WO2011086643A1 - パターン認識装置、パターン認識方法及びパターン認識用プログラム - Google Patents
パターン認識装置、パターン認識方法及びパターン認識用プログラム Download PDFInfo
- Publication number
- WO2011086643A1 WO2011086643A1 PCT/JP2010/007360 JP2010007360W WO2011086643A1 WO 2011086643 A1 WO2011086643 A1 WO 2011086643A1 JP 2010007360 W JP2010007360 W JP 2010007360W WO 2011086643 A1 WO2011086643 A1 WO 2011086643A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- loss
- class
- recognition dictionary
- sum
- calculated
- Prior art date
Links
- 238000003909 pattern recognition Methods 0.000 title claims abstract description 56
- 238000012567 pattern recognition method Methods 0.000 title claims description 6
- 238000004364 calculation method Methods 0.000 claims abstract description 115
- 239000013598 vector Substances 0.000 claims abstract description 113
- 238000000034 method Methods 0.000 claims description 84
- 238000012545 processing Methods 0.000 claims description 34
- 238000000605 extraction Methods 0.000 claims description 31
- 230000006870 function Effects 0.000 description 50
- 230000014509 gene expression Effects 0.000 description 31
- 238000011156 evaluation Methods 0.000 description 26
- 238000010586 diagram Methods 0.000 description 14
- 230000007423 decrease Effects 0.000 description 4
- 239000000284 extract Substances 0.000 description 4
- 238000002790 cross-validation Methods 0.000 description 2
- 238000007429 general method Methods 0.000 description 2
- 238000000926 separation method Methods 0.000 description 2
- 241000257465 Echinoidea Species 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
Definitions
- the present invention relates to a pattern recognition device that recognizes a pattern of input data, a pattern recognition method, a program for pattern recognition, a recognition dictionary creation device that creates a recognition dictionary used for pattern recognition, a recognition dictionary creation method, and The present invention relates to a recognition dictionary creation program.
- Patent Document 1 As a general method for classifying input data into two groups, the techniques described in Patent Document 1 and Non-Patent Document 1 are known.
- the soft margin classification system described in Patent Document 1 determines parameters including weight vectors and biases in all data vectors in a training set, and the minimum non-negative number of slack variables in each data vector based on a plurality of constraints. To decide. Further, the soft margin classification system described in Patent Document 1 determines the minimum value of the cost function so as to satisfy a plurality of constraints.
- Non-Patent Document 1 is to perform linear separation on a feature space by mapping a pattern to a finite or infinite dimensional feature space when input data cannot be linearly separated.
- FIG. 17 is an explanatory diagram showing a general pattern recognition apparatus.
- the pattern recognition apparatus shown in FIG. 17 includes a data input unit 201, a feature extraction unit 202, a recognition dictionary creation unit 203, an identification unit 206, and a result output unit 208.
- the recognition dictionary creation unit 203 includes a loss calculation unit 204 and a margin calculation unit 205.
- the feature extraction unit 202 converts the data input from the data input unit 201 into a d-dimensional feature vector, and the recognition dictionary creation unit 203 creates the recognition dictionary 207.
- the feature extraction unit 202 converts the data input from the data input unit 201 into a d-dimensional feature vector, the identification unit 206 identifies the data using the recognition dictionary 207, and then the result output unit 208 outputs the identification result.
- the recognition dictionary creation unit 203 creates the recognition dictionary 207 so that the evaluation function obtained by adding the loss calculated by the loss calculation unit 204 and the reciprocal of the margin calculated by the margin calculation unit 205 is minimized.
- FIG. 18 is an explanatory diagram showing processing for creating a recognition dictionary from linearly separable data.
- a black circle (hereinafter referred to as “ ⁇ ”) shown in FIG. 18 indicates data belonging to the negative class
- a white circle hereinafter referred to as “ ⁇ ” indicates data belonging to the positive class.
- a parallel margin boundary that is, a solid line 302 and a solid line 303 that maximizes a margin between negative data and positive data is obtained, and a broken line 301 that is equidistant therefrom is identified as an identification boundary. Create a recognition dictionary.
- FIG. 19 is an explanatory diagram showing a process of creating a recognition dictionary from data that cannot be linearly separated. Similar to the contents shown in FIG. 18, the solid line 402 and the solid line 403 are parallel margin boundaries, and the broken line 401 is an identification boundary equidistant from these. However, unlike the case shown in FIG. 18, the data 406 and 407 enclosed by the squares are included in the data set, and thus linear separation cannot be performed.
- the recognition dictionary creation unit 203 creates a recognition dictionary so as to have a margin boundary and an identification boundary that minimize this loss and increase the margin as much as possible.
- the recognition dictionary creation unit 203 obtains an identification boundary that minimizes the value L defined by Equation 1.
- the coefficient C shown in Equation 1 is a parameter that determines the balance between margin and loss, and its value is determined by trial and error by a cross-validation method or the like.
- FIG. 20 is an explanatory diagram showing a set of data that cannot be linearly separated.
- the data 504 shown in FIG. 20 is data that exists at a position away from the original distribution due to noise, or data that appears to be away from the distribution because the number of data is small.
- a general pattern recognition apparatus sets an identification boundary (broken line 501) at the same position from the solid line 502 and the solid line 503 that maximize the margin between data.
- FIG. 21 is an explanatory diagram showing a case where new unlearned data is added to the data set shown in FIG.
- the identification boundary is set at the position of the broken line 602 shown in FIG. 21, the number of errors of data indicated by “ ⁇ ” is 1, the number of errors of data indicated by “ ⁇ ” is 1, and the total number of errors is 2. .
- the number of data errors indicated by “ ⁇ ” is 3 (that is, the total number of errors is 3).
- the identification accuracy is lowered. As described above, it is desirable that high identification accuracy can be realized for new unlearned data even when the data used for learning includes noise or the number of data is small.
- the present invention provides a pattern recognition that can perform pattern recognition with high identification accuracy on new data that has not been learned even when the data used for learning includes noise or when the number of data is small. It is an object to provide an apparatus, a pattern recognition method, a pattern recognition program, a recognition dictionary creation device, a recognition dictionary creation method, and a recognition dictionary creation program for creating a recognition dictionary used for pattern recognition.
- the pattern recognition apparatus includes a loss calculation means for calculating the loss of a feature vector indicating the characteristics of data belonging to each class for each class, and the loss between the classes based on the loss calculated for each class.
- a loss difference calculating means for calculating a sum of differences, a recognition dictionary creating means for creating a recognition dictionary based on the sum of losses calculated for each class by the loss calculating means and the sum of the differences in losses between the classes;
- a pattern identification unit for identifying a data pattern using a recognition dictionary, and the recognition dictionary creation unit calculates the sum of losses for each class calculated by the loss calculation unit based on the input feature vector, and a loss
- the recognition dictionary is modified so that the sum of the difference of losses between classes calculated by the difference calculation means is minimized, and the pattern identification means recognizes the data pattern using the corrected recognition dictionary. Characterized in that it.
- a recognition dictionary creation device is a recognition dictionary creation device for creating a recognition dictionary used by a pattern recognition device for recognizing a pattern of data, and the loss of a feature vector indicating the characteristics of data belonging to each class for each class.
