Information processing method, information processing device, and computer program

Abstract

To provide an information processing method, an information processing apparatus, and a storage medium, by which "a hard-to-classify positive sample" is taken into consideration and performance of a classification model is enhanced as a result of introducing a hard-to-classify positive sample to an existing loss function while the classification model is trained.SOLUTION: An information processing method includes: extracting a feature vector of a sample concerning a training set including labels; recognizing hard-to-classify positive samples based on an angle between a feature vector of a standardized sample and a classification border of a class to which the sample belongs; weighting the hard-to-classify positive samples and accurate classification samples other than error classification samples and the hard-to-classify positive samples; and training the classification model based on the weighted hard-to-classify positive samples and the accurate classification samples other than the error classification samples and the hard-to-classify positive samples.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This invention relates to the field of information processing, and more particularly to an information processing method, an information processing device, and a computer program for face (human face) recognition. [Background technology]

[0002] Currently, convolutional neural network (CNN) classification models are widely used in the field of facial recognition. In the training process of classification models, the Softmax function is applied as the loss function for training CNN classification models.

[0003] When classifying samples using a CNN classification model, samples that are incorrectly classified into other categories (classes) are often called "misclassified samples," samples that are correctly classified are called "accurately classified samples," and among the accurately classified samples, those whose feature vectors are close to the classification boundary are called "difficult-to-classify positive samples." In conventional techniques, during the training process of a classification model, more attention is usually paid to misclassified samples, and the performance of the classification model is improved by introducing misclassified samples into the loss function.

[0004] However, conventional techniques do not consider "difficult-to-classify positive samples" during the training process. In reality, difficult-to-classify positive samples usually have relatively large intraclass differences compared to accurately classified samples, and are prone to being misclassified. Therefore, it is necessary to further improve the performance of classification models by considering "difficult-to-classify positive samples" during training and introducing difficult-to-classify positive samples into the existing loss function. [Overview of the project] [Problems that the invention aims to solve]

[0005] The object of the present invention is to provide an information processing method, an information processing device, and a computer program in order to solve the problems described above that exist in the prior art. Compared with classification models of the prior art, the classification model trained by the information processing method disclosed in the present invention has higher recognition accuracy. [Means for solving the problem]

[0006] According to one aspect of the present invention, an information processing method is provided, which is, Extract the feature vectors of the samples from the training set that includes the labels; Recognize hard-to-classify positive samples based on the angle between the normalized sample's feature vector and the classification boundary of the class to which the sample belongs; Weighting is applied to difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than difficult-to-classify positive samples; and This includes training a classification model based on weighted, difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the difficult-to-classify positive samples.

[0007] According to one aspect of the present invention, a sample is recognized as a difficult-to-classify positive sample when the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs is smaller than an angular threshold, and the sample is correctly classified.

[0008] According to one aspect of the present invention, the classification boundary of a class is determined based on the included angle between the center of the class to which the sample belongs and the center of the class that is closest to the class to which the sample belongs (i.e., the class with the smallest distance between the two classes).

[0009] According to one aspect of the present invention, when the cosine distance between the normalized feature vector of the sample and the center of the class to which the sample belongs is smaller than a first distance threshold and larger than a second distance threshold, the angle between the normalized feature vector of the sample and the classification boundary is determined to be smaller than an angle threshold. Based on the included angle between the center of the class to which the sample belongs and the center of the class closest to the class to which the sample belongs, and the angle threshold, the first distance threshold is determined, and based on the included angle, the second distance threshold is determined.

[0010] According to one aspect of the present invention, the included angle is maximized by determining the included angle based on an included angle optimization target function.

[0011] According to one aspect of the present invention, the included-angle optimization target function is a function that calculates the average of the sum of the included angles between the centers of all adjacent classes in the classification model.

[0012] According to one aspect of the present invention, weighting difficult-to-classify positive samples includes adaptively weighting them based on an adaptive weighting factor that changes in response to the cosine distance between the feature vector of the difficult-to-classify positive sample and the center of the class to which the difficult-to-classify positive sample belongs.

