Feature determination device, feature determination method, and computer-readable storage medium

Abstract

The similarity calculation unit (213) takes a plurality of types as objects, respectively, and calculates the similarity between a reference feature quantity and a recognition feature quantity, wherein the reference feature quantity is a feature quantity extracted from image data of the object type, and the recognition feature quantity is a feature quantity extracted from image data as the recognition object, i.e., recognition object data. The influence calculation unit (214) takes the similarity and recognition feature quantity related to each type as input, and calculates the influence degree on the similarity related to each partial image of the recognition object data. The feature determination unit (215) changes the recognition feature quantity according to the influence degree.

Description

Technical Field

[0001] The present invention relates to techniques for determining features of image data. Background Art

[0002] In financial institutions and government agencies, paper documents such as application forms are used to register, change, and delete data. Paper documents are processed according to the following procedures, for example: (1) document type, (2) defect confirmation, (3) content registration (if there is no defect), and (4) return to the recorder (if there is a defect). Performing these processes requires manpower costs, so it is desirable to automate these processes.

[0003] As a countermeasure to automate these processes, there is a system that digitizes paper documents, recognizes recorded items and recorded contents by computers, and determines defects (for example, see Patent Document 1). In such a system, it is necessary to define initial settings such as recording areas in paper documents before the system is operated. Performing this initial setting manually takes a lot of effort and time.

[0004] As a method for automatically extracting a recording area from the image data of a paper document, the following method is considered. A template is printed or recorded in advance in a paper document. Then, the template is determined based on the image data of the paper document, and the area outside the template is determined as the recording area. In order to realize this method, a technology for determining the template contained in the image data is required.

[0005] Here, in a plurality of recorded documents recorded by multiple persons, the recorded area is sometimes recorded and sometimes not recorded. In addition, the recorded content and the shape of the characters in the recorded area are different. Therefore, it is highly possible that the characteristic parts in the image data of the plurality of paper documents are templates. In addition, in the case where there are a plurality of types of paper documents, it is highly possible that the parts that help identify each type, that is, the characteristic parts inherent to each type in the image data of each type of paper document, are templates.

[0006] Patent Document 2 describes a technique for extracting features of image data using CNN (Convolutional Neural Network). Non-Patent Document 1 describes a technique for specifying a part in image data that is a basis for the type of CNN.

[0007] Prior art literature

[0008] Patent Literature

[0009] Patent Document 1: Japanese Patent Application Publication No. 2018-067180

[0010] Patent Document 2: Japanese Patent Application Publication No. 2016-110232

[0011] Non-patent literature

[0012] Non-patent literature 1: R.Selvaraju et al, "Grad-CAM: Visual Explanations from DeepNetworks via Gradient-based Localization" Summary of the invention

[0013] Problems to be solved by the invention

[0014] When determining the characteristic parts inherent to each type using the technology described in Patent Document 2 and Non-Patent Document 1, it is necessary to construct a CNN that learns the parameters related to each type. Therefore, every time the number of types increases, the parameters of the CNN need to be learned again, which requires computer resources and time.

[0015] An object of the present invention is to appropriately determine characteristic parts inherent to each type without relearning CNN parameters when the number of types increases.

[0016] Means for solving problems

[0017] The feature determination device of the present invention comprises: a similarity calculation unit, which takes a plurality of types as objects respectively, and calculates the similarity between a baseline feature quantity and a recognition feature quantity, wherein the baseline feature quantity is a feature quantity extracted from image data of the object type, and the recognition feature quantity is a feature quantity extracted from image data serving as a recognition object, i.e., recognition object data; an influence calculation unit, which takes the similarity related to each type calculated by the similarity calculation unit and the recognition feature quantity as input, and calculates the influence of each partial image of the recognition object data on the similarity; and a feature determination unit, which changes the recognition feature quantity according to the influence calculated by the influence calculation unit.

[0018] Effects of the Invention

[0019] In the present invention, the similarity between the reference feature and the identification feature is calculated for each type, and the influence of each partial image of the identification target data on the similarity is calculated. Then, the identification feature is changed according to the influence.