- the loss calculation means to calculate, the loss difference calculation means to calculate the sum of the difference of loss between each class based on the loss calculated for each class, and the total loss calculated by the loss calculation means for each class
- a recognition dictionary creation means for creating a recognition dictionary based on the sum of the differences in losses between classes, and the recognition dictionary creation means calculated by the loss calculation means based on the input feature vector
- the recognition dictionary is modified so that the sum of the sum of losses for each class and the sum of the differences of losses between classes calculated by the loss difference calculation means is minimized.
- the loss of feature vectors indicating the characteristics of data belonging to each class is calculated for each class, and based on the loss calculated for each class, the sum of the difference in loss between the classes is calculated.
- a recognition dictionary is created, and the loss calculated for each class based on the input feature vector The recognition dictionary is modified so that the sum of the sum of the differences and the sum of the difference in loss between classes is minimized, and the data pattern is identified using the modified recognition dictionary.
- a recognition dictionary creation method is a recognition dictionary creation method for creating a recognition dictionary used by a pattern recognition device for recognizing a pattern of data, and the loss of a feature vector indicating the characteristics of data belonging to each class is classified for each class. Based on the loss calculated for each class, calculate the sum of the difference in loss between each class, and calculate the sum of the loss calculated for each class and the sum of the difference in loss between each class. Based on the input feature vector, the recognition dictionary is created so that the sum of the loss calculated for each class and the sum of the difference of losses between classes is minimized. It is characterized by correcting.
- the pattern recognition program allows a computer to calculate a loss of a feature vector indicating the characteristics of data belonging to each class for each class, based on the loss calculated for each class.
- Loss difference calculation process that calculates the total difference of losses
- recognition dictionary creation process that creates a recognition dictionary based on the total loss calculated for each class in the loss calculation process and the total difference of losses between classes
- a pattern identification process for identifying a pattern of data using a recognition dictionary, and the recognition dictionary creation process, based on the input feature vector, the loss total for each class calculated by the loss calculation process
- the recognition dictionary is corrected so that the sum of the difference of losses between classes calculated by the loss difference calculation process is minimized, and the recognition dictionary is corrected by the pattern identification process. Characterized in that to identify patterns of data using.
- a recognition dictionary creation program is a recognition dictionary creation program applied to a computer that creates a recognition dictionary used by a pattern recognition device for recognizing a pattern of data, and shows the characteristics of data belonging to each class to the computer.
- Loss calculation processing that calculates the loss of feature vectors for each class
- loss difference calculation processing that calculates the sum of the differences between classes based on the loss calculated for each class
- class for loss calculation processing Based on the total loss calculated for each class and the total difference in loss between classes, a recognition dictionary creation process is executed to create a recognition dictionary.
- the sum of the total loss calculated by the loss calculation process and the total loss difference calculated by the loss difference calculation process is minimized. Characterized in that to fix the urchin recognition dictionary.
- pattern recognition can be performed with high identification accuracy on new data that has not been learned.
- FIG. 1 is a block diagram showing an embodiment of a pattern recognition apparatus according to the present invention.
- the pattern recognition apparatus according to the present invention includes a data input unit 101, a feature extraction unit 102, a recognition dictionary creation unit 103, an identification unit 106, and a result output unit 108.
- the recognition dictionary creation unit 103 includes a continuous loss calculation unit 104, a loss difference calculation unit 105, and a recognition dictionary determination unit 109.
- the data input unit 101 notifies the feature extraction unit 102 of the recognition target data input to the pattern recognition device.
- Examples of the recognition target data include image data captured by a camera.
- the feature extraction unit 102 extracts d feature values (hereinafter also referred to as d-dimensional feature vectors) based on the data notified from the data input unit 101.
- d-dimensional feature vectors d feature values
- the entire image is divided into 10 ⁇ 10 areas, and the average of the luminance values of the images in each area is obtained.
- the method by which the feature extraction unit 102 extracts multidimensional feature vectors is not limited to the above-described method. Since a method of extracting a multidimensional feature vector from input data is widely known, detailed description thereof is omitted.
- the recognition dictionary creation unit 103 inputs the d-dimensional feature vector extracted by the feature extraction unit 102 to the continuous loss calculation unit 104 and the loss difference calculation unit 105 in the stage of creating the recognition dictionary 107 called “learning”, and calculates the calculation.
- a recognition dictionary 107 is created based on the result.
- the continuous loss calculation unit 104 calculates the loss for each class based on the d-dimensional feature vector. Then, the continuous loss calculation unit 104 calculates the total loss calculated for each class. In the following description, the loss calculated by the continuous loss calculation unit 104 is referred to as a continuous loss in order to distinguish it from a loss calculated by a general method.
- the loss difference calculation unit 105 calculates the difference between the loss of one class and the loss of another class based on the d-dimensional feature vector. Then, the loss difference calculation unit 105 calculates the sum of all differences between classes.
- the recognition dictionary determination unit 109 determines a recognition dictionary to be created based on the sum of continuous losses calculated by the continuous loss calculation unit 104 and the sum of loss differences between classes calculated by the loss difference calculation unit 105. .
- the identification unit 106 performs data pattern identification processing using the d-dimensional feature vector extracted by the feature extraction unit 102 and the recognition dictionary 107 and notifies the result output unit 108 of the identification result.
- the identification unit 106 may identify a pattern of input data and recognize a class to which the data belongs.
- the identification method using a d-dimensional feature vector and a recognition dictionary is widely known, detailed description is omitted.
- the result output unit 108 outputs the identification result received from the identification unit 106.
- the continuous loss calculation unit 104 in the recognition dictionary creation unit 103 calculates the continuous loss for the kth class according to Equation 2 shown below.
- Equation 2 P k is a prior probability for the k th class, N k is the number of feature vectors belonging to the k th class, a vector x kn is an n th feature vector belonging to the k th class, and a vector ⁇ is a recognition dictionary. This is a discriminator parameter used as 107.
- the prior probability indicates the existence probability or the appearance frequency regarding data for which the correct answer is known.
- the proportion of class k data included in the data is the prior probability of class k.
- a value that is already statistically known may be set in advance.
- P k N 1 + ⁇ + N K.
- the prior probability P k in this case is the ratio of the learning data correctly assigned as belonging to the class k to the total learning data.
- the discriminator parameter is a parameter related to discrimination defined according to the discriminator to be used.
- a kernel function is used as the discriminant function
- a coefficient used when weighting each kernel function may be used as the parameter ⁇ illustrated in Expression 2.
- r (•) represents a risk level indicating the degree to which a feature vector given as an argument becomes an identification error.
- r (•) represents a risk level indicating the degree to which a feature vector given as an argument becomes an identification error.
- g k (•) an identification function for calculating the similarity between the k-th class and the feature vector x is a function shown below.
- the discriminant function g k is a function that increases as the degree of belonging to the k-th class increases.
- a class that is most likely to be erroneously recognized for the vector x kn is a j-th class, and an identification function g j (hereinafter referred to as g j ( ⁇ )) that calculates the similarity between the j-th class and the feature vector x may be described. Is a function shown below.
- the continuous loss calculation unit 104 calculates the degree of risk that the feature vector x becomes an identification error using Expressions 3 to 6 illustrated below.
- f (hereinafter also referred to as f (•)) is an arbitrary monotonically increasing function, and is defined as, for example, Expression 7 illustrated below.