[0013] According to one aspect of the present invention, the loss function of the classification model is a softmax loss function, and the classification model is trained by optimizing the softmax loss function based on weighted hard-to-classify positive samples, misclassified samples, and accurately classified samples other than hard-to-classify positive samples.

[0014] According to one aspect of the present invention, an information processing apparatus is provided, which is, A feature extraction unit configured to extract feature vectors from samples for a training set that includes labels; A recognition unit configured to recognize difficult-to-classify positive samples based on the angle between the feature vector of a normalized sample and the classification boundary of the class to which the sample belongs; A weighting unit configured to perform weighting on difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than difficult-to-classify positive samples; and A training unit configured to train a classification model based on the weighted difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than difficult-to-classify positive samples.

[0015] According to one aspect of the present invention, there is provided a computer-readable storage medium storing a computer program, which can realize an information processing method when executed by a computer, and the information processing method includes: Extracting the feature vector of a sample for a training set including labels; Recognizing difficult-to-classify positive samples based on the angle between the feature vector of a normalized sample and the classification boundary of the class to which the sample belongs; Performing weighting on difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than difficult-to-classify positive samples; and Training a classification model based on the weighted difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than difficult-to-classify positive samples.

[0016] According to another aspect of the present invention, there are further provided corresponding computer program codes and computer program products.

Brief Description of the Drawings

[0017] [Figure 1] It is a flowchart of the information processing method in an embodiment of the present invention. [Figure 2] It is a diagram showing a part of the unit circle in which the centers of each class after normalization and the feature vectors of samples are distributed. [Figure 3] This figure shows how adaptive weighting is applied to difficult-to-classify positive samples in an embodiment of the present invention to bring them closer to the center of the class to which the sample belongs. [Figure 4] This is a configuration diagram of an information processing device in an embodiment of the present invention. [Figure 5] This block diagram shows an exemplary configuration of a general-purpose personal computer that can implement the method and / or apparatus according to the embodiments of the present invention. [Modes for carrying out the invention]

[0018] Hereinafter, preferred embodiments for carrying out the present invention will be described in detail with reference to the attached drawings. Note that these embodiments are merely illustrative and do not limit the present invention.

[0019] Before describing embodiments of the present invention, we will first explain the meanings of some terms used in this specification.

[0020] The term "training set with labels" used here refers to a training set where the class of the sample data is known. For example, if the sample data in the training set is image data of faces (human faces), then the inclusion of labels in the training set means that the identity of the person is known.

[0021] Those skilled in the art should understand that the sample data described herein is not limited to image data of human faces, but may also include other image data or other types of data, such as audio data.

[0022] The term "classification model" as used herein may refer to a machine learning model that can determine the class of unlabeled data by training on a labeled training set. A neural network model is one example of such a model, but the present invention is not limited to this.

[0023] The term "center of the class" used here refers to the class's weight vector.

[0024] It should be understood that here, a general explanation of technical solutions related to convolutional neural networks is omitted, and the focus is on the technical solutions specific to the present invention.

[0025] Figure 1 is a flowchart of the information processing method in an embodiment of the present invention.

[0026] Information processing method 100 starts with the initialization of the classification model.

[0027] In step S110, feature vectors of samples are extracted from the training set of face images that include labels.

[0028] Subsequently, in step S120, difficult-to-classify positive samples, misclassified samples, and other accurately classified samples other than the difficult-to-classify positive samples are identified.

[0029] In step S130, the softmax loss function of the classification model is optimized by weighting the difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the difficult-to-classify positive samples.

[0030] In step S140, the classification model is trained based on the softmax loss function of the optimized classification model.

[0031] Next, the information processing method determines whether the training objective has been achieved. If the training objective has not been achieved, steps S110-S140 are repeated. If the training objective has been achieved, the information processing method terminates.

[0032] The following describes each step shown in Figure 1 in detail.

