventional major classification identification processing and detailed identification processing, there is a major classification method using a template method.

  In the above-described method, a plurality of representative vectors representing one or a plurality of categories are prepared in advance in the feature space by learning using the learning pattern DB for the large classification identification process.

  During the recognition phase, the pattern recognition apparatus calculates the distance between the representative vector and the input pattern, and sets the category included in the category group represented by the representative vector that is close to the input pattern as the correct candidate category. Detailed identification processing is executed for each of the correct candidate categories calculated as described above. In this case, a template method is usually used for the detailed identification process. That is, representative vectors representing each category are created in advance by learning, and in the recognition phase, the pattern recognition device calculates the similarity between the input pattern and each category from the distance between the input pattern and the representative vector. .

  FIG. 8 is a flowchart showing a flow of a series of processes of the conventional pattern recognition apparatus. The processing executed by each module is the same as that shown in FIGS.

(First embodiment)
Embodiments of the pattern recognition apparatus of the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram showing an example of the configuration of the pattern recognition apparatus according to the first embodiment of the present invention.

  The pattern recognition device 101 includes an input device 102, a display device 103, a pattern acquisition device 104, a communication device 105, a computing device (CPU) 106, and an external storage device 107.

  The input device 102 is a device for inputting a command executed for controlling a program executed by the arithmetic unit (CPU) 106 and other connected devices. The input device 102 is, for example, a keyboard or a mouse for inputting a command or the like.

  The display device 103 is a device such as a display that displays processing contents.

  The pattern acquisition device 104 is a device for acquiring a pattern such as a scanner or a microphone. The acquired pattern may be stored in the external storage device 107 or the like.

  The communication device 105 is a device used for exchanging data with an external device such as a PC or a server. The communication device 105 acquires an execution command transmitted from an external device, and acquires a pattern such as an image or sound from the external device. In addition, the communication device 105 transmits the content of processing executed in the pattern recognition device 101 to an external device.

  The arithmetic device (CPU) 106 is a device that executes a program stored in the external storage device 107 and executes recognition processing. For example, the arithmetic unit (CPU) 106 creates a feature selection dictionary 503 (see FIG. 2) and an identification dictionary 504 (see FIG. 2) using the learning pattern DB 207 (see FIG. 2), and also selects features. A recognition target pattern recognition process or the like is executed using the dictionary 503 (see FIG. 2) and the identification dictionary 504 (see FIG. 2).

  The external storage device 107 is an external storage device such as an HDD and a memory. The external storage device 107 stores a learning pattern DB 207 (see FIG. 2), a recognition target pattern DB 304 (see FIG. 3), a feature selection dictionary 503 (see FIG. 2), and an identification dictionary 504 (see FIG. 2). The The external storage device 107 stores a program (module) for the processing unit (CPU) 106 to execute processing, and temporarily stores processing results of processing executed by the processing unit (CPU) 106. To do.

  The pattern recognition apparatus 101 may not include the input device 102, the display device 103, the pattern acquisition device 104, or the communication device 105.

  When the pattern recognition device 101 does not include the input device 102, it is possible to use a method of instructing the start of processing from an external device using the communication device 105, or a method of automatically executing processing by time designation or the like. It is done.

  When the pattern recognition device 101 does not include the display device 103, a method of transmitting the processing result to the external device using the communication device 105 or a method of storing the processing result in the external storage device 107 can be considered.

  Output and input to a module that executes processing may be performed via the external storage device 107. For example, when the processing module is the processing unit 1 and the processing unit 2 and the processing unit 2 receives the processing result executed by the processing unit 1 as an input, the processing unit 1 stores the processing result in the external storage device 107. Alternatively, the processing unit 2 may acquire the processing result stored in the external storage device 107 as an input.

  The user uses the input device 102 to control a module that executes processing. In addition, the execution result of the process is displayed via the display device 103.

  Next, processing executed by the pattern recognition apparatus 101 according to the embodiment of the present invention will be described.

  The process executed by the pattern recognition apparatus 101 includes a learning phase and a recognition phase. In the learning phase, a feature selection dictionary 503 (see FIG. 2) and an identification dictionary 504 (see FIG. 2) are created using the learning pattern DB 207 (see FIG. 2). In the recognition phase, the input pattern is recognized using the feature selection dictionary 503 (see FIG. 2) and the identification dictionary 504 (see FIG. 2).

  FIG. 2 is a flowchart illustrating the configuration of a module and a DB (database) that execute the learning phase process in the pattern recognition apparatus 101 according to the first embodiment of the present invention. FIG. 3 is a flowchart illustrating the configuration of a module and a DB (database) that execute processing in the recognition phase in the pattern recognition apparatus 101 according to the first embodiment of the present invention.

  Note that the pattern recognition apparatus 101 may include an apparatus (recognition dictionary generation apparatus) that executes processing in the learning phase and an apparatus (recognition apparatus) that executes processing in the recognition phase. In that case, the recognition dictionary generation apparatus includes the module shown in FIG. 2 and generates a feature selection dictionary 503 and an identification dictionary 504 using the learning pattern DB 207. Further, the recognition apparatus includes the module shown in FIG. 3 and recognizes an input pattern using the feature selection dictionary 503 and the identification dictionary 504 generated by the recognition dictionary generation apparatus.

  In the present invention, in order to solve the conventional problem, in the learning phase, an identification function main part extraction unit that outputs a large classification feature conversion function and a large classification identification function to the feature selection dictionary 503 and the identification dictionary 504, respectively. The pattern recognition apparatus 101 includes 502.

  Further, according to the present invention, in the recognition phase, a large classification feature selection unit 601 that generates a feature for large classification using the feature selection dictionary 503 generated in the learning phase, and a large classification identification process using the identification dictionary 504 The pattern recognition apparatus 101 includes a large classification identifying unit 602 that executes a detailed classification and a detailed identifying unit 603 that performs a detailed identification process on a correct candidate category acquired in the large classification identifying process.

  More specifically, in the present invention, a function generated by limiting the detailed classification identification function to a low-dimensional partial feature space is used as the large classification identification function. The low-dimensional partial feature space is selected so as to well describe the behavior of the discriminant function for detailed classification. Therefore, the large classification discriminant function in the present invention can be regarded as an approximation of the detailed classification discriminant function.

  In the present invention, since a large classification discriminant function is created from a discriminant function for detailed classification created by an arbitrary learning algorithm, the pattern recognition apparatus 101 improves recognition accuracy by using a discriminating function with high accuracy. be able to. In addition, since the classification function in the present invention is a classification function for detailed classification limited to a partial feature space, it is possible to avoid the complexity of the configuration in which the learning algorithm is different between the large classification identification process and the detailed identification process. .

  Below, the process of each phase is demonstrated using FIG. First, the learning phase will be described.

  The pattern recognition apparatus 101 includes a pattern input unit 201, a feature extraction unit 202, a feature selection dictionary generation unit 203, a feature selection unit 204, an identification function generation unit 501, an identification function main part extraction unit 502, as a learning phase module and DB. A learning pattern DB 207, a feature selection dictionary 503, and an identification dictionary 504 are provided.

  In the learning phase, the pattern recognition apparatus 101 uses the learning pattern DB 207 to generate a feature selection dictionary 503 and an identification dictionary 504 that are used in the recognition phase.

  The learning pattern DB 207 is a set of patterns that are created in advance for learning and are given a correct answer label indicating an affiliation category. The learning pattern is created using the pattern acquisition device 104 or the like.

  The pattern is, for example, image data or audio data. The number of patterns is usually several tens or more, and may be tens of millions. For example, in the case of kanji recognition, a learning pattern DB 207 storing tens of millions or more patterns is used for learning.

  For example, the correct label may be expressed by associating a number with each recognition target category, or in the case of character recognition, a character code such as an EUC code, a JIS code, or an SJIS code may be used. In the processing to be described later, the correspondence between the pattern and the correct label is not lost so that the correct label of the pattern being processed can be understood. For example, a label indicating the category to which the pattern belongs may be recorded in the header portion of the pattern.