- ⁇ and ⁇ are hyper parameters, and desired values are set.
- values may be set using a method similar to the method for setting values in a general classifier. In this way, the continuous loss calculation unit 104 calculates the continuous loss for the class k by calculating the sum of the risks indicating how easily the input vector x belonging to the class k is erroneous.
- the continuous loss calculation unit 104 calculates the risk using the discrimination function g k that calculates the similarity between the k-th class and the feature vector x.
- the discriminant function g k used when calculating the risk is not limited to a function for calculating the similarity between the kth class and the feature vector x.
- the continuous loss calculation unit 104 may calculate the risk level using an identification function that calculates the distance between the kth class and the feature vector x.
- the smaller the output value calculated by the discriminant function that is, the shorter the distance
- the continuous loss calculation unit 104 may calculate the degree of risk using an expression in which g k (•) and g j (•) illustrated in Expressions 3 to 6 are interchanged.
- the continuous loss calculation unit 104 calculates the total continuous loss calculated for each class. That is, assuming that the number of classes is K, the continuous loss calculation unit 104 calculates the sum of continuous losses using Equation 8 illustrated below.
- the loss difference calculation unit 105 calculates the sum of the difference in loss between classes. For example, the loss difference calculation unit 105 calculates the sum of the difference in loss between the j-th class and the k-th class using Expression 9 illustrated below.
- the recognition dictionary determination unit 109 minimizes the weighted linear sum between the sum of the continuous losses calculated by the continuous loss calculation unit 104 and the sum of the difference of losses between classes calculated by the loss difference calculation unit 105.
- the discriminator parameter ⁇ is determined. For example, the recognition dictionary determination unit 109 weights the continuous loss calculated by the continuous loss calculation unit 104 using Equation 8 and the sum of the difference in loss between classes calculated by the loss difference calculation unit 105 using Equation 9.
- the discriminator parameter ⁇ that minimizes the added linear sum is determined.
- the recognition dictionary determination unit 109 may determine the discriminator parameter ⁇ so that the value L ( ⁇ ) calculated by Expression 10 illustrated below is minimized.
- the recognition dictionary determination unit 109 may determine the discriminator parameter ⁇ that minimizes L ( ⁇ ) by cross-validation.
- the identification unit 106 performs identification processing on the input data using the classifier parameter ⁇ determined in this way. Therefore, the determination of the discriminator parameter ⁇ by the recognition dictionary determination unit 109 means creation of a recognition dictionary used for input data identification processing.
- the recognition dictionary determination unit 109 sets the discriminator parameter ⁇ so that the evaluation value L ( ⁇ ) is minimized based on the input new data by the above method. Correct it. From this, it can be said that the recognition dictionary determination unit 109 corrects the recognition dictionary so that the evaluation value L ( ⁇ ) is minimized based on the input data.
- the output unit 108 is realized by a CPU of a computer that operates according to a program (pattern recognition program).
- the program is stored in a storage unit (not shown) of the pattern recognition device, and the CPU reads the program, and in accordance with the program, the data input unit 101, the feature extraction unit 102, the recognition dictionary creation unit 103 (more specifically, , Continuous loss calculation unit 104, loss difference calculation unit 105 and recognition dictionary determination unit 109), identification unit 106, and result output unit 108.
- Each of the result output units 108 may be realized by dedicated hardware.
- the recognition dictionary creation unit 103 (more specifically, the continuous loss calculation unit 104, the loss difference calculation unit 105, and the recognition dictionary determination unit 109) may operate as one device (a recognition dictionary creation device).
- FIG. 2 is a flowchart illustrating an example of processing for creating a recognition dictionary.
- Expression 10 is used as the evaluation function.
- it may be described as learning to create a recognition dictionary.
- the recognition dictionary is initialized using a learning data set (that is, used when creating a recognition dictionary) (step S701). Specifically, the recognition dictionary creation unit 103 sets an initial value of the discriminator parameter ⁇ , and puts the pattern recognition device into a state where it can be used for learning. Further, the recognition dictionary creation unit 103 sets a sufficiently large value for the variable L to be compared in the process described later.
- the data input unit 101 reads the input data (step S702) and notifies the feature extraction unit 102.
- the feature extraction unit 102 performs feature extraction from the input data to convert it into a d-dimensional feature vector (step S703).
- the recognition dictionary determination unit 109 corrects the discriminator parameter ⁇ , which is a recognition dictionary, so that the value of the evaluation function L ( ⁇ ) defined by Expression 10 decreases (steps S704 and S705). Specifically, the recognition dictionary determination unit 109 minimizes the weighted linear sum of the sum of the continuous losses calculated by the continuous loss calculation unit 104 and the sum of the difference of losses between classes calculated by the loss difference calculation unit 105.
- the discriminator parameter ⁇ is corrected so that
- the recognition dictionary determination unit 109 compares the value of the evaluation function L ( ⁇ ) with the value of the variable L. If the difference between the value of the evaluation function L ( ⁇ ) and the value of the variable L is sufficiently small (that is, converges to a constant value) (Yes in step S706), the recognition dictionary determination unit 109 at this time Is determined as a discriminator parameter, and the process is terminated.
- step S706 when the difference between the value of the evaluation function L ( ⁇ ) and the value of the variable L cannot be said to be sufficiently small (that is, has not converged to a certain value) (No in step S706), the recognition dictionary determination unit 109 Then, the value of L ( ⁇ ) at this time is substituted into the variable L, and the processing after step S704 is repeated.
- FIG. 3 is a flowchart illustrating an example of recognition processing using a recognition dictionary.
- the identification unit 106 initializes a recognition dictionary (step S801). Specifically, the identification unit 106 makes a state where the recognition dictionary created by the recognition dictionary creation unit 103 can be used.
- the data input unit 101 reads the input data (step S802) and notifies the feature extraction unit 102.
- the feature extraction unit 102 performs feature extraction from the input data to convert it into a d-dimensional feature vector (step S803).
- the identification unit 106 performs identification processing of the converted feature vector using the recognition dictionary (step S804), and notifies the result output unit 108 of the identification result. Then, the result output unit 108 outputs the identification result received from the identification unit 106 (step S805).
- the identification unit 106 determines whether or not the input data has been read (step S806). If the input data has not been read (No in step S806), the processes in and after step S802 are repeated. On the other hand, when the input data has been read (Yes in step S806), the identification unit 106 ends the process.
- the continuous loss calculation unit 104 calculates the continuous loss of feature vectors indicating the characteristics of data belonging to each class for each class. Further, the loss difference calculation unit 105 calculates the sum of the difference in loss between the classes based on the loss calculated for each class. Then, the recognition dictionary determination unit 109 creates a recognition dictionary based on the total loss calculated for each class and the total difference in loss between classes. Further, the recognition dictionary determination unit 109 sets the recognition dictionary so that the sum of the total loss calculated for each class based on the input feature vector and the total difference in loss between classes is minimized. Correct it. Then, the identification unit 106 identifies the data pattern using the corrected recognition dictionary. Therefore, even when the data used for learning includes noise or when the number of data is small, pattern recognition can be performed with high identification accuracy on new data that has not been learned.