[0033] First, the classification model is initialized, for example, by inputting a labeled training set and initializing the relevant parameters of the classification model (e.g., angular threshold, number of classes, class weights, learning rate, etc.).

[0034] In step S110, the feature vectors of the samples are extracted from the training set of face images containing labels, and a fully connected process is performed on the extracted sample feature vectors. Then, the feature vectors of the center of each class and the samples in the classification model are normalized so that the feature vectors of the center of each class and the samples are distributed on a unit circle.

[0035] The following explains, with reference to Figure 2, how difficult-to-classify positive samples, misclassified samples, and other accurately classified samples other than difficult-to-classify positive samples are recognized in step S120 shown in Figure 1.

[0036] Figure 2 shows a portion of the unit circle where the centers of each class and the feature vectors of the samples are distributed after normalization. The labels of samples S1, S2, and S3 are all class C1, meaning that samples S1, S2, and S3 all belong to class C1, and w1 is the center of class C1 (i.e., the weight vector of class C1). In Figure 2, the dotted line shows the classification boundary B1 between class C1, to which sample S1 belongs, and class C2, which is closest to class C1, and w2 is the center of class C2 (i.e., the weight vector of class C2).

[0037] As shown in Figure 2, samples S1 and S2 are correctly classified to belong to class C1, while sample S3 is incorrectly classified to belong to class C2. Typically, S1 and S2 are referred to as correctly classified samples (i.e., positive samples), and S3 is referred to as a misclassified sample. Correctly classified and misclassified samples can be identified by methods commonly used by those skilled in the art. For convenience, a detailed explanation of this is omitted here.

[0038] According to an embodiment of the present invention, a sample is a difficult-to-classify positive sample when the angle between the feature vector of a normalized accurately classified sample and the classification boundary of the class to which the sample belongs is smaller than an angular threshold.

[0039] As shown in FIG. 2, for example, for class C1, if the angle between the feature vector of the accurately classified sample after normalization and the classification boundary B1 of class C1 is smaller than the angle threshold A thre then the sample is a difficult-to-classify positive sample. That is, when the feature vector of the sample is within the angular interval (range) based on the angle threshold A thre (i.e., the feature vector of the sample is in the shaded part shown in FIG. 2), the sample is a difficult-to-classify positive sample. Therefore, the positive sample S2 in FIG. 3 is a difficult-to-classify positive sample.

[0040] According to an embodiment of the present invention, for the sake of calculation convenience, a comparison is made between the cosine distance between the feature vector of a sample and the center of the class to which the sample belongs, and the first distance threshold, and a comparison is made between the cosine distance and the second distance threshold, and based on the comparison result, it can be determined whether the angle between the feature vector of the sample and the classification boundary is smaller than the angle threshold.

[0041] For convenience, hereinafter, the cosine distance between the feature vector of a sample and the center of the class to which the sample belongs is referred to as the logit of the sample. The logit of sample S i (i.e., L si ) can be expressed by the following formula (1).

Equation

[0042] Among them, 1 ≤ j ≤ N, where N is the number of classes in the classification model, 1 ≤ i ≤ M, where M is the number of samples in the classification model. θ wj,Si represents the included angle between the feature vector of sample S i and the center w i of the class C j to which sample S j belongs.

[0043] The first distance threshold L t1 and the second distance threshold L t2 [[ID=

number

[0044] Eventually, θ n is the included angle between the center w1 of class C1 and the center w2 of the adjacent class C2, and A thre θ is the angular threshold. When the centers of each class are uniformly distributed, n This will be the maximum value.

[0045] In this invention, the θ when the centers of each class are uniformly distributed is given by the following equation (4). n Calculate.

number

[0046] Eventually,

number

[0047] Here, N is the number of classes and d is the dimension of the features. w θ is the included-angle optimization target function, which is a function that calculates the average of the sum of the included angles between the centers of all adjacent classes in the classification model. n The following can be calculated, and the class weight matrix W is updated using the gradient ascent method based on equation (5) below.

number

[0048] Of these, μ is the corresponding learning rate, and w j is Class C j It is the center, W j_nearest is Class C j Class C has the smallest distance to it. j_nearest It is the center of.