  The learning pattern DB 207, the feature selection dictionary 503, and the identification dictionary 504 are realized by the external storage device 107.

  The pattern input unit 201 acquires a pattern used for learning from the learning pattern DB 207, and outputs the acquired pattern to the feature extraction unit 202.

  The feature extraction unit 202 extracts an n-dimensional vector as a component effective for recognition from each pattern input from the pattern input unit 201. The n-dimensional vector generated at this time is called an n-dimensional feature. Thus, each pattern is expressed as an n-dimensional feature. The extracted n-dimensional features are output to the feature selection dictionary generation unit 203 and the feature selection unit 204.

  The subsequent feature selection dictionary generation unit 203 may require n-dimensional features of a plurality of patterns in order to generate the feature selection dictionary 503. In this case, every time an n-dimensional feature is required, the pattern input unit 201 and the feature extraction unit 202 may execute processing, and the feature selection dictionary generation unit 203 may acquire the required n-dimensional feature.

  Further, the pattern input unit 201 and the feature extraction unit 202 process all the learning patterns in advance to convert them into n-dimensional features, store the n-dimensional features in the external storage device 107 or the like, and the feature selection dictionary generation unit A method of acquiring n-dimensional features from the external storage device 107 every time 203 is required may be used.

  The pattern is expressed as an n-dimensional feature as the feature extraction unit 202 executes the process. With this processing, the pattern recognition apparatus 101 applies the same processing regardless of the type of the pattern because the pattern is expressed as an n-dimensional vector value regardless of whether the input data is sound or image. be able to.

  Here, as an example, processing of the feature extraction unit 202 in character recognition will be described.

  FIG. 5 is a flowchart illustrating an example of processing executed by the feature extraction unit 202 in character recognition according to the first embodiment of this invention.

  In input step 801, the feature extraction unit 202 captures the image output from the pattern input unit 201.

  In pre-processing step 802, the feature extraction unit 202 performs noise removal and blurring processing on the input image that has been taken in, and removes noise, blurring, and the like that cause obstacles to character recognition. For example, in the noise removal process, an isolated point having a size equal to or smaller than a certain threshold is removed.

  In the normalization step 803, the feature extraction unit 202 converts each of the preprocessed images into a fixed size image designated in advance. By this processing, the sizes of input images of various sizes can be made uniform, and processing depending on the image size can be unified.

  Examples of the normalization method include a linear normalization method, a non-linear normalization method, and a moment normalization method (see Non-Patent Document 1). For example, when the input image is a binary image, in the linear normalization method, a normalized image is generated by enlarging or reducing a partial image surrounded by a minimum rectangle surrounding a black pixel indicating a character portion to a fixed size image. Is done.

  In the character feature extraction step 804, the feature extraction unit 202 converts the image generated by normalization into an n-dimensional feature.

  As an example, the simplest pixel feature extraction will be described. In pixel feature extraction, the pixel value of each pixel is used as a feature. For example, when the input image is a 20 × 20 gray image, and the pixel value of each pixel is expressed by an integer value of 0 to 255, the number of pixels is 400, and the extracted features are 400-dimensional features. Each component of the 400-dimensional feature is a pixel value of 0 to 255 of the input image.

  In the output step 805, the feature extraction unit 202 outputs the converted n-dimensional feature to the feature selection dictionary generation unit 203 and the feature selection unit 204.

  The above is an example of the process of the feature extraction unit 202 in character recognition.

  Returning to the description of FIG.

  The feature selection dictionary generation unit 203 generates a feature selection dictionary used by the feature selection unit 204 described later to extract m-dimensional features from n-dimensional features.

  Here, first, the reason for extracting the m-dimensional feature from the n-dimensional feature will be described.

  The purpose of extracting the m-dimensional feature from the n-dimensional feature extracted by the feature extraction unit 202 is to realize high accuracy and high speed of the recognition process.

  First, speeding up will be described. The amount of calculation required when the pattern recognition apparatus 101 learns or recognizes is at least an order of the power of the number of dimensions.

  For example, in a calculation in which an n × n covariance matrix is used, a calculation amount in the order of the square of n is required. Further, when the order of the discriminant function is s, calculation of the discriminant function requires a calculation amount in the order of s-th power.

  Therefore, in order to reduce the amount of calculation, it is necessary to reduce feature components that contribute less to the identification process and to reduce the number of dimensions. Further, as the number of dimensions increases, the ratio of features with high correlation to each other increases, and an effect commensurate with the amount of calculation cannot be obtained. Therefore, it is effective to reduce the amount of calculation by integrating features having high correlation.

  Next, high accuracy will be described. When the pattern recognition apparatus 101 generates a discriminant function from a finite number of learning patterns in a high-dimensional space, an increase in the number of dimensions causes a decrease in accuracy. This is because the number of estimated parameters of the discriminant function increases as the number of dimensions increases, and parameter estimation using a finite number of learning patterns is statistically unreliable. Therefore, removing features that contribute little to the identification processing and appropriately reducing the number of dimensions of the features is effective in increasing the accuracy of the recognition processing.

  For the reasons described above, n-dimensional features are converted to m-dimensional features.

  The feature selection dictionary generation unit 203 uses the n-dimensional feature of the learning pattern generated by the feature extraction unit 202 to generate a conversion function that converts an n-dimensional feature into an m-dimensional feature. The generated conversion function is stored in the feature selection dictionary 503. In the case of linear conversion, the conversion function is represented by a matrix.

  For example, principal component analysis or linear discriminant method is used to generate the conversion function. The conversion into m-dimensional features may be performed by a function that is different for each category. In that case, a conversion function is created for each category.

  Here, generation of a feature selection function by principal component analysis will be described as an example. In the principal component analysis, first, the feature selection dictionary generation unit 203 calculates a covariance matrix of the learning pattern distribution. Next, the feature selection dictionary generation unit 203 selects m eigenvectors in descending order of the eigenvalues of the covariance matrix, and selects m-dimensional features whose components are m features obtained by projecting n-dimensional features onto the respective eigenvectors. To do. When m eigenvectors are denoted by pi (i = 1,..., m), the i-th component yi of the m-dimensional feature y is given by the inner product yi = x · pi of x and pi. Therefore, the conversion function generated using principal component analysis is represented by an m × n matrix Y having m eigenvectors as row vectors. Specifically, when an n-dimensional feature is represented as x and an m-dimensional feature is represented as y, the transformation matrix Y is represented as y = Yx.

The feature selection unit 204 converts an n-dimensional feature into an m-dimensional feature using a conversion function stored in the feature selection dictionary 503. The converted m-dimensional feature is output to the discriminant function generator 501. If the n-dimensional feature is x = (x1, x2,..., xn) and the m-dimensional feature after conversion is y = (y1, y2,..., ym), the conversion function is expressed as follows.
y1 = f1 (x1, x2,..., xn)
y2 = f2 (x1, x2,..., xn)
・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・
ym = fm (x1, x2,..., xn)
Further, in the case of linear conversion, y = Yx is expressed using an m × n conversion matrix Y.

When different conversion is performed for each category, the feature selection unit 204 calculates the m-dimensional feature yk = (yk1, yk2,..., Ykm) of the category k using the conversion function created for each category. To do. The conversion function of category k is expressed as follows.
yk1 = fk1 (x1, x2,..., xn)
yk2 = fk2 (x1, x2,..., xn)
・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・
ykm = fkm (x1, x2,..., xn)
In the case of linear conversion, yk = Ykx is expressed using an m × n conversion matrix Yk.

  The subsequent discriminant function generating unit 501 and discriminant function main part extracting unit 502 may require m-dimensional features of a plurality of patterns in order to generate a detailed discriminant function and a large class discriminant function.

  In this case, each time an m-dimensional feature is required, the pattern input unit 201, the feature extraction unit 202, and the feature selection unit 204 execute processing, and the discrimination function generation unit 501 and the discrimination function main part extraction unit 502 require m. A dimensional feature may be acquired.

