JP2020166397A - Image processing device, image processing method, and program - Google Patents

Abstract

To prevent wasteful increase in learning data when creating learning data for machine learning by expansion.SOLUTION: An image processing device includes: an identification unit (12) that identifies an expansion target from image data; an expansion unit (13) that generates expansion data obtained by expanding the expansion target of the image data by image processing; a calculation unit (14) that calculates a score of the expansion data by class-classification; and a selection unit (15) that selects whether or not to use the expansion data as learning data according to the obtained score.SELECTED DRAWING: Figure 1

Description

The present disclosure relates to image processing devices, image processing methods, and programs that extend learning data for machine learning.

Patent Document 1 describes an image processing device that generates teacher data (learning data) used in machine learning. In machine learning, it is preferable to use a wide variety of learning data because it affects the result of obtaining the distribution range of the learning data in the feature space. However, it is difficult to acquire and acquire a wide variety of learning data because it takes time and effort for shooting, and as described in Patent Document 1, a method of creating a wide variety of learning data is available. It is being considered.

JP-A-2018-163554

The present disclosure provides an image processing apparatus, an image processing method, and a program that select valid data and add only valid data when adding learning data for machine learning by data expansion.

The image processing apparatus of the present disclosure calculates a score of extended data by class classification, a specific unit that identifies an extended target from image data, an extended unit that generates extended data obtained by expanding the extended target of image data by image processing. It includes a calculation unit and a selection unit that selects whether or not to use the extended data as training data according to the obtained score.

These general and specific aspects may be realized by systems, methods, and computer programs, and combinations thereof.

According to the image processing apparatus, the image processing method, and the program of the present disclosure, it is possible to prevent an increase in unnecessary learning data when creating learning data for machine learning by data expansion.

It is a block diagram explaining the structure of the image processing apparatus which concerns on this disclosure. This is an example of the original image data to be expanded. This is an example of a mask image in which the mask area to be expanded is specified. This is an example of extended data generated from the image data of FIG. It is another example of the extended data generated from the image data of FIG. It is another example of the extended data generated from the image data of FIG. It is an example of the classification score calculated from the extended data of FIG. 4A. It is a flowchart explaining the processing of an image processing method. It is a flowchart explaining the selection process of the flowchart of FIG.

[Findings underlying this disclosure]
In recent years, machine learning has become widespread in various fields. In machine learning, a wide variety of learning data is required because the distribution range of the learning data in the feature space determines the result. However, it is difficult to create learning data by taking an image or the like due to physical problems. Therefore, data extensions such as data augmentation are used.

However, in data augmentation, image data that cannot be expected in reality may be generated. That is, since a random parameter is used in a data expansion technique such as data augmentation, there is a possibility that data irrelevant to the accuracy of machine learning is created as learning data. When data irrelevant to the accuracy of machine learning is created in this way, not only the accuracy cannot be improved, but also the processing capacity increases due to the increase in the number of learning data.

The present disclosure provides an image processing device, an image processing method, and a program capable of selecting valid data from the extended data and preventing an increase in unnecessary learning data when adding machine learning learning data by data expansion. provide. As a result, it is possible to realize the use of learning data that can improve the processing efficiency, instead of using an unnecessarily large amount of learning data.

[Embodiment]
Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. However, in the detailed description, unnecessary parts of the description of the prior art and substantially the same configuration may be omitted. This is for the sake of simplicity. In addition, the following description and accompanying drawings are disclosed so that those skilled in the art can fully understand the present disclosure, and are not intended to limit the subject matter of the claims.

The image processing apparatus, image processing method, and program according to the present disclosure generate learning data for machine learning by extension. In the following, the image processing device will be described as an example of expanding the learning data used for detecting a person. Therefore, in the following description, the target of image processing in the image processing apparatus is image data including a person. Specifically, a part of the original image data including a person will be described with an example of generating extended data in which the texture of any part of the person is changed.