[0020] Thus, even if the reference feature and the identification feature are extracted using a CNN without information about the type, the characteristic part can be properly determined. Therefore, when the number of types increases, the characteristic part can be properly determined without relearning the parameters of the CNN used to extract the reference feature and the identification feature. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1This is a configuration diagram of the feature identification system 100 according to the first embodiment.

[0022] Figure 2 This is a hardware configuration diagram of the machine learning device 10 and the feature identification device 20 according to the first embodiment.

[0023] Figure 3 This is a functional configuration diagram of the machine learning device 10 according to the first embodiment.

[0024] Figure 4 This is a functional configuration diagram of the feature identification device 20 according to the first embodiment.

[0025] Figure 5 This is a flowchart showing the operation of the machine learning device 10 according to the first embodiment.

[0026] Figure 6 This is a flowchart showing the operation of the feature identification device 20 according to the first embodiment.

[0027] Figure 7 This is a flowchart of the similarity calculation process in the first embodiment.

[0028] Figure 8 This is a flowchart of the influence calculation process according to the first embodiment.

[0029] Fig. 9 This is a flowchart of the feature identification process of implementation mode 1.

[0030] Fig.10 This is a flowchart of the similarity calculation process according to Modification Example 2. DETAILED DESCRIPTION

[0031] Implementation Method 1

[0032] ***Description of the structure***

[0033] Reference Figure 1 The configuration of the feature identification system 100 according to the first embodiment will be described.

[0034] The feature identification system 100 includes a machine learning device 10 and a feature identification device 20. The machine learning device 10 and the feature identification device 20 are connected via a communication path 30 such as a LAN (Local Area Network) to enable data transmission and reception.

[0035] The machine learning device 10 and the feature identification device 20 may be implemented by a single device. In this case, the machine learning device 10 and the feature identification device 20 can exchange data via a signal line or the like of the device.

[0036] Reference Figure 2The hardware configuration of the machine learning device 10 and the feature identification device 20 according to the first embodiment will be described.

[0037] The machine learning device 10 and the feature determination device 20 are computers.

[0038] The machine learning device 10 has a processor 11 , a memory 12 , an auxiliary storage device 13 , and a communication interface 14 . The feature determination device 20 has a processor 21 , a memory 22 , an auxiliary storage device 23 , a communication interface 24 , and an output interface 25 .

[0039] The processors 11 and 21 are ICs (Integrated Circuits) that perform processing. As specific examples, the processors 11 and 21 are CPUs (Central Processing Units), DSPs (Digital Signal Processors), and GPUs (Graphics Processing Units).

[0040] The memories 12 and 22 are storage devices for temporarily storing data. As specific examples, the memories 12 and 22 are SRAM (Static Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0041] The auxiliary storage devices 13 and 23 are storage devices for storing data. As a specific example, the auxiliary storage devices 13 and 23 are HDDs (Hard Disk Drives). In addition, the auxiliary storage devices 13 and 23 may be removable recording media such as SD (registered trademark; Secure Digital) memory cards, CF (CompactFlash; registered trademark), NAND flash memory, floppy disks, optical disks, high-density disks, Blu-ray (registered trademark) disks, and DVDs (Digital Versatile Disks).

[0042] The communication interfaces 14 and 24 are interfaces for communicating with external devices. As specific examples, the communication interfaces 14 and 24 are ports of Ethernet (registered trademark) or USB (Universal Serial Bus).

[0043] The output interface 25 is an interface for communicating with a device such as a display device. As a specific example, the output interface 25 is a port of HDMI (registered trademark; High-Definition Multimedia Interface).

[0044] Reference Figure 3 The functional configuration of the machine learning device 10 according to the first embodiment will be described.

[0045] The machine learning device 10 includes, as functional components, a sampling unit 111, a binary image conversion unit 112, an image generation unit 113, a feature extraction unit 114, an image reconstruction unit 115, and a parameter update unit 116. The functions of the functional components of the machine learning device 10 are implemented by software.

[0046] A program for realizing the functions of each functional component of the machine learning device 10 is stored in the auxiliary storage device 13. The program is read into the memory 12 by the processor 11 and executed by the processor 11. Thus, the functions of each functional component of the machine learning device 10 are realized.