  In addition, the pattern input unit 201, the feature extraction unit 202, and the feature selection unit 204 execute processing for all learning patterns in advance to convert them into m-dimensional features, and store the m-dimensional features in the external storage device 107 or the like. Alternatively, a method may be used in which the discriminant function generation unit 501 and the discriminant function main part extraction unit 502 are acquired from the external storage device 107 each time they are required.

  The discriminant function generation unit 501 acquires m-dimensional features and generates a detailed discriminant function uk (x) for calculating the similarity of the recognition target pattern with respect to the category k. The generated detailed discriminant function is output to the discriminant function main part extracting unit 502 and stored in the discriminating dictionary 504.

  The detailed identification function uk (x) is expressed as a function of x in order to unify the notation, but is actually a function depending on the m-dimensional feature y or yk. That is, the detailed discrimination function using an arbitrary function h is uk (x) = hk (y) = hk (f (x)) or uk (x) = hk (yk) = hk (fk (x) )It can be expressed as. The detailed discriminant function for the category k of the pattern is expressed as uk (x) = hk (y) using the m-dimensional feature y.

  When the feature selection unit 204 generates different m-dimensional features yk for each category, the detailed identification function of the identification function generation unit 501 is expressed as uk (x) = hk (yk).

  Algorithms used to generate the detailed discriminant function include nearest neighbor method, perceptron, improved projection distance method, modified quadratic discrimination, general learning vector quantization, polynomial network, or support vector machine.

  The discrimination function main part extraction unit 502 generates a major classification function vk in which the detailed discrimination function is limited to the L-dimensional partial feature space, and stores the generated major classification function vk in the identification dictionary 504. Further, the discriminant function main part extraction unit 502 generates a large classification feature conversion function g for converting an n-dimensional feature into an L-dimensional feature z in the L-dimensional partial feature space, and generates the generated major classification feature conversion. The function g is output to the feature selection dictionary 503.

  Note that the L-dimensional subspace is a subspace of the n-dimensional feature space and also a subspace of the m-dimensional feature space.

  The present invention is characterized by the discrimination function main part extraction unit 502. Details of the processing in the discrimination function main part extraction unit 502 will be described later.

  The above is the description of the learning phase process. Next, the recognition phase will be described with reference to FIG.

  The pattern recognition apparatus 101 includes, as a recognition phase module and DB, a pattern input unit 201, a feature extraction unit 202, a large classification feature selection unit 601, a large classification identification unit 602, a feature selection unit 204, a detailed identification unit 603, and a recognition result output. A unit 303, a recognition target pattern DB 304, a feature selection dictionary 503, and an identification dictionary 504.

  The pattern input unit 201 acquires a recognition target pattern and outputs the acquired recognition target pattern to the feature extraction unit 202. The recognition target pattern may be stored in advance in the recognition target pattern DB 304 and may be taken in from the recognition target pattern DB 304 or may be taken in directly from the pattern acquisition device 104 or the communication device 105. Note that the recognition target pattern DB 304 may be the external storage device 107, for example.

  The feature extraction unit 202 extracts n-dimensional features from the pattern input by the pattern input unit 201. The process in which the feature extraction unit 202 extracts n-dimensional features from the pattern is the same as the learning phase. The extracted n-dimensional features are output to the large classification feature selection unit 601 and the feature selection unit 204.

  The major classification feature selection unit 601 converts an n-dimensional feature into an L-dimensional feature z using the major classification conversion function g stored in the feature selection dictionary 503. Here, when the n-dimensional feature x and the large classification conversion function g are used, the L-dimensional feature z is expressed as z = g (x). The converted L-dimensional feature is output to the large classification identifying unit 602.

  The large classification identifying unit 602 uses the large classification identification function vk stored in the identification dictionary 504 to calculate a rough similarity for each category of the pattern. Further, the large classification identifying unit 602 calculates a category having a high similarity as a correct candidate category using the calculated similarity. Note that the correct answer candidate categories are calculated by the number designated by the user. For example, in the case of Kanji recognition for character types of several thousand categories, the correct answer candidates are limited to about several tens of categories.

  The large classification identifying unit 602 outputs the correct answer candidate category to the feature selecting unit 204. Details of the processing executed by the large classification identifying unit 602 will be described later.

  The feature selection unit 204 uses the conversion function f stored in the feature selection dictionary 503 to convert the n-dimensional feature into an m-dimensional feature. The process of converting n-dimensional features into m-dimensional features is the same as in the learning phase. When different m-dimensional features are extracted for each category, each correct candidate category obtained by the process executed by the large classification identifying unit 602 is converted into an m-dimensional feature. The converted m-dimensional feature is output to the detailed identification unit 603. The correct answer candidate category is also output to the detail identifying unit 603.

  The detailed identification unit 603 calculates the degree of similarity of the authentication target pattern with respect to the correct candidate category using the detailed identification function. The calculated similarity to the correct candidate category is output to the recognition result output unit 303. Details of the processing executed by the detail identifying unit 603 will be described later.

  The recognition result output unit 303 outputs the final recognition result by using the similarity of the authentication target pattern with the calculated correct candidate category. Usually, the recognition result output unit 303 outputs the category having the highest similarity as the recognition result.

  In addition, the recognition result output unit 303 may also output the second and subsequent categories of similarity as second candidates, third candidates, and the like.

  When the similarity is smaller than the specified threshold value, the recognition result output unit 303 may reject it as not corresponding to any category. For example, in the case of digit recognition, this applies to the case where a kanji is input.

  In addition, when the difference between the first and second similarities is smaller than the specified threshold, the recognition result output unit 303 may reject it because it is difficult to determine which category it is. For example, it is difficult to discriminate depending on the font such as alphabet “I” or “l” and number “1”.

  The recognition result is displayed on the display device 103 and transmitted to the outside using the communication device 105 or output to the external storage device 107.

  Hereinafter, the details of processing executed by the discrimination function main part extraction unit 502, the major classification feature selection unit 601, the major classification identification unit 602, and the detailed identification unit 603, which are features of the present invention, will be described.

  First, a method in which the discriminant function main part extraction unit 502 generates a major classification function vk and a major classification feature conversion function will be described.

  In the following description, a case will be described in which the feature selection unit 204 extracts different m-dimensional features yk for each category. When the feature selection unit 204 extracts the m-dimensional feature y that does not depend on the category, the discriminant function main part extraction unit 502 may perform the same processing by replacing yk with y.

  Further, as a condition in the present embodiment, the feature selection unit 204 converts an n-dimensional feature into an m-dimensional feature by a linear function conversion yk = Ykx, and the detailed identification function is a quadratic as shown in [Equation 1]. Let us assume the following function.

  First, after describing the outline of the process executed by the discriminant function main part extraction unit 502, each process shown in FIG. 4 will be described.

  The discriminant function main part extraction unit 502 generates a large class discriminant function vk (x) in which the detailed discriminant function uk (x) is limited to the L-dimensional partial feature space. In the present embodiment, as the large classification discriminant function vk (x), the detailed discriminant function uk (x) is well approximated and can be calculated at high speed.

  The discriminant function main part extraction unit 502 first selects feature axes that characterize the detailed discriminant function uk (x) of each category, and then sets the importance of each selected feature axis.

  Next, the discriminant function main part extraction unit 502 calculates a main axis for integrating feature axes based on the importance. Finally, the discriminant function main part extraction unit 502 generates a function in which the detailed discriminant function uk (x) is limited to the main axis, and outputs the function as the large class discriminant function vk.

  As can be seen from the detailed discriminant function shown in [Equation 1], the calculation amount of the secondary discriminant function is the order of the square of the dimension number m of the feature space (the number of terms in [Equation 1] is the dimension number m Square order). Therefore, by limiting the number of dimensions of the feature space, it is possible to reduce the amount of calculation of the detailed identification function.

  The large classification discriminating function in which the dependent region of the detailed classification discriminating function is limited to the L-dimensional partial feature space can be calculated faster than the detailed classification discriminating function. In the present embodiment, a method will be described in which a detailed classification discriminant function is a function that is limited to an L-dimensional partial feature space of an m-dimensional feature space as a major classification discriminant function.