In the present disclosure, "extension" means inversion, rotation, resizing, color of the target portion, change of texture, and the like. Inversion, rotation, and resizing of a person's image include changing the tilt of the person and changing the size of the person. Changing the size of a person includes changing a person of normal physique to a thinner or thicker physique, or changing the person to be taller or shorter. Changing the color of a person's image shall include changing the color of the person's skin and changing the color of the person's clothing and belongings.
Further, "texture" means changing the appearance texture of the image, and includes changing the pattern. For example, changing the texture of a person's image includes changing the material, pattern, pattern, etc. of clothing or the person's belongings.

<Image processing device>
As shown in FIG. 1, the image processing device 1 is an information processing device including a control unit 10, a storage unit 20, a communication interface (I / F) 21, an input unit 22, an output unit 23, and the like.

The control unit 10 is a controller that controls the entire image processing device 1. For example, the control unit 10 reads and executes the image processing program P stored in the storage unit 20 to serve as the reception unit 11, the specific unit 12, the expansion unit 13, the calculation unit 14, the selection unit 15, and the estimation unit 16. Achieve the processing of. Further, the control unit 10 is not limited to one that realizes a predetermined function by the cooperation of hardware and software, and may be a hardware circuit specially designed to realize a predetermined function. That is, the control unit 10 can be realized by various processors such as a CPU, MPU, GPU, FPGA, DSP, and ASIC.

The storage unit 20 is a recording medium for recording various information. The storage unit 20 is realized by, for example, a RAM, a ROM, a flash memory, an SSD (Solid State Device), a hard disk, another storage device, or a combination thereof as appropriate. In addition to the image processing program P executed by the control unit 10, various information acquired for image processing and the like are stored in the storage unit 20. For example, the storage unit 20 stores the condition data D1 used in the image processing, the original image data D2, the extended data D3 used in the image processing, and the learning data D4.

The communication I / F 21 is an interface circuit (module) for enabling data communication with an external device (not shown). The input unit 22 is an input means such as an operation button, a keyboard, a mouse, a touch panel, and a microphone used for operation and data input. The output unit 23 is an output means such as a display or a speaker used for outputting processing results and data.

The image processing device 1 may be realized by a plurality of information processing devices. Further, a part of the data stored in the storage unit 20 may be stored in an external storage device and read from the external storage device for use.

The condition data D1 is data input as a condition for data expansion in the image processing apparatus 1, and a specific example will be described later.

The original image data D2 is image data used for data expansion in the image processing device 1. For example, as shown in FIG. 2, the original image data D2 includes a person as a subject.

The extended data D3 is data expanded by image processing in the image processing apparatus 1, and specific examples will be described later.

The learning data D4 is data that serves as teacher data in machine learning. In the image processing device 1, the extended data D3 expanded by the image processing is added to the learning data D4.

The reception unit 11 accepts conditions used for image processing, specifically, expansion of image data. Further, the reception unit 11 stores the received condition as the condition data D1 in the storage unit 20.

For example, the condition can include a method of identifying the target of extension in the specific unit 12. The identification of the expansion target can include what kind of area of the image data is to be expanded. The condition can include a method of expanding data in the expansion unit 13. Extension methods include image inversion, rotation, resizing, texture change, and the like. The condition can include a method of selecting the extended data D3 in the selection unit 15. The selection method includes the accuracy of the data required for the learning data D4.

Further, the reception unit 11 receives the original image data to be expanded and stores it in the storage unit 20 as the original image data D2. For example, the reception unit 11 can use the image data received from an external device (not shown) via the communication I / F 21 as the original image data. The number of image data received by the reception unit 11 is not limited. In this case, the image processing device 1 may process a plurality of image data.

The identification unit 12 specifies the target of expansion from the original image data D2 of the expansion target. Here, an example of generating extended data with a modified texture will be described for a certain area of image data. In such a case, the specific unit 12 extracts the region to be expanded from the original image data D2. Further, the specific unit 12 outputs the specific result to the expansion unit 13. As for the region extraction method, the existing region extraction technology is used. For example, an example of a region extraction technique is semantic segmentation. Further, the specific unit 12 may correspond to a plurality of region extraction technologies and may be able to select the region extraction technology to be used according to the condition data D1.