- the recognition dictionary is determined so as to reduce the continuous loss and reduce the difference in loss between classes. Therefore, even if the data used for learning includes noise or the number of data is small, it is not learned. High identification accuracy can be realized for the new data. This effect will be described in detail below using specific examples. In the following specific example, a case where the number of classes is two will be described for ease of explanation. When the number of classes is two, the above formula 10 can be defined as the following formula 11.
- FIG. 4 and 5 are explanatory diagrams showing an example of an operation for determining an identification boundary for the data in the state shown in FIG.
- black circles hereinafter referred to as “ ⁇ ” in the figure are data belonging to class 1
- white circles hereinafter referred to as “ ⁇ ”
- a broken line 901 illustrated in FIG. 4 indicates an identification boundary set by the classifier.
- a solid line 902 indicates a margin boundary set at a certain distance ⁇ near the class 1 from the broken line 901. Further, the sum of the continuous losses of the data “ ⁇ ” included in the shaded area in FIG. 4 becomes the class 1 continuous loss L 1 ( ⁇ ).
- a broken line 1001 shown in FIG. 5 indicates an identification boundary set at the same position as the broken line 901 in FIG. 4, and a solid line 1002 is set closer to class 2 from the broken line 1001 and is also set at a constant distance ⁇ . Indicates the margin boundary. Further, the sum of the continuous losses of the data “ ⁇ ” included in the shaded area in FIG. 5 becomes the class 2 continuous loss L 2 ( ⁇ ).
- the continuous loss L 1 ( ⁇ ) of class 1 becomes a smaller value as the identification boundary moves away from the class 1 data indicated by “ ⁇ ”. Further, the continuous loss L 2 ( ⁇ ) of class 2 becomes smaller as the identification boundary is further away from the data of class 2 indicated by “ ⁇ ”.
- Expression 2 Expression 3, Expression 7, and Expression 11 are used.
- the value of ⁇ in Expression 7 is set to a sufficiently large value.
- the class 1 continuous loss L 1 ( ⁇ ) is included in the shaded region in FIG. 4 and the class 2 continuous loss L 2 ( ⁇ ) is included in the shaded region in FIG. It becomes almost equal to the number of “ ⁇ ”.
- obtaining ⁇ that minimizes the first term and the second term on the right side of Equation 11 means obtaining an identification boundary that minimizes the sum of these (ie, continuous loss).
- the number of “ ⁇ ” included in the shaded area in FIG. 4 and the number of “ ⁇ ” included in the shaded area in FIG. Means to find equal identification boundaries. Therefore, when the value of ⁇ in Equation 11 is sufficiently large, the number of “ ⁇ ” included in the shaded area in FIG. 4 is equal to the number of “ ⁇ ” included in the shaded area in FIG. Thus, an identification boundary that minimizes the sum of these is obtained.
- FIGS. 6 and 7 are explanatory diagrams showing an example of an operation for determining an identification boundary for the data in the state shown in FIG.
- class 1 data indicated by “ ⁇ ” is added immediately adjacent to class 2 data indicated by “ ⁇ ”. Even in such a case, the sum of these is obtained under the condition that the number of “ ⁇ ” included in the shaded area in FIG. 6 is equal to the number of “ ⁇ ” included in the shaded area in FIG.
- the identification boundary that minimizes is determined. That is, the identification boundary is set at the position of the broken line 1101 shown in FIG. 6 or the broken line 1201 shown in FIG.
- FIGS. 8 and 9 are explanatory diagrams showing another example of the operation for determining the identification boundary for the data in the state shown in FIG.
- an identification boundary (broken line 1301 in FIG. 8 or broken line 1401 in FIG. 9) is set at a position strongly dependent on data existing in the vicinity of the identification boundary, and is determined by a general pattern recognition apparatus. A result similar to the identification boundary (for example, the broken line 601 shown in FIG. 21) can be obtained.
- the present invention will be described with reference to specific examples, but the scope of the present invention is not limited to the contents described below.
- the number of classes is 2, and a kernel discriminant function is used as the discriminant function for each class. That is, the class k discrimination function is defined as shown in Equation 12 below.
- K is a kernel function (hereinafter also referred to as K (•)), and a Gaussian type kernel function defined as in the following Expression 13 is used.
- ⁇ (where ⁇ > 0) is a parameter that defines the size of the Gaussian kernel, and a desired value is set in advance in ⁇ .
- Expressions 2 and 6 are used as expressions for calculating the continuous loss for the kth class. Further, Equation 11 is used as the evaluation function.
- input vectors belonging to class 1 are (x [1], x [2],..., X [N1]), and inputs belonging to class 2 A vector is expressed as (x [N1 + 1], x [N1 + 2],..., X [N1 + N2]).
- a desired value is set for the prior probability P k in Equation 2.
- ( ⁇ [1], ⁇ [2],..., ⁇ [N1]) is used as a recognition dictionary for class 1, and ( ⁇ [N1 + 1], ⁇ 2 [N1 + 2],. ..., ⁇ 2 [N1 + N2]) are prepared, and all values are set to 1. Further, a minute value ⁇ indicating the change amount of the parameter value is prepared, and for example, 0.01 is set as the value of ⁇ .
- FIG. 10 is a flowchart illustrating an example of processing for creating a recognition dictionary in the present embodiment.
- sufficiently large values are set in the variables L new and L old (step S1501), and further, 1 is set in the variable i (step S1502).
- the recognition dictionary creation unit 103 stores the value of the i-th parameter ⁇ [i] in the variable ⁇ ′, decreases ⁇ [i] by ⁇ , and then calculates the evaluation value L ( ⁇ ) of Expression 11. Saved in the variable L ′ (step S1503).
- the process proceeds to step S1506.
- step S1505 the recognition dictionary creation unit 103 saves the value of the variable L ′ in the variable L new and also saves the value of the variable i in the variable j (step S1505).
- the recognition dictionary creation unit 103 returns the value stored in the variable ⁇ ′ to ⁇ [i] (step S1506). If the value of the variable i is smaller than the total number of parameters N1 + N2 (Yes in step S1507), the recognition dictionary creation unit 103 increases the value of i by 1 (S1508), and then returns to S1503 and repeats the subsequent processing. On the other hand, if the value of the variable i is not smaller than the total number N1 + N2 of parameters in step S1507 (No in step S1507), the process proceeds to step S1509.
- step S1509 If the value of variable L new is not smaller than L old (No in step S1509), the process ends. On the other hand, if the value of the variable L new is smaller than L old (Yes in step S1509), the process proceeds to step S1510.
- step S1510 the recognition dictionary creation unit 103 stores the value of L new in L old and changes the value of the jth parameter ⁇ [j] to decrease by ⁇ (step S1510). Then, the process returns to step S1502, and the subsequent processing is repeated.
- the value of the evaluation value L ( ⁇ ) is obtained by calculating the values of the continuous losses L 1 ( ⁇ ) and L 2 ( ⁇ ) of class 1 and class 2, respectively, as illustrated in Equation 11.
- FIG. 11 is a flowchart illustrating an example of processing for calculating a continuous loss of class 1.
- the recognition dictionary creation unit 103 sets the value of the variable L1 to 0 and the value of the variable n to 1 (step S1601). Further, the recognition dictionary creation unit 103 sets the value of the variable i to 1 and the value of the variable G1 to 0 (step S1602). Next, the recognition dictionary creation unit 103 calculates the value of the kernel function exemplified in Expression 13 using the input vectors x [n] and x [i], and multiplies the calculation result by the i-th parameter ⁇ [i]. The added value is added to the variable G1 (step S1603).