[0047] Furthermore, the auxiliary storage device 13 stores a plurality of learning image data 131 , a first parameter 132 , and a second parameter 133 .

[0048] Reference Figure 4 The functional structure of the feature identification device 20 according to the first embodiment will be described.

[0049] As functional components, the feature identification device 20 includes a feature extraction unit 211 and a feature modification unit 212. The feature modification unit 212 includes a similarity calculation unit 213, an influence calculation unit 214, and a feature identification unit 215. The functions of the functional components of the feature identification device 20 are implemented by software.

[0050] A program for realizing the functions of each functional component of the feature identification device 20 is stored in the auxiliary storage device 23. The program is read into the memory 22 by the processor 21 and executed by the processor 21. Thus, the functions of each functional component of the feature identification device 20 are realized.

[0051] Furthermore, the auxiliary storage device 23 stores one or more registered image data 231 associated with each of the plurality of categories, one or more reference feature quantities 232 associated with each of the plurality of categories, and one or more identification target data 233 .

[0052] exist Figure 2Only one processor 11 is shown. However, there may be multiple processors 11, and multiple processors 11 may cooperate to execute programs that implement the various functions of the machine learning device 10. Similarly, there may be multiple processors 21, and multiple processors 21 may cooperate to execute programs that implement the various functions of the feature determination device 20.

[0053] In the following description, data transmission and reception between functional components in each device may be performed through inter-process communication or through the memories 12 and 22 .

[0054] ***Description of the action***

[0055] Reference Figures 5 to 9 The operation of the feature identification system 100 according to the first embodiment will be described.

[0056] The operation of the feature identification system 100 in the first embodiment corresponds to the feature identification method in the first embodiment. In addition, the operation of the feature identification system 100 in the first embodiment corresponds to the processing of the feature identification program in the first embodiment.

[0057] Reference Figure 5 The operation of the machine learning device 10 according to the first embodiment will be described.

[0058] The machine learning device 10 takes the plurality of learning image data 131 stored in the auxiliary storage device 13 as objects and executes Figure 5 Thus, the first parameter 132 of the first neural network and the second parameter 133 of the second neural network are updated.

[0059] Each learning image data 131 is image data of any type among the multiple types to be learned. In Embodiment 1, each learning image data 131 is image data obtained by imaging a paper document of any type among the multiple types. In addition, each learning image data 131 does not include information related to the type.

[0060] (Step S11: Sampling Process)

[0061] The sampling unit 111 samples pixel data from the target learning image data 131 and converts the pixel data into image data of a reference size. Here, it is assumed that each learning image data 131 is grayscale.

[0062] (Step S12: Binary Image Conversion Process)

[0063] The binary image conversion unit 112 binarizes the grayscale learning image data 131 converted into the reference size in step S11 , and generates binary image data, that is, converted binary data.

[0064] (Step S13: Image generation processing)

[0065] The image generating unit 113 performs pre-processing necessary for learning on the grayscale learning image data 131 converted into the reference size in step S11 .

[0066] As a specific example, the image generation unit 113 removes noise when the image is acquired from the learning image data 131 to make the state of the learning image data 131 uniform. Alternatively, the image generation unit 113 adds noise to the learning image data 131 to make the learning image data 131 have changes. As a method of adding noise, it is possible to add Gaussian noise, salt and pepper noise, Gaussian blur, etc.; perform operations such as rotation, shifting, and enlargement and reduction; or adjust setting values ​​such as brightness value, contrast, and sharpness.

[0067] (Step S14: Feature Extraction Process)

[0068] The feature extraction unit 114 extracts the feature of each pixel as a learning feature from the learning image data 131 preprocessed in step S13 by the first neural network. At this time, the first neural network extracts the learning feature with reference to the first parameter 132 stored in the auxiliary storage device 13. In the first embodiment, it is assumed that the first neural network is a CNN.

[0069] (Step S15: Image reconstruction processing)

[0070] The image reconstruction unit 115 uses the learning feature quantity extracted in step S14 as input, binarizes the learning image data 131 through the second neural network, and generates binary image data, i.e., learning binary data. At this time, the second neural network generates the learning binary data with reference to the second parameter 133 stored in the auxiliary storage device 13. In the first embodiment, it is assumed that the second neural network is a CNN.