FIG. 3 is an example showing an region to be expanded specified from the original image data D2 shown in FIG. In the example shown in FIG. 3, a mask image in which the skirt portion, which is the undergarment worn by a person, is extracted as the mask region M from the original image data D2 in FIG. 2 is shown as an example, but the mask region specified by the specific portion 12 is shown. Is not limited. For example, a portion such as a T-shirt, which is a top worn by a person, may be extracted as a mask area.

The expansion unit 13 generates the expansion data D3 in which the expansion target of the original image data D2 is expanded by image processing. Specifically, the expansion unit 13 executes processing such as inversion, rotation, resizing, color change, and texture change on the original image data D2 according to the condition data D1. Here, an example of changing the color and texture of the mask area M specified by the specific unit 12 will be described. The extension unit 13 adds, for example, a color or texture randomly selected using a random parameter to the mask area M. As for the data expansion method, the existing data expansion technology is used. For example, alpha blending is an example of a data extension technology that changes colors and textures.

4A to 4C show an example of extended data D3 in which the color or texture of the mask region M selected in FIG. 3 is randomly changed. FIG. 4A is an example in which the color of the portion of the mask area M is changed, and FIGS. 4B and 4C are examples in which the pattern of the mask area M is changed. Although 3 patterns of expansion are shown in FIGS. 4A to 4C, the expansion unit 13 can generate expansion data D3 of more patterns, for example, 1 using a random parameter. It is possible to generate 10,000 patterns of image data. For example, the expansion unit 13 generates the expansion data D3 in the number specified by the condition data D1. That is, the user can freely set the number of extended data D3 to be generated in the extended unit 13.

At this time, the condition data D1 may include generating or not generating the extended data D3 under predetermined conditions. Specifically, when the texture is given in the extended data D3, it can be a condition that black is not included, or it can be a condition that black is included. Further, it may be a condition that the extended data D3 contains a large number of predetermined textures. Specifically, the condition can be such that the extended data D3 including black color is generated at a predetermined ratio (for example, 50%) when the texture is applied.

The calculation unit 14 calculates the classification score of the classification for the extended data D3. Specifically, the calculation unit 14 is a class classifier that has been trained by the reference learning data, and calculates the classification score of each class of the extended data D3. An existing classifier can be used for the calculation unit 14. For example, as an example of the classifier, VGG16, Inception V3 and the like can be used.

It is preferable that the calculation unit 14 has already been trained by the image group that statistically covers the attributes of the input image. Specifically, when the input image is an image of a "pedestrian", the calculation unit 14 needs to have been trained by a wide variety of images of "humans". Further, when the input image is an image of "carrot", the calculation unit 14 needs to have learned from a wide variety of images of "vegetables". Further, when the input image is an image of a "sports car", the calculation unit 14 needs to have been trained by a large number of various "automobile" images. Here, the calculation unit 14 calculates the classification score for all the extended data D3 expanded by the expansion unit 13. If the learning data D4 before the extension data D3 is added is used as the learning data used for learning by the calculation unit 14, the score of the data similar to the learning data D4 before the extension data D3 is added becomes high. Ultimately, if you choose a high score, you may overfit, and if you choose a low score, you may select completely unnecessary data. Therefore, it is preferable that the calculation unit 14 avoids using the learning data D4 before adding the extended data D3 as the learning data. Specifically, the learning data used as the reference in the calculation unit is just a natural image, or the above-mentioned examples (“pedestrian” and “human”, “carrot” and “vegetable”, “sports car”, etc. It is necessary to have a relationship of "automobile" etc.).

FIG. 5 shows an example of the classification score obtained from the extended data D3 of FIG. 4A. Since the extended data D3 of FIG. 4A was obtained from the original image data D2 of a person, a high value of "0.9043" was calculated as the classification score of the class of the person ('human'). In addition, high values were calculated for the classification score of the glasses ('glasses') class and the classification score of the skirt ('skirt') worn by the person in the image. On the other hand, since the other classes shown in FIG. 5 are classes unrelated to the extended data D3, the calculated classification score value is low.