  First, the detailed identification function shown in [Formula 1] is modified. Here, the matrix Wk and the vector wk are defined by [Equation 2] and [Equation 3], respectively.

  [Equation 1] is transformed into [Equation 5] using [Equation 2], [Equation 3], and [Equation 4].

  Since Wk is a symmetric matrix, there is an arbitrary unitary matrix Pk, and Wk can be diagonalized using Pk as shown in [Formula 6]. Here, as the diagonal component, as shown in [Equation 7], Pk that is arranged in descending order of the absolute value of the diagonal component is used.

  Since Pk is a unitary matrix, [Equation 5] can be transformed into [Equation 8].

  Further, when the conversion function yk = Ykx is used, uk (yk) can be re-expressed as a function of x, and [Equation 8] is expressed as shown in [Equation 9].

  When the row vector of the matrix PkYk is replaced by qki (i = 1,..., M) as shown in [Equation 10] and Pkwk is given as [Equation 11], the discriminant function main part extraction unit 502 [Equation 9] is transformed into [Equation 12].

That is, if the matrix Pk that diagonalizes Wk can be obtained, [Equation 1] is transformed into [Equation 12] using [Equation 6], [Equation 10], and [Equation 11]. . In order to obtain the matrix Pk from Wk, the eigenvalue problem of the symmetric matrix Wk may be solved.

  In the present embodiment, qki is used as a feature axis that characterizes the detailed discrimination function uk.

  Next, the importance degree hki of each feature axis qki is set. There are various methods for determining the importance, but the simplest method is to determine the magnitude of the coefficient of the term including the feature axis qki as the importance as shown in [Equation 17].

  Next, a calculation method of a main axis that integrates selected feature axes will be described.

  Here, consider a case where the n-dimensional vector a is selected as the integrated axis. At this time, the projection length of the feature axis qki onto the vector a is given by the inner product a · qki.

  It can be considered that as the value of the inner product is larger, the information amount of the feature axis qki lost by the integration into the vector a is smaller. Therefore, as shown in [Equation 18], it can be said that the larger the sum of the inner product value multiplied by the importance of the feature axis, the smaller the amount of information in the feature axis direction lost by the integration into the vector a.

Here, K is the number of categories. In the present embodiment, a vector a having a large amount as shown in [Equation 18] is selected as the main axis.

  Furthermore, [Equation 18] is transformed into [Equation 19].

  Here, the parenthesis in the result of the equation modification of [Equation 19] is defined as a matrix Q like [Equation 20].

The matrix Q is an n × n symmetric matrix. Accordingly, when the eigenvalues of the matrix Q are d1 ≧ d2 ≧... Dn in descending order and the normalized eigenvectors corresponding to the eigenvalues are u1, u2,... Un, the unitary matrix U is expressed as [Equation 22]. In other words, the matrix Q can be diagonalized as [Equation 21].

  [Equation 19] is transformed into [Equation 23] using [Equation 21].

Here, di is an evaluation value when a plurality of feature axes are integrated into one main axis. A larger value of di indicates that a plurality of feature axes are more integrated.

  Since u1, u2,..., un are orthogonal systems, the principal axis with the largest value of [Equation 18] from [Equation 23] is the eigenvector a = u1 corresponding to the largest eigenvalue. Hereinafter, the second major axis is the eigenvector a = u2 corresponding to the second eigenvalue, the third major axis is the eigenvector a = u3 corresponding to the third eigenvalue, and the Lth major axis is the matrix Q The eigenvector uL corresponding to the L-th largest eigenvalue.

  Finally, a large class discriminant function in which the detailed discriminant function is limited to the main axis is obtained. A vector obtained by limiting the n-dimensional feature x to the j-th major axis uj is transformed as shown in [Equation 24].

  When the vector x of the detailed discriminant function shown in [Equation 12] is replaced with the restriction vector shown in [Equation 24], it is transformed as [Equation 25].

  Here, when [Equation 26], [Equation 27], [Equation 28], and [Equation 29] are set, [Equation 25] is transformed into [Equation 30].

  Therefore, the large classification discriminant function is expressed as [Equation 30] and is expressed as z = UL (x) using [Equation 31].

  Therefore, the conversion function used for feature selection for large classification is [Equation 31].

  Using the equations described above, the discriminant function main part extraction unit 502 obtains a major classification discriminant function vk as shown in [Equation 30] and a major classification feature conversion function UL as shown in [Equation 31]. Can be generated.

  The details of the processing executed by the discriminant function main part extraction unit 502 will be described below.

  FIG. 4 is a flowchart illustrating processing executed by the discriminant function main part extraction unit 502 according to the first embodiment of this invention.

  In the discrimination function input step 701, the discrimination function main part extraction unit 502 acquires a detailed discrimination function from the discrimination function generation unit 501.

  In the feature axis selection step 702, the discriminant function main part extraction unit 502 selects a feature axis that is a reference for calculating the main axis.

  Specifically, the discriminant function main part extraction unit 502 transforms the detailed discriminant function shown in [Equation 1] into [Equation 12]. The modification is obtained by defining a symmetric matrix Wk as [Equation 2], obtaining a matrix Pk having an eigenvector of Wk as a row vector, and using [Equation 6], [Equation 10], and [Equation 11]. .

  The discriminant function main part extraction unit 502 selects qki as a feature axis.

  In the axis importance setting step 703, the discrimination function main part extraction unit 502 calculates the importance of each feature axis using the feature axis and the detailed discrimination function.

  The simplest method of calculating the importance of the feature axis is a method of defining the importance hki of the feature axis qki as shown in [Equation 17]. There are various ways of defining the importance, and some other examples will be described later.

  In the main axis calculation step 704, the discriminant function main part extracting unit 502 integrates the feature axes using the calculated importance and feature axes, and calculates L main axes.

  Specifically, the L major axes are calculated as upper L normalized eigenvectors u1, u2,..., UL having a large eigenvalue of the n × n matrix Q shown in [Equation 20].

  In the large classification discriminant function generation step 705, the discriminant function main part extraction unit 502 uses the main axis and the detailed discriminant function to detail the L-dimensional feature space generated by the L eigenvectors u1, u2,. A large class discriminant function with a limited discriminant function is generated.

  Specifically, the large classification discriminant function is given as [Equation 30] using [Equation 26], [Equation 27], [Equation 28], and [Equation 29]. A feature selection function for large classification is given by [Equation 31].

  In the output step 706, the discriminant function main part extraction unit 502 stores the feature selection function [Equation 31] for the major classification discriminant function in the feature selection dictionary 503, and also identifies the major classification discriminant function [Equation 30]. Store in dictionary 504.

  The above is the detailed description of the processing executed by the discriminant function main part extraction unit 502.

  Next, processing executed by the major classification feature selection unit 601, the major classification identification unit 602, and the detailed identification unit 603 in the recognition phase will be described.

  The major classification feature selection unit 601 extracts the L-dimensional feature z [Equation 29] using the feature selection function [Equation 31] for the major classification identification function stored in the feature selection dictionary 503, and the major classification identification unit To 602.

  The large classification identifying unit 602 calculates the similarity using the large classification identification function [Equation 30] stored in the identification dictionary 504. Based on the calculated similarity, the large classification identifying unit 602 calculates correct candidate categories for the number specified by the user in descending order of similarity. The calculated correct candidate category is output to the feature selection unit 204.

  The detailed identification unit 603 calculates the similarity to the correct candidate category using the detailed identification function uk (x) stored in the feature selection dictionary 503, and outputs the similarity to the calculated correct candidate category as a recognition result. The data is output to the unit 303.

  Hereinafter, an example of a method for setting the importance of the feature axis in the axis importance setting step 703 will be described.

Example 1
The importance of the feature axis qki is defined as [Equation 17], assuming that it is the magnitude of the absolute values of the coefficients λkii and ζki of [Equation 12].