[0049] Next, the logit of the sample is the second distance threshold L t2 Larger than the first distance threshold L, and the sample logit is greater than the first distance threshold L. t1 When the angle between the sample's feature vector and the classification boundary of the class to which the sample belongs is smaller than the angular threshold, the sample can be recognized as a difficult-to-classify positive sample.

[0050] In step S130, when a sample is recognized as a difficult-to-classify positive sample, the softmax loss function of the classification model is further optimized by applying adaptive weighting to the difficult-to-classify positive sample, and the optimized softmax loss function L is shown in equation (6) below.

number

[0051] In equation (6), α is an adaptive weighting factor, where α = t*(logit+1.0) if the sample is recognized as a “difficult-to-classify positive sample”, and α = 1.0 otherwise, where t is a weight that adjusts the weighting factor for difficult-to-classify positive samples. s is a scale parameter used to measure the radius of the circumference where the feature vectors of the center of each class and the sample are distributed, and e sβ is the weight value of the misclassified sample, N is the number of classes, and S is the sample size. i The class is C j And, w j is Class C j It is the center of.

[0052] When a sample is not recognized as a "difficult-to-classify positive sample" (i.e., the sample is recognized as a "misclassified sample" or "another accurately classified sample other than a difficult-to-classify positive sample"), α = 1.0, and in this case, the softmax loss function is equivalent to the loss function based on positive and misclassified samples in the prior art.

[0053] The softmax loss function is optimized by incorporating a loss function for difficult-to-classify positive samples based on the existing softmax loss function.

[0054] Figure 3 shows how adaptive weighting is applied to difficult-to-classify positive samples in an embodiment of the present invention to bring them closer to the center of the class to which the sample belongs. As shown in Figure 3, by applying adaptive weighting to the logit of a difficult-to-classify positive sample S2, the included angle between the sample's feature vector and the center w1 of the class to which the sample belongs is further reduced. In other words, the feature vector of the difficult-to-classify positive sample is brought closer to the center of the class to which the sample belongs, thereby improving the performance of the classification model.

[0055] In step S140, the classification model is trained based on the optimized softmax loss function, and the performance of the classification model is improved by updating the relevant parameters of the classification model.

[0056] Although the optimized softmax loss function mentioned here does not include a margin, the present invention's method of processing information using a classification model can be similarly applied even when a margin exists, and when a margin exists, the optimized softmax loss function L m The expression is as shown in equation (7), where m is the margin.

number

[0057] The information processing apparatus and information processing method of the present invention can recognize difficult-to-classify positive samples and optimize the softmax loss function by applying adaptive weighting to these difficult-to-classify positive samples, thereby training a classification model and improving the performance of the classification model.

[0058] The table below shows that the recognition accuracy (recognition precision) of classification models trained using the information processing technology of the present invention is clearly improved compared to classification models of the prior art (for most training sample sets), using different training sample sets, namely the facial image databases CFP, AGEDB, COX-Cam1, COX-Cam2, and COX-Cam3. [Table 1]

[0059] Figure 4 is a diagram showing the configuration of an information processing device 400 according to an embodiment of the present invention. As shown in Figure 4, the information processing device 400 may include a feature extraction unit 410, a recognition unit 420, a weighting unit 430, and a training unit 440.

[0060] The feature extraction unit 410 is configured as follows: it extracts feature vectors of samples from a training set of face images including labels; the recognition unit 420 is configured as follows: it recognizes hard-to-classify positive samples, misclassified samples, and other accurately classified samples other than hard-to-classify positive samples; the weighting unit 430 is configured as follows: it optimizes the softmax loss function of the classification model by weighting hard-to-classify positive samples, misclassified samples, and other accurately classified samples other than hard-to-classify positive samples; and the training unit 440 is configured as follows: it trains the classification model based on the optimized softmax loss function of the classification model.