The selection unit 15 selects the data to be used as the learning data D4 for the expansion data D3 generated by the expansion unit 13. Specifically, the selection unit 15 selects data to be used as learning data D4 from the extended data D3 according to the classification score calculated by the calculation unit 14 and the conditions specified in advance. Further, the selection unit 15 stores the selected extended data D3 in the storage unit 20 as learning data D4 together with the calculation result of the calculation unit 14.

《Classification score》
The selection unit 15 classifies using the classification score calculated by the calculation unit 14, and there are a plurality of possible methods for using the classification score. Specifically, it is conceivable to use the classification score of a single class calculated by the calculation unit 14, or to calculate and use a new score from the classification scores of a plurality of classes calculated by the calculation unit 14. Further, whether to use the classification score of a single class or the classification score of a plurality of classes is specified in the condition data D1.

(1.1) When Referencing Only the Classification Score of a Single Class The selection unit 15 may select the extended data D3 by using the classification score of one class. In this case, the selection unit 15 uses only the classification score of the class specified by the condition data D1 among the classification scores of each class calculated by the calculation unit 14.

(1.2) When referring to the classification scores of a plurality of classes The selection unit 15 may select the extended data D3 by using the classification scores of a plurality of classes. In this case, the selection unit 15 calculates and uses a new score using the classification scores of a plurality of classes specified in the condition data D1 among the classification scores of each class calculated by the calculation unit 14. For example, the selection unit 15 uses the value obtained by weighting addition as shown in the following equation (1) as a new score.

'human_new'
= Whuman *'human' + Wglasses *'glasses' + Wskirt *'skirt'
... (1)
However,
'human_new': New score calculated by selection 15
'human','glasses','skirt': Classification score calculated by the calculation unit 14 Whuman, Wglasses, Wskirt: Weight value of each classification score.

Equation (1) is a score of a person weighted by a weight value ('human'), a score of glasses ('glasses'), and a score of a skirt ('skirt'), respectively, when calculating a new person's score. Is an example in which is added.

<< Selection method >>
The selection unit 15 selects the extended data D3 using the classification score, and a plurality of selection methods can be considered. This selection method differs depending on the purpose for which the learning data D4 is increased. Specifically, it is conceivable that only rare image data is to be added, only common image data is to be added, or randomly selected image data is added. The purpose of adding the extended data D3 to the training data D4 is specified by the condition data D1.

(2.1) When adding only rare image data For example, when it is desired to add rare image data to the training data D4, it is conceivable to add image data having a low classification score. A rare image is an image of a situation that is generally not expected. In such a case, for example, the selection unit 15 arranges a plurality of extended data D3s in descending order, selects the extended data D3 having a predetermined ratio (for example, 15%) that is lower, and adds it to the learning data D4. The predetermined ratio used here is specified in the condition data D1. Further, although the example in which the extended data D3 is arranged in descending order of the classification score has been described here, the same applies even if the extended data D3 is processed in ascending order.

In addition, the selection unit 15 may select the extended data D3 in which the score of the designated class is equal to or less than a predetermined value. In this case, the number of extended data D3 to be selected may be determined in advance, or all may be selected. The score and selection method used in this case are specified in the condition data D1.

(2.2) When adding common image data For example, when it is desired to add common image data to the training data D4, it is conceivable to add image data having a high classification score. A common image is an image of a common situation that is not uncommon in general. In such a case, the selection unit 15 arranges the plurality of extended data D3s in descending order, selects the extended data D3 having a higher predetermined ratio (for example, 15%), and adds it to the learning data D4. The predetermined ratio used here is specified in the condition data D1. Further, although the example in which the extended data D3 is arranged in descending order of the classification score has been described here, the same applies even if the extended data D3 is processed in ascending order.

In addition, the selection unit 15 may select the extended data D3 in which the score of the designated class is equal to or higher than a predetermined value. In this case, the number of extended data D3 to be selected may be determined in advance, or all may be selected. The score and selection method used in this case are specified in the condition data D1.