[0071] (Step S16: Parameter update processing)

[0072] The parameter updating unit 116 calculates the difference between the converted binary data generated in step S12 and the learned binary data generated in step S15. The parameter updating unit 116 updates the first parameter 132, which is a parameter of the first neural network, and the second parameter 133, which is a parameter of the second neural network, based on the calculated difference.

[0073] Reference Figure 6 The operation of the feature identifying device 20 according to the first embodiment will be described.

[0074] As Figure 6As a premise of the processing shown, the feature extraction unit 211 takes each of the plurality of registered image data 231 stored in the auxiliary storage device 23 as an object, extracts the feature amount of each pixel from the registered image data 231 as the object as a reference feature amount 232 , and stores it in the auxiliary storage device 23 .

[0075] Here, one or more registered image data 231 are stored in the auxiliary storage device 23 for each of the plurality of types to be processed. The registered image data 231 is associated with information indicating the type. The plurality of registered image data 231 may include the same image data as the learning image data 131 or may only include different image data.

[0076] Specifically, the feature extraction unit 211 uses the first neural network generated by the machine learning device 10 as a feature extraction model to extract the reference feature quantity 232 from the target registered image data 231. At this time, the first neural network extracts the reference feature quantity 232 with reference to the first parameter 132. Then, the feature extraction unit 211 stores the reference feature quantity 232 extracted from each registered image data 231 in association with the information indicating the type in the auxiliary storage device 23.

[0077] Thus, one or more reference feature quantities 232 are stored in the auxiliary storage device 23 for each type.

[0078] (Step S21: Feature extraction processing)

[0079] The feature extraction unit 211 extracts a feature amount for each pixel as an identification feature amount from the identification target data 233 which is image data to be identified.

[0080] Here, the recognition target data 233 is image data of any type among a plurality of types to be processed. The recognition target data 233 may be the same image data as the learning image data 131 or may be different image data.

[0081] Specifically, the feature extraction unit 211 uses the first neural network generated by the machine learning device 10 as a feature extraction model to extract identification features from the identification target data 233. At this time, the first neural network refers to the first parameter 132 to extract the identification features.

[0082] (Step S22: Similarity calculation process)

[0083] The similarity calculation unit 213 targets each of the plurality of categories and calculates the similarity between the reference feature 232 , which is a feature extracted from the image data of the target category, and the identification feature extracted in step S21 .

[0084] (Step S23: Influence calculation process)

[0085] The influence calculation unit 214 receives the similarity for each type calculated in step S22 and the identification feature extracted in step S21 as input, and calculates the influence on the similarity for each partial image of the identification target data.

[0086] (Step S24: Feature determination processing)

[0087] The feature identification unit 215 changes the identification feature extracted in step S21 according to the influence calculated in step S23, and identifies the feature part of the identification target data 233. Then, the feature identification unit 215 identifies that the identified feature part is the template in the paper document and the other parts are the entry area.

[0088] Reference Figure 7 Similarity calculation processing of implementation mode 1 ( Figure 6 Step S22) is described below.

[0089] The processing of step S31 to step S36 is executed for each type.

[0090] In step S31, the similarity calculation unit 213 obtains the reference feature quantity 232 of the object type from the auxiliary storage device 23. In step S32, the similarity calculation unit 213 obtains the reference feature quantity 232 of the object type. Figure 6 The identification feature quantity extracted in step S21.

[0091] In step S33, if the similarity calculation unit 213 reads a plurality of reference feature quantities 232 in step S31, the process proceeds to step S34, and if the similarity calculation unit 213 reads a single reference feature quantity 232 in step S31, the process proceeds to step S35. In step S34, the similarity calculation unit 213 processes the average of the reference feature quantities 232 as the reference feature quantity 232 of the object type. In addition, the similarity calculation unit 213 may determine the reference feature quantity 232 of the object type by some method without being limited to the average.

[0092] In step S35 , the similarity calculation unit 213 calculates the degree of deviation between the reference feature value 232 and the identification feature value acquired in step S32 .