[0061] The information processing device 400 can implement the processing corresponding to the information processing method 100 shown in Figure 1. Specifically, the feature extraction unit 410, recognition unit 420, weighting unit 430, and training unit 440 in Figure 4 can implement the processing of steps S110-S140 in Figure 1, respectively. A detailed explanation of the processing performed by each unit in the information processing device 400 has already been given in the explanation of the information processing method 100 in Figure 1, so a detailed explanation is omitted here.

[0062] Each component or unit in the above-described device may be composed of software, firmware, hardware, or a combination thereof. Since the specific means and methods used for configuration are well known to those skilled in the art, a detailed explanation is omitted here. When implemented by software or firmware, the programs constituting the software are installed from a storage medium or network to a computer having a dedicated hardware structure (for example, the general-purpose computer 500 shown in Figure 5), and the computer can realize various functions when various programs are installed.

[0063] Figure 5 is a diagram showing the configuration of a hardware configuration (general-purpose computer) 500 that can realize the method and apparatus in the embodiment of the present invention.

[0064] The general-purpose computer 500 may be, for example, a computer system. Note that the general-purpose computer 500 is merely an example and does not limit the scope or function of the method and apparatus according to the present invention. Furthermore, the general-purpose computer 500 does not depend on any module, assembly, or combination thereof in the method and apparatus described above.

[0065] In Figure 5, the central processing unit (CPU) 501 performs various processes based on the program stored in the ROM 502 or the program routed from the memory unit 508 to the RAM 503. The RAM 503 can also store data necessary for the CPU 501 to perform various processes, as needed. The CPU 501, ROM 502, and RAM 503 are connected to each other via the bus 504. The input / output interface 505 is also connected to the bus 504.

[0066] Furthermore, the input / output interface 505 is connected to the following components: an input unit 506 including a keyboard, an output unit 507 including a display such as a liquid crystal display (LCD) and speakers, a storage unit 508 including a hard disk, and a communication unit 509 including a network interface card, such as a LAN card or modem. The communication unit 509 performs communication processing via a network such as the Internet or a LAN. The drive 510 may be connected to the input / output interface 505 as needed. A removable medium 511, such as a semiconductor memory, can be set in the drive 510 as needed, allowing computer programs read from it to be installed in the storage unit 508.

[0067] Furthermore, the present invention also provides a program product including a machine-readable instruction code. When such an instruction code is read and executed by a machine, it can perform the methods of the embodiments of the present invention described above. Accordingly, various storage media that carry such program products, such as magnetic disks (including floppy disks®), optical disks (including CD-ROMs and DVDs), magneto-optical disks (including MDs®), and semiconductor memory devices, are also included in the present invention.

[0068] The storage media mentioned above may include, but are not limited to, magnetic disks, optical disks, magneto-optical disks, and semiconductor memory devices.

[0069] Furthermore, each operation (process) in the above-described method can also be implemented as a computer-executable program stored on various machine-readable storage media.

[0070] Furthermore, the above examples and other details are disclosed as follows:

[0071] (Note 1) Information processing method, Extract the feature vectors of the samples from the training set that includes the labels; Recognize hard-to-classify positive samples based on the angle between the normalized sample's feature vector and the classification boundary of the class to which the sample belongs; Weighting is applied to the aforementioned difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the aforementioned difficult-to-classify positive samples; and A method comprising training a classification model based on weighted, difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the difficult-to-classify positive samples.

[0072] (Note 2) The information processing method described in Appendix 1, A sample is recognized as a difficult-to-classify positive sample when the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs is smaller than the angle threshold, and the sample is correctly classified.

[0073] (Note 3) The information processing method described in Appendix 1, A method for determining a class classification boundary based on the included angle between the center of the class to which the sample belongs and the center of the class closest to the class to which the sample belongs.