(2.3) When randomly selected and added For example, when the purpose is to simply increase the training data D4 without limiting the classification score of the image data to be added, the image data is randomly selected and added to the training data D4. to add. For example, when 10,000 expansion data D3s are generated in the expansion unit 13, the selection unit 15 arranges all the expansion data D3s in descending order of the classification score, for example, randomly from the top 0 to 20% of 2000 sheets. 500 sheets, 400 sheets randomly from 2000 sheets of the top 20-40%, 400 sheets randomly from 2000 sheets of the top 40-60%, 300 sheets randomly from 2000 sheets of the top 60-80%, top 80-100% 200 sheets are randomly selected from 2000 sheets. As a result, the selection unit 15 can add various variations of the extended data D3 to the learning data D4 with a small number of sheets. In this case as well, the same applies when the extended data D3 is arranged in ascending order of the score. The selection method when the extended data is randomly selected is defined in advance in, for example, the image processing program P. Alternatively, the selection method may be specified by the condition data D1.

The estimation unit 16 executes the classification of the test data D5 by using the training data D4 after the extension data D3 is added by the selection unit 15. Further, the estimation unit 16 compares the obtained classification score with a predetermined threshold value or more, determines that the learning data D4 having the desired estimation accuracy is obtained when the threshold value or more is equal to or higher than the threshold value, and determines that the learning data D4 having the desired estimation accuracy is not obtained. It is determined that the learning data D4 with the estimation accuracy of is not obtained. For example, the estimation unit 16 sets a threshold value to be compared with the classification score of a certain class and compares it with the classification score obtained for this class. Thereby, it can be determined whether or not the updated learning data D4 can execute the process with the desired estimation accuracy. Further, when the estimation unit 16 determines that the learning data D4 having the desired estimation accuracy has not been obtained, the extended data D3 can be newly created and the process can be repeated.

As described above, when the image processing device 1 generates the extended data to be used as the training data, only the extended data selected according to the calculated score can be added as the training data. Therefore, it is possible to prevent unnecessary increase of learning data, and efficient processing can be realized.

<Image processing method>
Next, the image processing method executed by the image processing apparatus 1 will be described with reference to the flowcharts shown in FIGS. 6 and 7.

The image processing device 1 accepts a condition relating to data expansion (S1). The conditions include information such as the target and method of data expansion, the number of extended data to be generated, and the accuracy required for the extended data. At this time, the image processing device 1 registers the received condition as the condition data D1 in the storage unit 20.
Further, the image processing device 1 receives the original image data and registers it in the storage unit 20 (S2).

The image processing device 1 reads the original image data D2 from the storage unit 20 and identifies the target of expansion (S3). For example, the image processing device 1 generates an image that specifies a mask area of an area to be expanded.

The image processing device 1 generates the extended data D3 (S4). At this time, the image processing device 1 generates a predetermined number of extended data D3 specified as a condition in step S1 and stores them in the storage unit 20.

The image processing device 1 calculates a classification score for each extended data D3 generated in step S4 (S5). Further, the image processing device 1 stores the calculated classification score in the storage unit 20 in association with each extended data D3.

After that, the image processing apparatus 1 refers to the classification score calculated in step S5 and selects the extended data D3 to be added to the learning data D4 (S6). The selection process of the extended data D3 will be described later using the flowchart shown in FIG.

The image processing device 1 adds the extended data D3 selected in step S6 to the learning data D4 (S7).

The image processing device 1 executes the learning process with the new learning data D4 to which the extended data D3 is added in step S7 (S8).

If the accuracy obtained in step S8 is not equal to or greater than the threshold value (NO in S9), the desired learning data D4 could not be obtained, so the image processing apparatus 1 returns to step S1 and repeats the processes of steps S1 to S9. When the same condition data D1 and original image data D2 are used, the process may be repeated by returning to step S3.

When the accuracy obtained in step S9 is equal to or greater than the threshold value (NO in S9), the image processing device 1 ends because the desired learning data D4 has been obtained.