Example 2
In [Equation 12], considering the difference in the degree between the coefficient λkii and the coefficient ζki, the importance of the feature axis qki is defined as in [Equation 32].

Example 3
When determining the importance of the feature axis qki, not only the coefficients λkii and ζki but also the scale of the amount of variation qki · x of x in the qki direction is taken into consideration. The scale of the fluctuation amount is calculated by the variance value of x in the qki direction.

  Assuming that there are N learning patterns and the average vector m is [Expression 33], the covariance matrix Σ of the distribution of xi (i = 1,..., N) is given by [Expression 34].

  The variance value vki of x in the qki direction can be calculated by [Equation 35].

Thus, the importance of the feature axis qki is defined by [Equation 36], for example. Also, a method of defining the importance of the feature axis qki using [Equation 37] or [Equation 38] may be considered.

Example 4
As shown in [Equation 39], the discriminant function [Equation 12] may be defined by defining the variance value of the portion limited to the feature axis qki as the importance. The variance value σki is given as [Equation 41] using [Equation 40]. The importance is defined as hki = σki.

(Second Embodiment)
In the first embodiment, as shown in [Equation 12], the discriminant function is assumed to be a function of second order or lower, and the feature selection function is assumed to be a linear function. In the second embodiment, the feature selection function f (x) or fk (x) is not limited to a linear function. In the second embodiment, it is assumed that the discriminant function is a function of second order or lower.

  Since the configuration of the pattern recognition apparatus 101 in the second embodiment and the processing executed by each module are the same as those in the first embodiment, the description thereof is omitted. Hereinafter, the difference from the first embodiment will be mainly described.

  In the second embodiment, the processing executed by the discrimination function main part extraction unit 502 in the learning phase, and the processing executed by each of the major classification feature selection unit 601, the major classification identification unit 602, and the detailed identification unit 603 in the recognition phase. Is different.

  Hereinafter, the identification function main part extraction unit 502, the major classification feature selection unit 601, the major classification identification unit 602, and the detailed identification unit 603 according to the second embodiment will be described.

  First, the process of the discriminant function main part extraction unit 502 in the learning phase will be described.

  The processing executed by the discriminant function main part extraction unit 502 is the same as that shown in FIG. 4, but the specific processing is different.

  In the discrimination function input step 701, the discrimination function main part extraction unit 502 acquires a detailed discrimination function from the discrimination function generation unit 501.

  In a feature axis selection step 702, the discriminant function main part extraction unit 502 selects a feature axis that is a reference for main axis selection.

  Specifically, the discriminant function main part extraction unit 502 transforms the detailed discriminant function shown in [Equation 1] into [Equation 45].

The modification is obtained by defining the symmetric matrix Wk as [Equation 2], obtaining a matrix Pk having the eigenvector of Wk as a row vector, and using [Equation 6], [Equation 44], and [Equation 11]. .

  The discriminant function main part extraction unit 502 selects qki as a feature axis. In the first embodiment, qki is an n-dimensional vector, but in the second embodiment, it is an m-dimensional vector.

  In the axis importance setting step 703, the discrimination function main part extraction unit 502 calculates the importance of each feature axis using the feature axis and the detailed discrimination function.

  There are various methods for calculating the importance. For example, the method shown in the first embodiment can be used. In this case, x is replaced with y.

  In the main axis calculation step 704, the discriminant function main part extracting unit 502 integrates the feature axes using the calculated importance and feature axes, and calculates L main axes.

  Specifically, L major axes are obtained as upper L normalized eigenvectors u1, u2,..., UL having large eigenvalues of the m × m matrix shown in [Equation 20]. In the first embodiment, the matrix shown in [Equation 20] is an n × n matrix and ui is an n-dimensional vector, whereas in the second embodiment, the matrix shown in [Equation 20]. Is an m × m matrix and ui is an m-dimensional vector.

  In the large classification discriminant function generation step 705, the discriminant function main part extraction unit 502 uses the main axis and the detailed discriminant function to detail the L-dimensional feature space generated by the L eigenvectors u1, u2,. A large class discriminant function with a limited discriminant function is generated.

  Specifically, the large classification identification function is given as [Equation 30] using [Equation 26], [Equation 27], [Equation 28], and [Equation 46]. Further, the feature selection function for the major classification function is given as [Equation 31].

  In the output step 706, the discrimination function main part extraction unit 502 stores the feature selection function for large classification [Equation 31] in the feature selection dictionary 503, and also stores the large classification discrimination function [Equation 30] in the discrimination dictionary 504. To store.

  Next, processing executed by the major classification feature selection unit 601, major classification identification unit 602, and detailed identification unit 603 in the recognition phase will be described.

  The major classification feature selection unit 601 extracts the L-dimensional feature z [Equation 46] using the major classification feature selection function [Equation 31] stored in the feature selection dictionary 503, and sends it to the major classification identification unit 602. Output.

  The large classification identifying unit 602 calculates the similarity using the large classification identification function [Equation 30] stored in the identification dictionary 504. Based on the calculated similarity vk (x), the large classification identifying unit 602 calculates correct candidate categories for the number specified by the user in descending order of similarity. The calculated correct candidate category is output to the feature selection unit 204.

  The detailed identification unit 603 calculates the similarity to the correct candidate category using the detailed identification function uk (x) stored in the feature selection dictionary 503, and recognizes the calculated similarity to the correct candidate category as a recognition result output unit. It outputs to 303.

(Third embodiment)
In the first embodiment, as shown in [Equation 12], the discriminant function is assumed to be a function of second order or lower, and the feature selection function is assumed to be a linear function. In the third embodiment, a case will be described in which the discriminant function is not limited to a function of second order or lower. Also in the third embodiment, the feature selection function is assumed to be a linear function.

  Since the configuration of the pattern recognition apparatus 101 in the third embodiment and the processing executed by each module are the same as those in the first embodiment, description thereof is omitted. Hereinafter, the difference from the first embodiment will be mainly described.

  In the third embodiment, processing executed by the identification function main part extraction unit 502 in the learning phase, and processing executed by each of the large classification feature selection unit 601, the large classification identification unit 602, and the detailed identification unit 603 in the recognition phase. Is different.

  Hereinafter, the identification function main part extraction unit 502, the major classification feature selection unit 601, the major classification identification unit 602, and the detailed identification unit 603 according to the third embodiment will be described.

  First, the process of the discriminant function main part extraction unit 502 in the learning phase will be described.

  The processing executed by the discriminant function main part extraction unit 502 is the same as that shown in FIG. 4, but the specific processing is different.

  In the discrimination function input step 701, the discrimination function main part extraction unit 502 acquires a detailed discrimination function from the discrimination function generation unit 501.

  In a feature axis selection step 702, the discriminant function main part extraction unit 502 selects a feature axis that is a reference for main axis selection.

  Specifically, the discriminant function main part extraction unit 502 selects m-dimensional vectors random as the number specified by the user as feature axes. Here, the feature axes are M selected vectors and are represented as aki (i = 1,..., M).

  In the axis importance setting step 703, the discrimination function main part extraction unit 502 calculates the importance of each feature axis using the feature axis and the detailed discrimination function.

  The following calculation methods can be considered. Hereinafter, in the case where the m-dimensional feature does not depend on the category k, it may be considered that the m-dimensional feature yk and the subscript of the conversion function Y are omitted.

  First, the m-dimensional vector aki is converted into an expression in the n-dimensional feature space. Ak is defined as [Equation 47], and AkYk is set as [Equation 48].

At this time, the vector qki is an expression in the n-dimensional feature space of aki.

  The variance of discriminant function values in the characteristic axis qki direction is defined as the importance. The function uki shown in [Equation 47] is obtained by limiting the detailed identification function to the qki axis. Specifically, when N learning patterns are set to xi, the variance value σki of the value of this function is calculated by [Equation 41] using [Equation 40]. This may be set as hki = σki.

  In the main axis calculation step 704, the discriminant function main part extracting unit 502 integrates the feature axes using the calculated importance and feature axes, and calculates L main axes.