- step S1604 If the value of the variable i is smaller than N1 (Yes in step S1604), the recognition dictionary creation unit 103 increases the value of the variable i by 1 (step S1605), returns to step S1603, and repeats the subsequent processing. . On the other hand, if the value of the variable i is not smaller than N1 (No in step S1604), the process proceeds to step S1606. In step S1606, the recognition dictionary creation unit 103 sets the value of i to N1 + 1 and the value of G2 to 0 (step S1606).
- the recognition dictionary creation unit 103 calculates the value of the kernel function exemplified in Expression 13 using the input vectors x [n] and x [i], and multiplies the calculation result by the i-th parameter ⁇ [i]. The obtained value is added to the variable G2 (step S1607). If the value of the variable i is smaller than N1 + N2 (Yes in step S1608), the recognition dictionary creation unit 103 increments the value of the variable i by 1 (step S1609), returns to step S1607, and repeats the subsequent processing. . On the other hand, if the value of the variable i is not smaller than N1 + N2 (No in step S1608), the process proceeds to step S1610.
- the value set in the variable G1 is the value of the class 1 discriminant function
- the value set in the variable G2 is the value of the class 2 discriminant function. Therefore, the recognition dictionary creation unit 103 obtains a continuous loss related to the input vector x [n] belonging to class 1 according to Equation 6 (step S1610). If the value of the variable n is smaller than N1 (Yes in step S1611), the recognition dictionary creation unit 103 increases the value of the variable n by 1 (step S1612), returns to step S1602, and repeats the subsequent processing. On the other hand, if the value of the variable n is not smaller than N1 (No in step S1611), the process proceeds to step S1613.
- Recognition dictionary creation unit 103 calculates a value obtained by multiplying the prior probability P 1 that is set to a desired value in a variable L1, further it divided by the calculation results in the input vector number N1 Class 1 values Is set to the variable L1 (S1613). Thereafter, the process ends.
- FIG. 12 is a flowchart illustrating an example of processing for calculating a continuous loss of class 2.
- the recognition dictionary creation unit 103 sets the value of the variable L2 to 0 and the value of the variable n to N1 + 1 (step S1701). Further, the recognition dictionary creation unit 103 sets the value of the variable i to 1 and the value of the variable G1 to 0 (step S1702). Next, the recognition dictionary creation unit 103 calculates the value of the kernel function exemplified in Expression 13 using the input vectors x [n] and x [i], and multiplies the calculation result by the i-th parameter ⁇ [i]. The obtained value is added to the variable G1 (step S1703).
- step S1704 If the value of the variable i is smaller than N1 (Yes in step S1704), the recognition dictionary creation unit 103 increments the value of the variable i by 1 (step S1705), returns to step S1703, and repeats the subsequent processing. . On the other hand, if the value of the variable i is not smaller than N1 (No in step S1704), the process proceeds to step S1706. In step S1706, the recognition dictionary creation unit 103 sets the value of i to N1 + 1 and the value of G2 to 0 (step S1706).
- the recognition dictionary creation unit 103 calculates the value of the kernel function exemplified in Expression 13 using the input vectors x [n] and x [i], and multiplies the calculation result by the i-th parameter ⁇ [i]. The obtained value is added to the variable G2 (step S1707). If the value of the variable i is smaller than N1 + N2 (Yes in step S1708), the recognition dictionary creation unit 103 increments the value of the variable i by 1 (step S1709), returns to step S1707, and repeats the subsequent processing. . On the other hand, if the value of the variable i is not smaller than N1 + N2 (No in step S1708), the process proceeds to step S1710.
- the recognition dictionary creation unit 103 obtains a continuous loss related to the input vector x [n] belonging to class 2 according to Equation 6 (step S1710). If the value of the variable n is smaller than N1 + N2 (Yes in step S1711), the recognition dictionary creation unit 103 increases the value of the variable n by 1 (step S1712), returns to step S1702, and repeats the subsequent processing. On the other hand, if the value of the variable n is not smaller than N1 + N2 (No in step S1711), the process proceeds to step S1713.
- Recognition dictionary creation unit 103 calculates a value obtained by multiplying the prior probability P 2, which is set to a desired value in a variable L2, further divided by the calculation results in the input vector number N2 of Class 2 value Is set to the variable L2 (S1713). Thereafter, the process ends.
- FIG. 13 is a flowchart illustrating an example of processing for calculating an evaluation value.
- the evaluation value L ( ⁇ ) is obtained according to Expression 11 will be described.
- the recognition dictionary creation unit 103 sets a value obtained by adding the variable L1 and the variable L2 calculated in the above process to the variable L (step S1801). In addition, the recognition dictionary creation unit 103 sets a value obtained by multiplying the square of the difference between the variable L1 and the variable L2 calculated in the above-described process by a desired value ⁇ to the variable L ′ (step S1802). The recognition dictionary creation unit 103 sets a value obtained by adding the value set to the variable L in step S1801 and the value set to the variable L ′ in step S1802 to the variable L (step S1803), and ends the process. . The value of L obtained in this way is used as the evaluation value L ( ⁇ ) in step S1503 in FIG.
- the processing at the stage of creating the recognition dictionary has been described above. Next, the processing at the stage of recognizing data using the recognition dictionary will be described.
- the input data is converted into a d-dimensional feature vector x by predetermined feature extraction.
- the feature vector converted in this way is referred to as an input vector.
- FIG. 14 is a flowchart showing an example of identification processing performed on one input data.
- the identification unit 106 sets the value of the variable i to 1 and the value of the variable G1 to 0 (step S1901).
- the identification unit 106 calculates the value of the kernel function exemplified in Expression 13 using the input vectors x and x [i], and uses the value obtained by multiplying the calculation result by the i-th parameter ⁇ [i] as the variable G1. (Step S1902).
- step S1903 when the value of the variable i is smaller than N1 (Yes in step S1903), the identification unit 106 increases the value of the variable i by 1 (step S1904), returns to step S1902, and repeats the subsequent processing. On the other hand, if the value of the variable i is not smaller than N1 (No in step S1903), the process proceeds to step S1905. In step S1905, the identification unit 106 sets the value of i to N1 + 1 and the value of G2 to 0 (step S1905).
- the identification unit 106 calculates the value of the kernel function exemplified in Expression 13 using the input vectors x and x [i], and uses the value obtained by multiplying the calculation result by the i-th parameter ⁇ [i] as the variable G2. (Step S1906). If the value of the variable i is smaller than N1 + N2 (Yes in step S1907), the identifying unit 106 increases the value of the variable i by 1 (step S1908), returns to step S1906, and repeats the subsequent processing. On the other hand, if the value of the variable i is not smaller than N1 + N2 (No in step S1907), the process proceeds to step S1909.
- the identification unit 106 outputs a value obtained by subtracting the value of the variable G2 from the value of the variable G1 (that is, the value of G1-G2), identifies the class to which the input vector belongs, and ends the process (step S1909). ).
- the output value is positive
- the input vector x is identified as belonging to class 1.
- the output value is not positive
- the input vector x is recognized as belonging to class 2.