[0074] (Note 4) The information processing method described in Appendix 2, A method for determining the angle between the normalized sample's feature vector and the classification boundary when the cosine distance between the normalized sample's feature vector and the center of the class to which the sample belongs is smaller than a first distance threshold and larger than a second distance threshold, and determining the first distance threshold based on the included angle between the center of the class to which the sample belongs and the center of the class closest to the sample's class and the angle threshold, and determining the second distance threshold based on the included angle.

[0075] (Note 5) The information processing method described in Appendix 4, A method for maximizing the included angle by determining the included angle based on an included angle optimization target function.

[0076] (Note 6) The information processing method described in Appendix 5, The aforementioned included angle optimization target function is a function that calculates the average of the sum of the included angles between the centers of all adjacent classes in the classification model, according to the method.

[0077] (Note 7) The information processing method described in Appendix 1, A method for weighting the difficult-to-classify positive samples, comprising performing adaptive weighting based on an adaptive weighting factor that changes according to the cosine distance between the feature vector of the difficult-to-classify positive sample and the center of the class to which the difficult-to-classify positive sample belongs.

[0078] (Note 8) The information processing method described in Appendix 1, A method for training a classification model, wherein the loss function of the classification model is a softmax loss function, and the softmax loss function is optimized based on the weighted hard-to-classify positive samples, the misclassified samples, and the accurately classified samples other than the hard-to-classify positive samples.

[0079] (Note 9) An information processing device, A feature extraction unit configured to extract feature vectors from samples for a training set that includes labels; A recognition unit configured to recognize hard-to-classify positive samples based on the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs; A weighting unit configured to assign weights to the aforementioned difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the aforementioned difficult-to-classify positive samples; and Apparatus including a training unit configured to train a classification model based on weighted, difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the difficult-to-classify positive samples.

[0080] (Note 10) The information processing device described in Appendix 9, The recognition unit is configured to recognize a sample as a difficult-to-classify positive sample when the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs is smaller than an angle threshold, and the sample is correctly classified.

[0081] (Note 11) The information processing device described in Appendix 9, The recognition unit is configured to determine the class classification boundary based on the included angle between the center of the class to which the sample belongs and the center of the class closest to the class to which the sample belongs.

[0082] (Note 12) The information processing device described in Appendix 10, The recognition unit is configured to determine that the angle between the normalized feature vector of the sample and the classification boundary is smaller than an angle threshold when the cosine distance between the normalized feature vector of the sample and the center of the class to which the sample belongs is smaller than a first distance threshold and larger than a second distance threshold, and the first distance threshold is determined based on the included angle between the center of the class to which the sample belongs and the center of the class closest to the class to which the sample belongs and the angle threshold, and the second distance threshold is determined based on the included angle.

[0083] (Note 13) The information processing device described in Appendix 12, The recognition unit is configured to maximize the included angle by determining the included angle based on an included angle optimization target function.

[0084] (Note 14) The information processing device described in Appendix 13, The aforementioned included-angle optimization target function is a function that calculates the average of the sum of the included angles between the centers of all adjacent classes in the classification model.

[0085] (Note 15) The information processing device described in Appendix 9, The weighting unit is configured to perform adaptive weighting based on an adaptive weighting factor that changes according to the cosine distance between the feature vector of the difficult-to-classify positive sample and the center of the class to which the difficult-to-classify positive sample belongs.

[0086] (Note 16) The information processing device described in Appendix 9, The apparatus wherein the loss function of the classification model is a softmax loss function, the weighting unit is configured to optimize the softmax loss function based on the hard-to-classify positive samples, the misclassified samples, and the accurately classified samples other than the hard-to-classify positive samples after adaptive weighting, and the training unit is configured to train the classification model using the optimized softmax loss function.

[0087] (Note 17) The information processing device described in Appendix 9, The aforementioned classification model is used for face recognition and is implemented using a convolutional neural network model.