  Specifically, the L major axes are obtained as upper L eigenvectors u1, u2,..., UL having a large eigenvalue of the n × n matrix shown in [Equation 20].

  In the large classification discriminant function generation step 705, the discriminant function main part extraction unit 502 uses the main axis and the detailed discriminant function to detail the L-dimensional feature space generated by the L eigenvectors u1, u2,. Generated as a large class discriminant function with limited discriminant function.

  Specifically, the large classification identification function is generated as shown in [Equation 50].

Further, the feature selection function for the large classification identification function is given by a matrix as shown in [Equation 53] using the matrix UL of [Equation 31].

  In the output step 706, the discriminant function main part extraction unit 502 stores the feature selection function [Equation 53] for large classification in the feature selection dictionary 503, and also stores the large classification discrimination function [Equation 50] in the discrimination dictionary 504. To store.

  Next, processing executed by the major classification feature selection unit 601, major classification identification unit 602, and detailed identification unit 603 in the recognition phase will be described.

  The major classification feature selection unit 601 extracts the L-dimensional feature z [Equation 46] using the major classification feature selection function [Equation 53] stored in the feature selection dictionary 503, and sends it to the major classification identification unit 602. Output.

  The large classification identifying unit 602 calculates the similarity using the large classification identification function [Equation 50] stored in the identification dictionary 504. Based on the calculated similarity vk (x), the large classification identifying unit 602 calculates correct candidate categories for the number specified by the user in descending order of similarity. The calculated correct candidate category is output to the feature selection unit 204.

  The detailed identification unit 603 calculates the similarity to the correct candidate category using the identification function uk (x) stored in the feature selection dictionary 503, and uses the recognition result output unit 303 to calculate the similarity to the calculated correct candidate category. Output to.

(Fourth embodiment)
In the first embodiment, as shown in [Equation 12], the discriminant function is assumed to be a function of second order or lower, and the feature selection function is assumed to be a linear function. In the fourth embodiment, a case where the discriminant function is not limited to a function of second order or lower will be described. Furthermore, in the fourth embodiment, the feature selection function is not limited to a linear function.

  Since the configuration of the pattern recognition apparatus 101 in the fourth embodiment and the processing executed by each module are the same as those in the first embodiment, the description thereof is omitted. Hereinafter, the difference from the first embodiment will be mainly described.

  In the fourth embodiment, processing executed by the identification function main part extraction unit 502 in the learning phase, and processing executed by each of the large classification feature selection unit 601, the large classification identification unit 602, and the detailed identification unit 603 in the recognition phase. Is different.

  Hereinafter, the identification function main part extraction unit 502, the major classification feature selection unit 601, the major classification identification unit 602, and the detailed identification unit 603 according to the fourth embodiment will be described.

  First, the process of the discriminant function main part extraction unit 502 in the learning phase will be described.

  The processing executed by the discriminant function main part extraction unit 502 is the same as that shown in FIG. 4, but the specific processing is different.

  In the discrimination function input step 701, the discrimination function main part extraction unit 502 acquires a detailed discrimination function from the discrimination function generation unit 501.

  In a feature axis selection step 702, the discriminant function main part extraction unit 502 selects a feature axis that is a reference for main axis selection.

  Specifically, the discriminant function main part extraction unit 502 selects m-dimensional vectors random as the number specified by the user as feature axes. Here, the wiretapping axis is a vector selected from M pieces, and is represented as aki (i = 1,..., M).

  In the axis importance setting step 703, the discrimination function main part extraction unit 502 calculates the importance of each feature axis using the feature axis and the detailed discrimination function.

  The importance calculation method is as follows. Hereinafter, in the case where the m-dimensional feature does not depend on the category k, it may be considered that the m-dimensional feature yk and the subscript of the conversion function Y are omitted.

  The variance of the discriminant function value in the characteristic axis qki direction is taken as the importance. The function uki shown in [Formula 51] is obtained by limiting the detailed identification function to the qki axis.

Specifically, when N learning patterns are set to xi, the variance value σki of the value of this function is calculated by [Equation 52] using [Equation 51]. This may be set as hki = σki.

  In the main axis calculation step 704, the discriminant function main part extraction unit 502 integrates the feature axes using the calculated importance and feature axes, and selects L main axes.

  Specifically, the L major axes are obtained as upper L eigenvectors u1, u2,..., UL having a large eigenvalue of the m × m matrix shown in [Equation 20].

  Thereafter, the discrimination function main part extraction unit 502 proceeds to the large classification discrimination function generation step 705.

  In the large classification discriminant function generation step 705, the discriminant function main part extraction unit 502 uses the main axis and the detailed discriminant function to detail the L-dimensional feature space generated by the L eigenvectors u1, u2,. Generated as a large class discriminant function with limited discriminant function.

  Specifically, the large classification discriminant function is given as [Equation 52]. The feature selection function for the major classification function is given by the matrix of [Equation 31].

  In the output step 706, the discriminant function main part extraction unit 502 stores the large classification feature selection function [Equation 31] in the feature selection dictionary 503 and the large classification discriminant function [Equation 52]. To store.

  Next, processing executed by the major classification feature selection unit 601, major classification identification unit 602, and detailed identification unit 603 in the recognition phase will be described.

  The major classification feature selection unit 601 extracts the L-dimensional feature z [Equation 46] using the major classification feature selection function [Equation 31] stored in the feature selection dictionary 503, and sends it to the major classification identification unit 602. Output.

  The large classification identifying unit 602 calculates the similarity using the large classification identification function [Equation 52] stored in the identification dictionary 504. Based on the calculated similarity vk (x), the large classification identifying unit 602 calculates correct candidate categories for the number specified by the user in descending order of similarity. The calculated correct candidate category is output to the feature selection unit 204.

  The detailed identification unit 603 calculates the similarity to the correct candidate category using the identification function uk (x) stored in the feature selection dictionary 503, and uses the calculated similarity to the correct candidate category to the recognition result output unit. Output.

  According to one aspect of the present invention, since the large classification discriminant function is generated as the limit function of the detailed discriminant function, the accuracy of the recognition process can be maintained and the speed can be increased.

  In addition, since the large classification function in the present invention is generated as a limit function of the detailed classification function, it is not necessary to use different learning algorithms for the large classification process and the detailed classification process. Therefore, complexity of the configuration can be avoided.

  In addition, according to one aspect of the present invention, since a large classification discriminant function is generated from a detailed discriminant function created by an arbitrary learning algorithm, it is possible to improve recognition accuracy by using a high-precision discriminant function. Can do.

101 Pattern Recognition Device 102 Input Device 103 Display Device 104 Pattern Acquisition Device 105 Communication Device 106 Arithmetic Device (CPU)
107 External storage device (HDD, memory)
201 pattern input unit 202 feature extraction unit 203 feature selection dictionary generation unit 204 feature selection unit 205 identification dictionary generation unit 207 learning pattern DB
208 Feature Selection Dictionary 209 Identification Dictionary 301 Major Classification Identification Unit 302 Detailed Identification Unit 303 Recognition Result Output Unit 304 Recognition Target Pattern DB
501 Discrimination function generation unit 502 Discrimination function main part extraction unit 503 Feature selection dictionary 504 Discrimination dictionary 601 Major classification feature selection unit 602 Major classification identification unit 603 Detailed classification unit 701 Discrimination function input step 702 Feature axis selection step 703 Axis importance Setting step 704 Main axis calculation step 705 Major classification function generation step 706 Output step 801 Input step 802 Preprocessing step 803 Normalization step 804 Character feature extraction step 805 Output step

Claims (24)