<< Selection process >>
As shown in FIG. 7, in the selection process in step S6 of the flowchart of FIG. 6, the image processing device 1 first determines whether or not to use the classification scores of a plurality of classes (S61). Specifically, the image processing device 1 determines by referring to the condition of the condition data D1 stored in the storage unit 20.

When the classification scores of a plurality of classes are used (YES in S61), the image processing apparatus 1 calculates a new score by using the classification scores of the plurality of classes (S62). Specifically, the image processing device 1 calculates a new score by a calculation method included in the condition data D1 stored in the storage unit 20.

On the other hand, when the classification scores of the plurality of classes are not used (NO in S61), the image processing apparatus 1 proceeds to step S63. At this time, the classification score of one class specified by the condition data D1 stored in the storage unit 20 is used.

The image processing device 1 uses the new score calculated in step S62 or the classification score of one class specified in the condition data D1 stored in the storage unit 20, and a plurality of extensions stored in the storage unit 20. The data D3 is arranged in descending order of score (S63).

When the condition data D1 specifies that a rare image should be added to the training data D4 (YES in S64), the image processing apparatus 1 selects the extended data D3 having a lower predetermined ratio arranged in descending order in step S63. (S65). Here, the predetermined ratio to be selected is also defined by the condition data D1.

When it is not specified to add a rare image to the training data D4 (YES in S64), whether or not the image processing device 1 is specified to add a common image in the condition data D1 to the training data D4. (S66).

When it is specified to add a common image to the training data D4 in the condition data D1 (YES in S66), the extended data D3 having a higher predetermined ratio arranged in descending order is selected in step S63 (S67). Here, the predetermined ratio to be selected is also defined by the condition data D1.

On the other hand, when it is not specified to add a common image in the condition data D1 to the training data D4 (NO in S66), the extended data D3 is selected by the method specified in the condition data D1 (S68).

When the extended data D3 is selected by any of the methods (S65, S67, S68), the process proceeds to step S7 of the flowchart of FIG.

As described above, when the image processing device 1 generates the extended data to be used as the training data, only the extended data selected according to the calculated score can be added as the training data. Therefore, it is possible to prevent unnecessary increase of learning data, and efficient processing can be realized.

[Effects and supplements]
As described above, the above-described embodiment has been described as an example of the technology disclosed in this application. However, the technique in the present disclosure is not limited to this, and can be applied to embodiments in which changes, replacements, additions, omissions, etc. are made as appropriate. Therefore, other embodiments will be illustrated below.

<< Modification example >>
(1) Improving the accuracy of class classification By training the calculation unit 14 in advance using desired image data, the accuracy of classification can be improved. Therefore, an additional learning unit may be provided to access an external database or the like that holds a large amount of images to acquire effective image data, and the acquired plurality of image data may be used for learning. In this case, the additional learning unit acquires image data under the specified conditions from a plurality of external devices and the like, and learns the acquired image data as learning data.

(2) Judgment of estimation accuracy In the image processing device 1 described above, the estimation unit 16 estimates with the test data D5 using the training data D4 to which the extended data D3 selected by the selection unit 15 is added, but the reception unit 11 The original image data D2 received in 1 may be estimated by the estimation unit 16, and expansion by the expansion unit 13 may be executed when the original image data D2 is equal to or greater than a predetermined threshold value. That is, when data expansion is executed using image data that does not originally satisfy the conditions, there is a high possibility that the obtained extended data D3 does not satisfy the conditions required for the learning data D4. Therefore, by using only the original image data D2 that satisfies the conditions in advance, it is possible to prevent the execution of unnecessary processing. Specifically, the estimation unit 16 obtains the classification score of each class for the original image data D2 by class classification, and when the classification score for a predetermined class is equal to or higher than a predetermined threshold value, the extension unit 13 processes it. May be executed. As a result, unnecessary processing can be reduced, and efficient addition of learning data D4 can be realized.

<< Outline of Embodiment >>
(1) The image processing apparatus of the present disclosure includes a specific unit that identifies an expansion target from image data, an extension unit that generates expansion data obtained by expanding the expansion target of image data by image processing, and a score of the expansion data by classification. It is provided with a calculation unit for calculating the above and a selection unit for selecting whether or not to use the extended data as training data according to the obtained score.