[0088] (Note 18) A computer-readable storage medium, A computer program is stored there. The computer program can implement an information processing method when executed by a computer, and the information processing method is Extract the feature vectors of the samples from the training set that includes the labels; Recognize hard-to-classify positive samples based on the angle between the normalized sample's feature vector and the classification boundary of the class to which the sample belongs; Weighting is applied to the aforementioned difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the aforementioned difficult-to-classify positive samples; and A storage medium comprising training a classification model based on weighted difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the difficult-to-classify positive samples.

[0089] (Note 19) A computer-readable storage medium as described in Appendix 18, A storage medium that recognizes a sample as a difficult-to-classify positive sample when the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs is smaller than an angle threshold, and the sample is correctly classified.

[0090] (Note 20) A computer-readable storage medium as described in Appendix 18, A storage medium that determines the classification boundary of a class based on the enclosed angle between the center of the class to which the sample belongs and the center of the class closest to the class to which the sample belongs.

[0091] Although preferred embodiments of the present invention have been described above, the present invention is not limited to these embodiments, and any modification to the present invention that does not deviate from the spirit of the invention falls within the technical scope of the present invention.

Claims

1. A computer-based information processing method, Extract the feature vectors of the samples from the training set that includes the labels; Recognizes hard-to-classify positive samples based on the angle between the normalized sample's feature vector and the classification boundary of the class to which the sample belongs; Weighting is applied to the aforementioned difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the aforementioned difficult-to-classify positive samples; and The classification model is trained based on weighted, difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the difficult-to-classify positive samples. An information processing method for recognizing a difficult-to-classify positive sample based on the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs, wherein the sample is recognized as a difficult-to-classify positive sample when the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs is smaller than an angle threshold and the sample is correctly classified.

2. The information processing method according to claim 1, An information processing method for determining a class classification boundary based on the enclosed angle between the center of the class to which the sample belongs and the center of the class closest to the class to which the sample belongs.

3. The information processing method according to claim 1, When the cosine distance between the normalized sample's feature vector and the center of the class to which the sample belongs is less than the first distance threshold and greater than the second distance threshold, it is determined that the angle between the normalized sample's feature vector and the classification boundary is less than the angle threshold. An information processing method that determines the first distance threshold based on the angle between the center of the class to which the sample belongs and the center of the class closest to the class to which the sample belongs, and the angle threshold, and determines the second distance threshold based on the angle.

4. The information processing method according to claim 3, An information processing method that maximizes the included angle by determining the included angle based on an included angle optimization target function.

5. The information processing method according to claim 4, The aforementioned included angle optimization target function is a function that calculates the average of the sum of the included angles between the centers of all adjacent classes in a classification model, in this information processing method.

6. The information processing method according to claim 1, An information processing method that includes performing adaptive weighting on the difficult-to-classify positive samples based on an adaptive weighting factor that changes according to the cosine distance between the feature vector of the difficult-to-classify positive sample and the center of the class to which the difficult-to-classify positive sample belongs.

7. The information processing method according to claim 1, The loss function of the classification model is the softmax loss function, and the weighted positive samples that are difficult to classify, An information processing method for training a classification model by optimizing the softmax loss function based on accurately classified samples other than the misclassified samples and the difficult-to-classify positive samples.

8. An information processing device, A feature extraction unit configured to extract feature vectors from samples for a training set that includes labels; A recognition unit configured to recognize hard-to-classify positive samples based on the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs; A weighting unit configured to weight the aforementioned difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the aforementioned difficult-to-classify positive samples; and Includes a training unit configured to train a classification model based on weighted difficult-to-classify positive samples, misclassified samples, and accurately classified samples other than the difficult-to-classify positive samples, The recognition unit is an information processing device that recognizes a sample as a difficult-to-classify positive sample when the angle between the normalized feature vector of the sample and the classification boundary of the class to which the sample belongs is smaller than an angle threshold, and the sample is correctly classified.

9. A program for causing a computer to execute the information processing method described in any one of claims 1 to 7.

Images