A dictionary recognition apparatus for pattern recognition comprising a processor and a storage medium connected to the processor,
The storage medium stores a learning pattern database composed of a plurality of learning patterns,
The pattern recognition dictionary generation device includes:
A pattern input unit for acquiring each of the learning patterns as one category from the learning pattern database;
A feature extraction unit that extracts n-dimensional features for each of the acquired categories;
Using the extracted n-dimensional feature, a feature selection function for converting the n-dimensional feature into an m-dimensional feature that is a dimension equal to or less than the n-dimension is generated, and the generated feature selection function is used as a feature selection dictionary. A feature selection dictionary generator for storing in the storage medium;
A feature selection unit that converts the extracted n-dimensional feature into the m-dimensional feature using the feature selection function;
Using the converted m-dimensional feature, a detailed identification function on the m-dimensional feature space for calculating the similarity of the recognition target pattern for each category is generated, and the generated detailed identification function is used as an identification dictionary. A discriminant function generator for storing in the storage medium as
Large classification that transforms the m-dimensional feature into an L-dimensional feature on the L-dimensional feature space that is a dimension less than the m-dimension, is a subspace of the n-dimensional feature space, and is a subspace of the m-dimensional feature space A large classification for generating a feature selection function and calculating the similarity of the pattern to be authenticated for each category on the L-dimensional feature space by converting the detailed identification function as a function on the L-dimensional feature space An identification function is generated, the generated major classification feature selection function is stored in the storage medium as the feature selection dictionary, and the generated major classification identification function is stored in the storage medium as the identification dictionary A function main part extractor;
A pattern recognition dictionary generating apparatus comprising: The discriminant function main part extraction unit includes:
Performing a detailed identification function acquisition step of acquiring the detailed function;
Performing a feature axis selection step of selecting M feature axes from the n-dimensional feature space using the acquired detailed identification function;
Performing an axis importance determining step for calculating the importance of the feature axis;
Performing a main axis calculating step of calculating the L main axes by integrating the feature axes;
The large classification discriminant function generation step of generating the large classification discriminant function by converting the detailed discriminant function as a function on the L-dimensional feature space generated by the main axis is performed. The pattern recognition dictionary generation device described. The feature selection function is a linear function, and the detailed discriminant function uk (x) for the n-dimensional feature x is a second-order or lower-order polynomial function expressed by Equation 1,
The feature axis selection step includes:
The discriminant function main part extraction unit uses the n-dimensional feature or the m-dimensional feature y to characterize the vector qki obtained by transforming the detailed discriminant function uk (x) as shown in Equation 2 or Equation 3. Including selecting as an axis,
The axis importance determination step is
The pattern according to claim 2, wherein the discriminant function main part extraction unit includes a step of setting the importance of the feature axis qki using the function hki of the coefficient λkii and the coefficient ζki in Expression 2 or 3. A recognition dictionary generator.

The main axis calculation step includes:
The discriminant function main part extraction unit calculating eigenvalues of the matrix Q shown in Formula 4 generated from the feature axis qki and the axis importance hki;
The discriminant function main part extraction unit selects L eigenvectors in descending order of the calculated eigenvalues among eigenvectors of the matrix Q, and calculates the selected eigenvectors as main axes. The pattern recognition dictionary generation device according to claim 3, further comprising:

The pattern recognition dictionary generation device includes an input unit for a user to operate the pattern recognition dictionary generation device,
In the feature axis selection step, the discriminant function main part extraction unit randomly selects M arbitrary vectors as the feature axes, or selects M vectors based on an instruction from the user. The pattern recognition dictionary generation device according to claim 2, further comprising a step of selecting as a pattern recognition dictionary.   In the axis importance determination step, the discriminant function main part extraction unit has transformed the detailed discriminant function for the N learning patterns xi as a function on the M-dimensional space generated by the feature axis qki. 3. The pattern recognition dictionary generation device according to claim 2, wherein a function d (xi) is calculated, and the axis importance h is calculated as a function of a variance value σ of the calculated function d (xi). . A pattern recognition apparatus comprising a processor and a storage medium connected to the processor,
The storage medium is
A recognition target pattern database composed of a plurality of recognition target patterns, and
A feature selection function that converts an n-dimensional feature extracted from the recognition target pattern into an m-dimensional feature that is a dimension of n or less, and a subspace of the n-dimensional feature space that is a dimension of m or less And a feature selection dictionary that stores a large-class feature selection function for converting the m-dimensional feature into an L-dimensional feature on an L-dimensional feature space that is a subspace of the m-dimensional feature space;
A detailed discriminant function on the m-dimensional feature space for calculating the similarity of the recognition target pattern to the correct candidate pattern, and the detailed discriminant function converted as a function on the L-dimensional feature space, and the L-dimensional feature An identification dictionary that stores a large classification identification function for calculating the similarity of the recognition target pattern to each correct answer candidate in space;
Store
The pattern recognition device includes:
A pattern input unit for obtaining the recognition target pattern from the recognition target pattern database;
A feature extraction unit for extracting an n-dimensional feature of the acquired recognition target pattern;
A major classification feature selection unit that converts the extracted n-dimensional feature into the L-dimensional feature using the major classification feature selection function;
A large classification identifying unit that calculates the similarity of the pattern to be recognized with respect to the correct candidate pattern using the converted L-dimensional feature and the large classification identifying function, and selects one or more correct candidate patterns; ,
A feature selection unit that converts the n-dimensional feature into the m-dimensional feature using the feature selection function;
A detailed identification unit that calculates the similarity of the recognition target pattern to the selected correct candidate pattern using the detailed identification function and the converted m-dimensional feature;
A recognition result output unit that outputs a recognition result for the recognition target pattern based on the calculated similarity;
A pattern recognition apparatus comprising: The major classification function is
Performing a detailed identification function acquisition step of acquiring the detailed function;
Performing a feature axis selection step of selecting M feature axes from the n-dimensional feature space using the acquired detailed identification function;
Performing an axis importance determining step for calculating the importance of the feature axis;
Performing a main axis calculating step of calculating the L main axes by integrating the feature axes;
The major axis is generated by executing a large classification function generating step for generating the large classification function by converting the detailed classification function as a function on the L-dimensional feature space generated by the main axis. The pattern recognition apparatus according to claim 7. The feature selection function is a linear function, and the detailed discriminant function uk (x) for the n-dimensional feature x is a second-order or lower polynomial function shown in Equation 5,
In the feature axis selection step, using the n-dimensional feature or the m-dimensional feature y, the vector qki obtained by transforming the detailed identification function uk (x) as shown in Equation 6 or Equation 7 is used as the feature axis. Selected
9. The pattern recognition apparatus according to claim 8, wherein, in the axis importance determination step, the importance of the feature axis qki is set using the function hki of the coefficient λkii and the coefficient ζki in Expression 6 or 7.

In the principal axis calculation step, the eigenvalue of the matrix Q shown in Equation 8 generated from the feature axis qki and the axis importance hki is calculated, and among the eigenvectors of the matrix Q, the calculated eigenvalue is large The pattern recognition apparatus according to claim 9, wherein L eigenvectors are selected in order, and the selected eigenvectors are calculated as main axes.