- the case where the number of classes is 2 has been described. However, the number of classes is not limited to two and may be three or more.
- the case where the classifier uses the kernel function defined by Expression 13 has been described. However, the function used by the classifier is not limited to the kernel function.
- FIG. 15 is a block diagram showing an example of the minimum configuration of the pattern recognition apparatus according to the present invention.
- the pattern recognition apparatus according to the present invention calculates a loss (for example, continuous loss) of a feature vector indicating the characteristics of data belonging to each class for each class (for example, the calculation is performed using Equation 8).
- the continuous loss calculation unit 104), and the loss difference calculation means 82 (for example, calculating using the equation 9) that calculates the sum of the differences of the losses between the classes based on the loss calculated for each class.
- the loss difference calculation unit 105 For example, the loss difference calculation unit 105), the total loss calculated for each class by the loss calculation unit 81 (for example, the total continuous loss calculated using Equation 8), and the total difference between the classes. Based on (for example, the sum of the difference of loss between classes calculated using Equation 9), a recognition dictionary creating means 83 (for example, recognition Dictionary decision section And 09), the pattern identifying means 84 for identifying the data pattern using a recognition dictionary (e.g., a recognition unit 106) and.
- a recognition dictionary creating means 83 for example, recognition Dictionary decision section And 09
- the pattern identifying means 84 for identifying the data pattern using a recognition dictionary (e.g., a recognition unit 106) and.
- the recognition dictionary creating unit 83 calculates the sum of losses calculated by the loss calculating unit 81 and the sum of differences calculated by the loss difference calculating unit 82.
- the recognition dictionary is corrected so that the sum of the evaluation values (e.g., the evaluation value L ( ⁇ ) of the evaluation function defined by Equation 10) is minimized, and the pattern identification unit 84 uses the corrected recognition dictionary to change the data pattern.
- FIG. 16 is a block diagram showing an example of the minimum configuration of the recognition dictionary creation apparatus according to the present invention.
- a recognition dictionary creation device is a recognition dictionary creation device for creating a recognition dictionary used by a pattern recognition device for recognizing a pattern of data, and loss of feature vectors indicating features of data belonging to each class (for example, continuous Loss) is calculated for each class (for example, calculated using Equation 8), for example, based on the loss calculated for each class, based on the loss calculated for each class.
- a loss difference calculation unit 92 (for example, the loss difference calculation unit 105) that calculates the sum of the differences of losses (for example, using Equation 9) and a loss sum calculated by the loss calculation unit 91 for each class (for example, , Based on the sum of the continuous losses calculated using Equation 8) and the sum of the differences in losses between classes (eg, the sum of the differences in losses calculated using Equation 9), Recognition dictionary Creating (e.g., to determine the identifier parameter alpha) and a recognition dictionary creating means 93 (e.g., the recognition dictionary determination section 109).
- the recognition dictionary creating unit 93 calculates the sum of the losses for each class calculated by the loss calculating unit 91 and the sum of the differences of losses calculated by the loss difference calculating unit 92.
- the recognition dictionary is corrected so that the sum (for example, the evaluation value L ( ⁇ ) of the evaluation function defined by Expression 10) is minimized.
- Pattern recognition is performed using a recognition dictionary created with such a configuration, even if the data used for learning contains noise or the number of data is small, for unlearned new data, Pattern recognition can be performed with high identification accuracy.
- Loss calculation means (for example, the continuous loss calculation unit 104) that calculates a loss (for example, continuous loss) of a feature vector indicating the characteristics of data belonging to each class for each class (for example, calculates using Equation 8). ) And the difference calculated between the classes based on the loss calculated for each class (for example, using the equation 9), the loss difference calculating means (for example, the loss difference calculating unit 105) ), The sum of losses calculated for each class by the loss calculation means (for example, the sum of continuous losses calculated using Equation 8), and the sum of the differences of losses between classes (for example, using Equation 9)
- a recognition dictionary creating means (for example, a recognition dictionary determining unit 109) for generating a recognition dictionary (for example, determining a discriminator parameter ⁇ ), and a recognition dictionary based on the calculated sum of difference of losses between classes)
- Use the data pattern Pattern recognition means (for example, an identification unit 106) for identifying the loss, and the recognition dictionary creation means calculates the sum of the losses for each class and the loss difference
- the recognition dictionary is corrected so that the sum (for example, the evaluation value L ( ⁇ ) of the evaluation function defined by Equation 10) with the sum of the difference of losses between the classes calculated by the means is minimized, and the pattern identification means
- a pattern recognition apparatus for identifying a data pattern using a corrected recognition dictionary.
- the loss calculation means calculates the sum of the risk levels (for example, the risk levels calculated by Formula 3 to Formula 6 and Formula 7) indicating the degree to which the class to which the feature vector belongs is an identification error (for example, Formula 2)
- Pattern recognition device that calculates the loss for each class based on the sum of the risk levels.
- the loss calculation means calculates the loss of each class using a kernel function (for example, calculation using Expressions 12 and 13), and the pattern identification means creates a recognition dictionary created based on the kernel function.
- a pattern recognition apparatus that uses data to identify patterns of data.
- a feature vector extraction unit (for example, a feature extraction unit 102) that extracts a feature vector from data input as data to be recognized is provided, and the loss calculation unit loses the feature vector extracted by the feature vector extraction unit. Is recognized for each class, and the pattern recognition unit identifies the feature vector pattern extracted by the feature vector extraction unit using a recognition dictionary.
- a recognition dictionary creation device for creating a recognition dictionary used by a pattern recognition device for recognizing a data pattern, wherein a loss of feature vectors (for example, continuous loss) indicating features of data belonging to each class is determined for each class.
- a loss calculation means for example, using the equation 8) to calculate (for example, the continuous loss calculation unit 104) and the total difference of losses between classes based on the loss calculated for each class
- a loss difference calculating means for example, calculating using the equation 9) (for example, the loss difference calculating unit 105) and a loss sum calculated by the loss calculating means for each class (for example, using the equation 8)
- a recognition dictionary is created (for example, identification) based on the sum of the difference of losses between classes (for example, the sum of the difference of losses between classes calculated using Equation 9).
- a recognition dictionary creation means for example, a recognition dictionary decision unit 109
- the recognition dictionary creation means based on the input feature vector, the sum of losses calculated by the loss calculation means for each class, Recognition that corrects the recognition dictionary so that the sum (for example, the evaluation value L ( ⁇ ) of the evaluation function defined by Expression 10) with the sum of the differences of the losses calculated by the loss difference calculation means is minimized.
- Dictionary creation device for example, a recognition dictionary decision unit 109
- the loss calculation means calculates the sum of the risk levels (for example, the risk levels calculated by Formula 3 to Formula 6 and Formula 7) indicating the degree of discrimination error in the class to which the feature vector belongs (for example, Formula 2).
- Recognition dictionary creation device that calculates the loss for each class based on the sum of the risk levels.
- Feature vectors are extracted from data input as recognition target data, the loss of the extracted feature vectors is calculated for each class, and the extracted feature vector pattern is identified using a recognition dictionary. Pattern recognition method.
- the present invention is preferably applied to a pattern recognition device that recognizes a pattern of input data.
- the pattern recognition apparatus according to the present invention is suitably applied to image recognition and the like.