As a result, when the expanded data is used as the learning data, it is determined whether or not the expanded data is used as the learning data according to the score, so that it is possible to prevent unnecessary increase of the learning data.

(2) In the image processing apparatus of (1), the specific unit may extract an area to be expanded from the image data, and the expansion unit may expand the area extracted by the specific unit from the image data.

As a result, it is possible to create learning data in which a specific area of image data is expanded.

(3) In the image processing apparatus of (2), the extension portion may change the color or texture of the area.

As a result, the variation of the training data can be increased only by changing only the color and texture of the area.

(4) In the image processing apparatus of (2), the expansion unit may invert, rotate, or resize the expansion target.

As a result, the variation of the training data can be increased only by reversing and rotating the target.

(5) In the image processing apparatus of (1), when a plurality of extended data are arranged in descending or ascending order of the score, the selection unit may select a predetermined number of upper or lower extended data.

This makes it possible to learn the desired accuracy.

(6) In the image processing apparatus of (1), when a plurality of extended data are arranged in descending or ascending order of the score, the selection unit may select a specified number of extended data from each range classified by the score. ..

This makes it possible to learn the desired accuracy.

(7) In the image processing apparatus of (1), when the extension condition is input by the user, the extension unit may process the image data under the extension condition.

As a result, it is possible to generate extended data under the specified conditions, and it is possible to generate data of a desired variation.

(8) The image processing method of the present disclosure includes a step of specifying an expansion target from image data, a step of generating expansion data obtained by expanding the expansion target of the image data by image processing, and a score of the expansion data by classification. It has a step of calculating the above and a step of selecting whether or not to use the extended data as training data according to the obtained score.

As a result, when the expanded data is used as the learning data, it is determined whether or not the expanded data is used as the learning data according to the score, so that it is possible to prevent unnecessary increase of the learning data.

The image processing apparatus and image processing method described in all claims of the present disclosure are realized by cooperation with hardware resources such as a processor, a memory, and a program.

The image processing apparatus and image processing method of the present disclosure are useful for extending learning data of machine learning.

1 Image processing device 10 Control unit 11 Reception unit 12 Specific unit 13 Expansion unit 14 Calculation unit 15 Selection unit 16 Estimate unit 20 Storage unit 21 Communication I / F
22 Input unit 23 Output unit D1 Condition data D2 Original image data D3 Extended data D4 Learning data

Claims (9)

A specific part that identifies the extension target from the image data,
An extension unit that generates extended data obtained by expanding the expansion target of the image data by image processing,
A calculation unit that calculates the score of the extended data by classification,
A selection unit that selects whether or not to use the extended data as training data according to the obtained score, and
An image processing device comprising. The specific unit extracts an area to be expanded from the image data and
The image processing apparatus according to claim 1, wherein the expansion unit expands a region extracted by the specific unit in the image data. The image processing apparatus according to claim 2, wherein the expansion unit changes the color or texture of the region. The image processing apparatus according to claim 2, wherein the expansion unit inverts, rotates, or resizes the expansion target. The image processing apparatus according to any one of claims 1 to 4, wherein the selection unit selects a predetermined number of upper or lower extended data when the plurality of extended data are arranged in descending or ascending order of the score. .. The selection unit according to any one of claims 1 to 4, wherein when a plurality of the extended data are arranged in descending or ascending order of the score, a specified number of the extended data is selected from each range classified by the score. Image processing equipment. The image processing apparatus according to any one of claims 1 to 6, wherein the expansion unit processes the image data under the expansion condition when the expansion condition is input by the user. Steps to identify the extension target from the image data and
A step of generating extended data obtained by expanding the extended target of the image data by image processing, and
Steps to calculate the score of the extended data by classification and
A step of selecting whether or not to use the extended data as training data according to the obtained score, and
Image processing method having. An image processing program that causes a computer to execute the image processing method according to claim 8.

Images