[0093] As a specific example, the similarity calculation unit 213 calculates the distance between the reference feature quantity 232 and the identification feature quantity, thereby calculating the deviation. Specifically, the similarity calculation unit 213 uses the average square error of each pixel to calculate the deviation between the reference feature quantity 232 and the identification target data 233. In this case, the similarity calculation unit 213 calculates the square error of the distance between the reference feature quantity 232 and the identification target feature quantity for each pixel, and calculates the average of the square errors of all pixels as the deviation. That is, the similarity calculation unit 213 calculates the deviation as shown in Formula 1.

[0094] (Formula 1) dissimilarity j =Σ k (feature k -basis jk ) 2

[0095] Here, basis jk is the reference feature 232 for pixel k of the registered image data 231 of type j. k is the identification feature quantity for pixel k of the identification target data 233. j is the deviation for type j.

[0096] Furthermore, as another specific example, the similarity calculation unit 213 weights the mean square error to calculate the degree of deviation between the reference feature quantity 232 and the identification target data 233 .

[0097] At this time, the weight is determined for each type so that the pixel with greater influence has a larger value. Alternatively, the weight is determined for each type so that the pixel with greater deviation (standard deviation) of the feature amount of the registered image data 231 has a smaller value. Alternatively, the weight may be determined by a combination of these two methods.

[0098] That is, the similarity calculation unit 213 calculates the degree of deviation as shown in Formula 2.

[0099] (Formula 2) dissimilarity j =Σ k w jk (feature k -basis jk ) 2

[0100] Here, w jk is the weight of pixel k in the registered image data 231 of type j. jk For example, |basis jk | 2 / σjk σ jk is the standard deviation associated with pixel k of the registration image data 231 of type j.

[0101] In step S36, the similarity calculation unit 213 converts the deviation calculated in step S35 into similarity. At this time, the similarity calculation unit 213 calculates the similarity in such a way that the larger the deviation, the smaller the similarity, and the smaller the deviation, the larger the similarity. As a specific example, the similarity calculation unit 213 normalizes the deviation within the range of 0 to 1, and subtracts the normalized deviation from 1 to calculate the similarity.

[0102] In step S37 , the similarity calculation unit 213 outputs a vector having the similarities calculated for each type as elements as a similarity vector.

[0103] Reference Figure 8 Influence calculation process of implementation mode 1 ( Figure 6 Step S23) is described below.

[0104] In step S41, the influence calculation unit 214 obtains Figure 7 In step S42, the influence calculation unit 214 obtains the similarity vector output in step S37. Figure 6 The identification feature quantity extracted in step S21.

[0105] In step S43, the influence calculation unit 214 uses the similarity vector obtained in step S41 and the identification feature quantity obtained in step S42 as input, and calculates the magnitude of the influence on the similarity when each pixel of each partial image as the identification target data 233 is changed, as the influence. Specifically, the influence calculation unit 214 inputs the similarity vector and the identification feature quantity to Grad-CAM, and calculates the influence related to each pixel. That is, the influence is a gradient. Grad-CAM is described in non-patent document 1.

[0106] In step S44, the influence calculation unit 214 outputs the influence calculated in step S43. In addition, the feature quantity has multiple layers. Therefore, the influence calculation unit 214 may output the influence for each layer of the feature quantity, or may output the average of the influences of each layer of the feature quantity.

[0107] Reference Fig. 9 The feature determination process of implementation mode 1 ( Figure 6 Step S24) is described below.

[0108] In step S51, the feature identification unit 215 obtains Figure 8 In step S52, the feature determination unit 215 obtains the influence degree output in step S44. Figure 6The identification feature quantity extracted in step S21.

[0109] In step S53, the feature determination unit 215 uses the influence obtained in step S51 as a weight to weight the identification feature obtained in step S52, thereby changing the identification feature. At this time, the feature determination unit 215 performs weighting according to each layer while outputting the influence according to each layer. In step S54, the feature determination unit 215 outputs the identification feature changed in step S53 to a display device or the like via the output interface 25. The feature part of the identification object data 233 is determined using the identification feature.