Abstract
Description
102 特徴抽出部
103 認識辞書作成部
104 連続損失計算部
105 損失差計算部
106 識別部
107 認識辞書
108 結果出力部
109 認識辞書決定部
Claims (10)
- 各クラスに属するデータの特徴を示す特徴ベクトルの損失をクラスごとに計算する損失計算手段と、
クラスごとに計算された損失をもとに、各クラス間の損失の差の総和を計算する損失差計算手段と、
前記損失計算手段がクラスごとに計算した損失の総和と、前記各クラス間の損失の差の総和とに基づいて、認識辞書を作成する認識辞書作成手段と、
前記認識辞書を用いてデータのパターンを識別するパターン識別手段とを備え、
前記認識辞書作成手段は、入力された特徴ベクトルをもとに、前記損失計算手段が計算したクラスごとの損失の総和と、前記損失差計算手段が計算した各クラス間の損失の差の総和との和が最小になるように認識辞書を修正し、
パターン識別手段は、修正された認識辞書を用いてデータのパターンを識別する
ことを特徴とするパターン認識装置。 - 損失計算手段は、特徴ベクトルが属するクラスが識別誤りである度合いを示す危険度の総和をもとにクラスごとの損失を計算する
請求項1記載のパターン認識装置。 - 損失計算手段は、各クラスの損失をカーネル関数を用いて計算し、
パターン識別手段は、前記カーネル関数に基づいて作成された認識辞書を用いてデータのパターンを識別する
請求項1または請求項2に記載のパターン認識装置。 - 認識対象になるデータとして入力されたデータから特徴ベクトルを抽出する特徴ベクトル抽出手段を備え、
損失計算手段は、前記特徴ベクトル抽出手段が抽出した特徴ベクトルの損失をクラスごとに計算し、
パターン識別手段は、前記特徴ベクトル抽出手段が抽出した特徴ベクトルのパターンを認識辞書を用いて識別する
請求項1から請求項3のうちのいずれか1項に記載のパターン認識装置。 - データのパターンを認識するパターン認識装置が用いる認識辞書を作成する認識辞書作成装置であって、
各クラスに属するデータの特徴を示す特徴ベクトルの損失をクラスごとに計算する損失計算手段と、
クラスごとに計算された損失をもとに、各クラス間の損失の差の総和を計算する損失差計算手段と、
前記損失計算手段がクラスごとに計算した損失の総和と、前記各クラス間の損失の差の総和とに基づいて、認識辞書を作成する認識辞書作成手段とを備え、
前記認識辞書作成手段は、入力された特徴ベクトルをもとに、前記損失計算手段が計算したクラスごとの損失の総和と、前記損失差計算手段が計算した各クラス間の損失の差の総和との和が最小になるように認識辞書を修正する
ことを特徴とする認識辞書作成装置。 - 損失計算手段は、特徴ベクトルが属するクラスが識別誤りである度合いを示す危険度の総和をもとにクラスごとの損失を計算する
請求項5記載の認識辞書作成装置。 - 各クラスに属するデータの特徴を示す特徴ベクトルの損失をクラスごとに計算し、
クラスごとに計算された損失をもとに、各クラス間の損失の差の総和を計算し、
クラスごとに計算された損失の総和と、前記クラス間の損失の差の総和とに基づいて、認識辞書を作成し、
入力された特徴ベクトルをもとに、クラスごとに計算された損失の総和と、前記クラス間の損失の差の総和との和が最小になるように認識辞書を修正し、
修正された認識辞書を用いてデータのパターンを識別する
ことを特徴とするパターン認識方法。 - データのパターンを認識するパターン認識装置が用いる認識辞書を作成する認識辞書作成方法であって、
各クラスに属するデータの特徴を示す特徴ベクトルの損失をクラスごとに計算し、
クラスごとに計算された損失をもとに、各クラス間の損失の差の総和を計算し、
クラスごとに計算された損失の総和と、前記各クラス間の損失の差の総和とに基づいて、認識辞書を作成し、
入力された特徴ベクトルをもとに、クラスごとに計算された損失の総和と、前記各クラス間の損失の差の総和との和が最小になるように認識辞書を修正する
ことを特徴とする認識辞書作成方法。 - コンピュータに、
各クラスに属するデータの特徴を示す特徴ベクトルの損失をクラスごとに計算する損失計算処理、
クラスごとに計算された損失をもとに、各クラス間の損失の差の総和を計算する損失差計算処理、
前記損失計算処理でクラスごとに計算した損失の総和と、前記各クラス間の損失の差の総和とに基づいて、認識辞書を作成する認識辞書作成処理、および、
前記認識辞書を用いてデータのパターンを識別するパターン識別処理を実行させ、
前記認識辞書作成処理で、入力された特徴ベクトルをもとに、前記損失計算処理で計算したクラスごとの損失の総和と、前記損失差計算処理で計算した各クラス間の損失の差の総和との和が最小になるように認識辞書を修正させ、
パターン識別処理で、修正された認識辞書を用いてデータのパターンを識別させる
ことを特徴とするパターン認識プログラム。 - データのパターンを認識するパターン認識装置が用いる認識辞書を作成するコンピュータに適用される認識辞書作成プログラムであって、
前記コンピュータに、
各クラスに属するデータの特徴を示す特徴ベクトルの損失をクラスごとに計算する損失計算処理、
クラスごとに計算された損失をもとに、各クラス間の損失の差の総和を計算する損失差計算処理、および、
前記損失計算処理でクラスごとに計算した損失の総和と、前記各クラス間の損失の差の総和とに基づいて、認識辞書を作成する認識辞書作成処理を実行させ、
前記認識辞書作成処理で、入力された特徴ベクトルをもとに、前記損失計算処理で計算したクラスごとの損失の総和と、前記損失差計算処理で計算した各クラス間の損失の差の総和との和が最小になるように認識辞書を修正させる
ことを特徴とする認識辞書作成プログラム。
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP10843002.6A EP2525306A4 (en) | 2010-01-14 | 2010-12-20 | Pattern recognition device, pattern recognition method and pattern recognition-use program |
JP2011549770A JP5621787B2 (ja) | 2010-01-14 | 2010-12-20 | パターン認識装置、パターン認識方法及びパターン認識用プログラム |
CN201080061441.2A CN102713945B (zh) | 2010-01-14 | 2010-12-20 | 模式识别器和模式识别方法 |
US13/521,227 US8750628B2 (en) | 2010-01-14 | 2010-12-20 | Pattern recognizer, pattern recognition method and program for pattern recognition |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2010-005447 | 2010-01-14 | ||
JP2010005447 | 2010-01-14 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2011086643A1 true WO2011086643A1 (ja) | 2011-07-21 |
Family
ID=44303949
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2010/007360 WO2011086643A1 (ja) | 2010-01-14 | 2010-12-20 | パターン認識装置、パターン認識方法及びパターン認識用プログラム |
Country Status (6)
Country | Link |
---|---|
US (1) | US8750628B2 (ja) |
EP (1) | EP2525306A4 (ja) |
JP (1) | JP5621787B2 (ja) |
KR (1) | KR20120104363A (ja) |
CN (1) | CN102713945B (ja) |
WO (1) | WO2011086643A1 (ja) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2021128474A (ja) * | 2020-02-12 | 2021-09-02 | 株式会社東芝 | 学習装置、学習方法、および学習プログラム |
Families Citing this family (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2014118978A1 (ja) | 2013-02-01 | 2014-08-07 | 富士通株式会社 | 学習方法、情報処理装置および学習プログラム |
WO2016181468A1 (ja) * | 2015-05-11 | 2016-11-17 | 株式会社東芝 | パターン認識装置、パターン認識方法およびプログラム |
CN107533672B (zh) * | 2015-05-11 | 2021-01-12 | 株式会社东芝 | 模式识别装置、模式识别方法以及记录介质 |
US9582715B2 (en) | 2015-06-30 | 2017-02-28 | International Business Machines Corporation | Feature selection algorithm under conditions of noisy data and limited recording |
CN107402914B (zh) * | 2016-05-20 | 2020-12-15 | 富士通株式会社 | 自然语言深度学习系统和方法 |
JP6199461B1 (ja) * | 2016-09-13 | 2017-09-20 | ヤフー株式会社 | 情報処理装置、情報処理方法、およびプログラム |
JP7114892B2 (ja) * | 2017-12-13 | 2022-08-09 | 京セラドキュメントソリューションズ株式会社 | 画像形成装置 |
US11829861B2 (en) * | 2019-07-17 | 2023-11-28 | Unist (Ulsan National Institute Of Science And Technology) | Methods and apparatus for extracting data in deep neural networks |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH0887572A (ja) | 1994-06-30 | 1996-04-02 | At & T Corp | ソフトマージン分類システム |
JPH08106295A (ja) * | 1994-10-05 | 1996-04-23 | Atr Onsei Honyaku Tsushin Kenkyusho:Kk | パターン認識方法及び装置 |