  In the feature axis selection step, M arbitrary vectors are randomly selected as the feature axes, or M vectors are selected as the feature axes based on a preset instruction. Item 9. The pattern recognition device according to Item 8.   In the axis importance determination step, the detailed discriminant function for the N patterns xi is calculated as a function d (xi) transformed as a function in the M-dimensional space generated by the feature axis qki, and is calculated. The pattern recognition apparatus according to claim 8, wherein the axis importance h is calculated as a function of a variance value σ of the function d (xi). A pattern recognition apparatus comprising a processor and a storage medium connected to the processor,
The storage medium stores a learning pattern database composed of a plurality of learning patterns, and a recognition target pattern database composed of a plurality of recognition target patterns,
The pattern recognition dictionary generation device includes:
A pattern input unit that acquires each of the learning patterns as one category from the learning pattern database, or a pattern input unit that acquires the recognition target pattern from the recognition target pattern database;
Extracting a n-dimensional feature for each of the acquired categories, and extracting an n-dimensional feature of the acquired recognition target pattern;
Using the extracted n-dimensional feature, a feature selection function that converts the n-dimensional feature into an m-dimensional feature that is a dimension equal to or less than the n-dimensional feature is generated, and the generated feature selection function is used as a feature selection dictionary. A feature selection dictionary generating unit for storing in the storage medium as
A feature selection unit that converts the n-dimensional feature into the m-dimensional feature using the feature selection function;
Using the converted m-dimensional feature, a detailed identification function on an m-dimensional feature space for calculating a pattern similarity to each category is generated, and the generated detailed identification function is stored as an identification dictionary. A discriminant function generator to be stored in the medium;
A large-scale transform that converts the m-dimensional feature into an L-dimensional feature on the L-dimensional feature space that is a dimension less than the m-dimension, a subspace of the n-dimensional feature space, and a subspace of the m-dimensional feature space. Generating a classification feature selection function, converting the detailed identification function as a function on the L-dimensional feature space, and calculating a large classification identification function for calculating the similarity of the pattern to each category on the L-dimensional feature space Main part of the discriminant function for generating and storing the generated major classification feature selection function in the storage medium as the feature selection dictionary and storing the generated major classification discrimination function in the storage medium as the discrimination dictionary An extractor;
A major classification feature selection unit that converts the extracted n-dimensional feature into the L-dimensional feature using the major classification feature selection function;
Using the converted L-dimensional features and the large classification identification function, a similarity of the pattern to be recognized for each category is calculated, and a large classification identification unit that selects one or more correct candidate patterns;
A feature selection unit that converts the n-dimensional feature into the m-dimensional feature using the feature selection function;
A detailed identification unit that calculates the similarity of the pattern to be recognized with respect to the selected correct candidate pattern using the detailed identification function and the converted m-dimensional feature;
A recognition result output unit that outputs a recognition result for the recognition target pattern based on the calculated similarity;
A pattern recognition apparatus comprising: The discriminant function main part extraction unit includes:
Performing a detailed identification function acquisition step of acquiring the detailed function;
Performing a feature axis selection step of selecting M feature axes from the n-dimensional feature space using the acquired detailed identification function;
Performing an axis importance determining step for calculating the importance of the feature axis;
Performing a main axis calculating step of calculating the L main axes by integrating the feature axes;
The large classification discrimination function generating step of generating the large classification identification function by converting the detailed identification function as a function on the L-dimensional feature space generated by the main axis is performed. The pattern recognition dictionary generation device described. The feature selection function is a linear function, and the detailed discriminant function uk (x) for the n-dimensional feature x is a second-order or lower polynomial function shown in Equation 9,
The feature axis selection step includes:
The discriminant function main part extraction unit uses the n-dimensional feature or the m-dimensional feature y to characterize the vector qki obtained by transforming the detailed discriminant function uk (x) as shown in Equation 10 or Equation 11. Including selecting as an axis,
The axis importance determination step is
15. The pattern according to claim 14, wherein the discriminant function main part extracting unit includes a step of setting the importance of the feature axis qki using the function hki of the coefficient λkii and the coefficient ζki in Expression 10 or 11. Recognition device.

The main axis calculation step includes:
The discriminant function main part extraction unit calculating an eigenvalue of the matrix Q shown in Formula 12 generated from the feature axis qki and the axis importance hki;
The discriminant function main part extraction unit selects L eigenvectors in descending order of the calculated eigenvalues among eigenvectors of the matrix Q, and calculates the selected eigenvectors as main axes; The pattern recognition apparatus according to claim 15, comprising:

The pattern recognition dictionary generation device includes an input unit for a user to operate the pattern recognition dictionary generation device,
In the feature axis selection step, the discriminant function main part extraction unit randomly selects M arbitrary vectors as the feature axes, or M vectors as the feature axes based on an instruction from the user. The pattern recognition apparatus according to claim 14, further comprising a selecting step.   In the axis importance determination step, the discriminant function main part extraction unit has transformed the detailed discriminant function for the N learning patterns xi as a function on the M-dimensional space generated by the feature axis qki. 3. The pattern recognition dictionary generation device according to claim 2, wherein a function d (xi) is calculated, and the axis importance h is calculated as a function of a variance value σ of the calculated function d (xi). . A pattern recognition dictionary generation method in a pattern recognition dictionary generation device comprising a processor and a storage medium connected to the processor,
The storage medium stores a learning pattern database composed of a plurality of learning patterns,
The method
The pattern recognition dictionary generation device acquires each learning pattern as one category from the learning pattern database;
A second step in which the pattern recognition dictionary generating device extracts an n-dimensional feature for each of the acquired categories;
The pattern recognition dictionary generation device generates a feature selection function that converts the n-dimensional feature into an m-dimensional feature that is a dimension equal to or less than the n-dimension using the extracted n-dimensional feature, and the generated A third step of storing a feature selection function in the storage medium as a feature selection dictionary;
A fourth step in which the dictionary recognition device for pattern recognition converts the extracted n-dimensional feature into the m-dimensional feature using the feature selection function;
The dictionary recognition device for pattern recognition generates a detailed identification function on an m-dimensional feature space for calculating a similarity of a pattern to be recognized for each category using the converted m-dimensional feature, and the generation A fifth step of storing the detailed identification function performed as an identification dictionary in the storage medium;
The pattern recognition dictionary generating device converts the dimension less than or equal to the m dimension into an L dimension feature on the L dimension feature space that is a subspace of the n dimension feature space and a subspace of the m dimension feature space. By generating a large-scale feature selection function for converting the m-dimensional feature and converting the detailed discriminant function as a function on the L-dimensional feature space, the similarity of the pattern to each category on the L-dimensional feature space is obtained. Generating a large classification identification function for calculation, storing the generated large classification feature selection function in the storage medium as the feature selection dictionary, and using the generated large classification identification function as the identification dictionary A sixth step of storing in a storage medium;
A pattern recognition dictionary generating method characterized by comprising: The sixth step includes
The pattern recognition dictionary generating device acquires a detailed identification function for acquiring the detailed function;
A feature axis selection step in which the dictionary recognition device for pattern recognition selects M feature axes from the n-dimensional feature space using the acquired detailed identification function;
The pattern recognition dictionary generating device calculates an importance of the feature axis, an axis importance determining step;
A principal axis calculating step in which the pattern recognition dictionary generating device calculates the L principal axes by integrating the feature axes;
The pattern recognition dictionary generation device generates a large classification function by generating the large classification function by converting the detailed classification function as a function on the L-dimensional feature space generated by the main axis;
The pattern recognition dictionary generation method according to claim 19, further comprising: The feature selection function is a linear function, and the detailed discriminant function uk (x) for the n-dimensional feature x is a second-order or lower polynomial function shown in Equation 13,
The feature axis selection step includes:
The pattern recognition dictionary generation device uses the n-dimensional feature or the m-dimensional feature y to transform the detailed identification function uk (x) as shown in Equation 14 or Equation 15 into a vector qki Including the step of selecting as
The axis importance determination step is
The pattern recognition dictionary generating apparatus includes the step of setting the importance of the feature axis qki using the function hki of the coefficient λkii and the coefficient ζki in the equation (14) or (15). Dictionary generation method.

The main axis calculation step includes:
The discriminant function main part extraction unit calculating an eigenvalue of the matrix Q shown in Formula 16 generated from the feature axis qki and the axis importance hki;
The discriminant function main part extraction unit selects L eigenvectors in descending order of the calculated eigenvalues among eigenvectors of the matrix Q, and calculates the selected eigenvectors as main axes. The pattern recognition dictionary generation method according to claim 21, further comprising:

The pattern recognition dictionary generation device includes an input unit for a user to operate the pattern recognition dictionary generation device,
In the feature axis selection step, the pattern recognition dictionary generating device randomly selects M arbitrary vectors as the feature axes, or selects M vectors as the feature axes based on an instruction from the user. 21. The pattern recognition dictionary generation method according to claim 20, further comprising the step of:   In the axis importance determination step, the discriminant function main part extraction unit has transformed the detailed discriminant function for the N learning patterns xi as a function on the M-dimensional space generated by the feature axis qki. 21. The method of claim 20, further comprising: calculating a function d (xi); and calculating the axis importance h as a function of a variance value σ of the calculated function d (xi). Pattern recognition dictionary generation method.

Images