[0110] ***Effects of Implementation Method 1***

[0111] As described above, the feature identification device 20 of the first embodiment calculates the similarity between the reference feature 232 and the identification feature for each type, and calculates the influence on the similarity related to each partial image (each pixel) of the identification target data 233. Then, the identification feature is changed according to the influence.

[0112] Thus, even if the reference feature 232 and the identification feature are extracted using a CNN that does not have information about the type, the characteristic part can be appropriately determined. Therefore, when the type increases, the characteristic part corresponding to the type can be appropriately determined without relearning the parameters of the CNN used to extract the reference feature and the identification feature. As a result, when the type increases, the template and the recording part can be appropriately determined without relearning the parameters of the CNN.

[0113] ***Other structures***

[0114] <Modification 1>

[0115] The machine learning device 10 only needs to satisfy the following condition: the learning image data 131 provided as the input of the feature extraction unit 114 does not contain information related to the type, and the first neural network is learned. If this condition is satisfied, the machine learning device 10 is not limited to the structure described in the first embodiment, and may be another structure.

[0116] <Modification 2>

[0117] In Embodiment 1, after integrating the plurality of reference feature quantities 232, the reference feature quantity 232 and the identification feature quantity are compared to calculate the similarity. However, after calculating the similarity with the identification feature quantity for each of the plurality of reference feature quantities 232, the similarities related to each of the plurality of reference feature quantities 232 may be integrated.

[0118] Reference Fig.10The similarity calculation process in this case ( Figure 6 Step S22) is described below.

[0119] The processing of step S61 to step S66 is executed for each type.

[0120] The processing of step S61 to step S62 is Figure 7 The processing of step S31 to step S32 is the same as that of step S31. In step S63, the similarity calculation unit 213 calculates the degree of deviation between each reference feature quantity 232 acquired in step S61 and the identification feature quantity acquired in step S62.

[0121] In step S64, if the similarity calculation unit 213 reads a plurality of reference feature quantities 232 in step S61, the process proceeds to step S65, and if one reference feature quantity 232 is read in step S61, the process proceeds to step S66. In step S65, the similarity calculation unit 213 processes the average of the deviations from the reference feature quantities 232 calculated in step S63 as the deviation of the object type. In addition, the similarity calculation unit 213 may determine the deviation of one object type by some method without being limited to the average.

[0122] The processing of steps S66 to S67 is Figure 7 The processing of steps S36 to S37 is the same.

[0123] <Variation 3>

[0124] In the first embodiment, each functional component is implemented by software. However, as a modification example 3, each functional component may be implemented by hardware. Regarding this modification example 3, the differences from the first embodiment will be described.

[0125] The structures of the machine learning device 10 and the feature identification device 20 according to the third modification are described.

[0126] When each functional structural element is realized by hardware, the machine learning device 10 has an electronic circuit instead of the processor 11, the memory 12, and the auxiliary storage device 13. The electronic circuit is a dedicated circuit that realizes the functions of each functional structural element, the memory 12, and the auxiliary storage device 13. Similarly, when each functional structural element is realized by hardware, the feature determination device 20 has an electronic circuit instead of the processor 21, the memory 22, and the auxiliary storage device 23. The electronic circuit is a dedicated circuit that realizes the functions of each functional structural element, the memory 22, and the auxiliary storage device 23.

[0127] As the electronic circuit, a single circuit, a complex circuit, a programmed processor, a parallel programmed processor, a logic IC, a GA (Gate Array), an ASIC (Application Specific Integrated Circuit), and an FPGA (Field-Programmable Gate Array) can be envisioned.

[0128] Each functional component may be realized by one electronic circuit, or may be realized by distributing the functional components by a plurality of electronic circuits.

[0129] <Variation 4>

[0130] As a fourth modification, some of the functional components may be implemented by hardware, and the other functional components may be implemented by software.

[0131] The processors 11 and 21, the memories 12 and 22, the auxiliary storage devices 13 and 23, and the electronic circuit are referred to as a processing circuit. That is, the functions of each functional component are realized by the processing circuit.

[0132] Description of symbols

[0133] 10: machine learning device; 11: processor; 12: memory; 13: auxiliary storage device; 14: communication interface; 111: sampling unit; 112: binary image conversion unit; 113: image generation unit; 114: feature extraction unit; 115: image reconstruction unit; 116: parameter updating unit; 131: learning image data; 132: first parameter; 133: second parameter; 20: feature determination device; 21: processor; 22: memory; 23: auxiliary storage device; 24: communication interface; 211: feature extraction unit; 212: feature change unit; 213: similarity calculation unit; 214: influence calculation unit; 215: feature determination unit; 231: registration image data; 232: baseline feature; 233: identification object data; 30: communication path; 100: feature determination system.

Claims

1. A feature determination device, the feature determination device comprising: a similarity calculation unit that takes each of the plurality of types as an object and calculates the similarity between a reference feature quantity extracted from image data of the object type and a recognition feature quantity extracted from image data that is a recognition object, i.e., recognition object data; an influence calculation unit that takes the similarities related to each type calculated by the similarity calculation unit and the recognition feature quantity as input, and calculates the influence of each partial image of the recognition target data on the similarities; as well as A feature specifying unit changes the identification feature amount according to the influence calculated by the influence calculating unit.

2. The feature determination device according to claim 1, wherein: The similarity calculation unit calculates the similarity by calculating the distance between the reference feature value and the identification feature value.

3. The feature determination device according to claim 1, wherein: The influence degree calculation unit calculates, for each of the types, the magnitude of the influence on the similarity when each of the partial images of the recognition target data is changed as the influence degree.

4. The feature determination device according to claim 2, wherein: The influence degree calculation unit calculates, for each of the types, the magnitude of the influence on the similarity when each of the partial images of the recognition target data is changed as the influence degree.

5. The feature determination device according to claim 1, wherein: The feature specifying unit weights the identification feature amount using the influence degree as a weight, thereby changing the identification feature amount.

6. The feature determination device according to claim 2, wherein: The feature specifying unit weights the identification feature amount using the influence degree as a weight, thereby changing the identification feature amount.

7. The feature determination device according to claim 3, wherein: The feature specifying unit weights the identification feature amount using the influence degree as a weight, thereby changing the identification feature amount.

8. The feature determination device according to claim 4, wherein: The feature specifying unit weights the identification feature amount using the influence degree as a weight, thereby changing the identification feature amount.

9. The feature determination device according to any one of claims 1 to 8, wherein: The reference feature amount and the identification feature amount are extracted by a feature extraction model having no information on the plurality of types.

10. The feature determination device according to claim 9, wherein: The feature extraction model is a first neural network generated by a machine learning device, and the machine learning device updates the parameters of the first neural network and the second neural network based on the difference between converted binary data and learned binary data. The converted binary data is binary image data obtained by converting image data as a learning object, namely, learning data, and the learned binary data is binary image data obtained by converting the learning data by the second neural network based on the feature quantity of the learning data extracted by the first neural network, namely, the learning feature quantity.

11. A method for determining a feature, wherein: The similarity calculation unit takes the plurality of types as objects, respectively, and calculates the similarity between a reference feature quantity and a recognition feature quantity, wherein the reference feature quantity is a feature quantity extracted from image data of the object type, and the recognition feature quantity is a feature quantity extracted from image data as a recognition object, i.e., recognition object data. The influence calculation unit takes the similarity related to each type and the identification feature as input, and calculates the influence of each partial image of the identification target data on the similarity. The feature specifying unit changes the identification feature amount according to the influence degree.

12. A computer-readable storage medium storing a feature determination program, the feature determination program causing a computer to function as a feature determination device, the feature determination device executing the following processing: A similarity calculation process is performed to calculate the similarity between a reference feature quantity and a recognition feature quantity, wherein the reference feature quantity is a feature quantity extracted from image data of the object type, and the recognition feature quantity is a feature quantity extracted from image data of the recognition object, i.e., recognition object data; an influence calculation process of taking the similarities related to each type calculated by the similarity calculation process and the identification feature quantity as input, and calculating the influence of each partial image of the identification target data on the similarities; and The feature specifying process changes the identification feature amount according to the influence calculated by the influence calculating process.

Images