JP2010005447A (ja) | 2009-10-13 | 2010-01-14 | Sophia Co Ltd | 遊技機 |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7076473B2 (en) * | 2002-04-19 | 2006-07-11 | Mitsubishi Electric Research Labs, Inc. | Classification with boosted dyadic kernel discriminants |
US7490071B2 (en) * | 2003-08-29 | 2009-02-10 | Oracle Corporation | Support vector machines processing system |
JP2005198970A (ja) * | 2004-01-19 | 2005-07-28 | Konica Minolta Medical & Graphic Inc | 医用画像処理装置 |
-
2010
- 2010-12-20 US US13/521,227 patent/US8750628B2/en active Active
- 2010-12-20 CN CN201080061441.2A patent/CN102713945B/zh active Active
- 2010-12-20 JP JP2011549770A patent/JP5621787B2/ja active Active
- 2010-12-20 KR KR1020127018452A patent/KR20120104363A/ko not_active Application Discontinuation
- 2010-12-20 EP EP10843002.6A patent/EP2525306A4/en not_active Ceased
- 2010-12-20 WO PCT/JP2010/007360 patent/WO2011086643A1/ja active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH0887572A (ja) | 1994-06-30 | 1996-04-02 | At & T Corp | ソフトマージン分類システム |
JPH08106295A (ja) * | 1994-10-05 | 1996-04-23 | Atr Onsei Honyaku Tsushin Kenkyusho:Kk | パターン認識方法及び装置 |
JP2010005447A (ja) | 2009-10-13 | 2010-01-14 | Sophia Co Ltd | 遊技機 |
Non-Patent Citations (4)
Title |
---|
LEARNING, vol. 20, 1995, pages 273 - 297 |
NORIKAZU TAKAHASHI: "Support Vector Machine and its Efficient Learning Algorithms", IPSJ SIG NOTES, vol. 2007, no. 96, 25 September 2007 (2007-09-25), pages 49 - 54, XP008170794 * |
See also references of EP2525306A4 |
SHOTARO AKO: "Kernel Machine", IEICE TECHNICAL REPORT, vol. 103, no. 228, 22 July 2003 (2003-07-22), pages 1 - 6, XP008170514 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2021128474A (ja) * | 2020-02-12 | 2021-09-02 | 株式会社東芝 | 学習装置、学習方法、および学習プログラム |
JP7221892B2 (ja) | 2020-02-12 | 2023-02-14 | 株式会社東芝 | 学習装置、学習方法、および学習プログラム |
Also Published As
Publication number | Publication date |
---|---|
JP5621787B2 (ja) | 2014-11-12 |
JPWO2011086643A1 (ja) | 2013-05-16 |
CN102713945A (zh) | 2012-10-03 |
KR20120104363A (ko) | 2012-09-20 |
US20130129220A1 (en) | 2013-05-23 |
EP2525306A1 (en) | 2012-11-21 |
EP2525306A4 (en) | 2018-01-10 |
CN102713945B (zh) | 2015-03-25 |
US8750628B2 (en) | 2014-06-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP5621787B2 (ja) | パターン認識装置、パターン認識方法及びパターン認識用プログラム | |
US7242810B2 (en) | Multimodal high-dimensional data fusion for classification and identification | |
US7724961B2 (en) | Method for classifying data using an analytic manifold | |
Parrish et al. | Classifying with confidence from incomplete information | |
US11521099B2 (en) | Dictionary generation apparatus, evaluation apparatus, dictionary generation method, evaluation method, and storage medium for selecting data and generating a dictionary using the data | |
US8612370B2 (en) | Pattern recognition device, pattern recognition method, and pattern recognition program | |
US11049045B2 (en) | Classification apparatus, robot, and classification method | |
CN111126482A (zh) | 一种基于多分类器级联模型的遥感影像自动分类方法 | |
CN111444937A (zh) | 一种基于集成tsk模糊分类器的众包质量提升的方法 | |
Symeonidis et al. | Neural attention-driven non-maximum suppression for person detection | |
Antosik et al. | New measures of classifier competence-heuristics and application to the design of multiple classifier systems | |
Beigi et al. | A hierarchical approach to large-scale speaker recognition. | |
CN107729945A (zh) | 基于类间稀疏表示的鉴别回归、分类方法及系统 | |
Chen et al. | Online vehicle logo recognition using Cauchy prior logistic regression | |
Roos et al. | Probabilistic SVM for open set automatic target recognition on high range resolution radar data | |
WO2017000118A1 (en) | Method and apparatus for predicting attribute for image sample | |
Sikdar et al. | Open-set metric learning for person re-identification in the wild | |
Sun et al. | Discriminative canonical correlation analysis with missing samples | |
Besrour et al. | Incremental subclass support vector machine | |
Cecotti et al. | Rejection schemes in multi-class classification--application to handwritten character recognition | |
JP2020181265A (ja) | 情報処理装置、システム、情報処理方法及びプログラム | |
Huang et al. | Unconstrained face verification by optimally organizing multiple classifiers | |
Kundu et al. | A modified RBFN based on heuristic based clustering for location invariant fingerprint recognition and localization with and without occlusion | |
Sangineto et al. | Learning discriminative spatial relations for detector dictionaries: an application to pedestrian detection | |
CN113221941B (zh) | 一种自适应对称图信息传播的鉴别回归分类方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 201080061441.2 Country of ref document: CN |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 10843002 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2011549770 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2010843002 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 13521227 Country of ref document: US |
|
ENP | Entry into the national phase |
Ref document number: 20127018452 Country of ref document: KR Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |