Machine learning image creation method, machine learning method, machine learning image creation device, and machine learning image creation program

Abstract

In machine learning image creation, when reducing the image quality of a teacher image that includes a bright spot to create a student image that forms a pair with the teacher image, this machine learning image creation method does not degrade the image quality within a prescribed range from the bright spot as compared with regions outside said prescribed range from the bright spot.

Description

Machine learning image creation method, machine learning method, machine learning image creation device, and machine learning image creation program

[0001] The present invention relates to a machine learning image creation method, a machine learning method, a machine learning image creation device, and a machine learning image creation program.

[0002] Conventionally, a technique for super-resolution using a machine learning model has been known. For example, the machine learning model performs machine learning using training data that pairs a high-resolution teacher image with a low-resolution student image created by blurring the teacher image. Using the machine learning model that performs machine learning using such training data, a high-resolution image that appears to have been captured by a high-resolution image sensor is generated from an input image captured by a low-resolution image sensor through image processing.

[0003] Furthermore, Japanese Patent Application Laid-Open Publication No. 2020-141908 discloses a machine learning model that performs machine learning using multiple pairs of learning data, each pair consisting of an image containing noise as input data and an image containing no noise as output data, to accurately reduce noise and improve image quality.

[0004] However, images captured using medical equipment may contain bright spots with extremely high pixel brightness. As mentioned above, when creating a student image, it is necessary to blur the teacher image to degrade it. Therefore, if the teacher image contains bright spots, the student image will be created with the bright spots also blurred.

[0005] However, adding blur to bright spots can cause problems such as dark coloring around the bright spots in the inferred image. Specifically, black ringing artifacts appear around the bright spots in the inferred image.

[0006] The present invention has been made in consideration of the above-mentioned circumstances, and aims to provide a machine learning image creation method, a machine learning method, a machine learning image creation device, and a machine learning image creation program that, when inferring a high-resolution image from a low-resolution image that includes bright spots, can make the area around the bright spots in the inferred image more natural.

[0007] One aspect of the present invention is a method for creating images for machine learning, in which when creating images for machine learning by lowering the image quality of a teacher image that includes a bright spot to create a student image that is a pair with the teacher image, the image quality within a specified range from the bright spot is not lowered compared to the image quality outside the specified range from the bright spot.

[0008] A machine learning method according to one aspect of the present invention performs machine learning using a teacher image containing the above-described bright spots and a student image containing bright spots created by the above-described method for creating images for machine learning.

[0009] In addition, one embodiment of the image creation device for machine learning of the present invention includes a bright spot detection unit that detects bright spots from a teacher image, and an image quality processing unit that, when reducing the image quality of the teacher image to create a student image that is paired with the teacher image, does not lower the image quality within a specified range from the bright spot compared to outside the specified range from the bright spot.

[0010] Another aspect of the image creation device for machine learning of the present invention includes a processor that detects bright spots from a teacher image and, when reducing the image quality of the teacher image to create a student image that is a pair with the teacher image, does not reduce the image quality within a specified range from the bright spot compared to outside the specified range from the bright spot.

[0011] In addition, one embodiment of the image creation program for machine learning of the present invention causes a computer to perform a process of detecting bright spots from a teacher image, and a process of not lowering the image quality of a specified range from the bright spot compared to an area outside the specified range from the bright spot when lowering the image quality of the teacher image to create a student image that is a pair with the teacher image.

[0012] 1 is a diagram showing a general flow when a machine learning model M1 learns to improve the resolution of an image using a machine learning method to become a trained model M2 and performs inference, according to the first embodiment; FIG. 2 is a diagram showing an example in which the machine learning model M is configured using a neural network; FIG. 3 is a flowchart for explaining an example of the flow of the image creation process of the first embodiment; FIG. 4 is a flowchart for explaining another example of the flow of the image creation process of the first embodiment; FIG. 5 is a block diagram showing an example of the configuration of an image creation device according to the first embodiment of the present invention; FIG. 6 is a diagram for explaining the configuration of a bright spot detection unit and an image quality processing unit of the first embodiment; FIG. 7 is a diagram for explaining the configuration of a bright spot detection unit and an image quality processing unit of the second embodiment; FIG. 8 is a diagram for explaining an example of an area in which bright spots are detected; FIG. 9 is a diagram for explaining the configuration of a bright spot detection unit and an image quality processing unit of the third embodiment.

[0013] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. (First Embodiment) Fig. 1 is a diagram showing a general flow according to the first embodiment, in which a machine learning model M1 undergoes learning to improve image resolution using a machine learning method to become a trained model M2, and then performs inference. Note that in the specification and drawings, for the sake of simplicity, the "machine learning model" may be simply referred to as the "model."

[0014] The machine learning model (learning model) M1 is a mathematical model such as a neural network (NN). Fig. 2 is a diagram showing an example of the machine learning model M1 configured using a neural network.

[0015] The neural network includes an input layer IL, a hidden layer HL, and an output layer OL. Each layer includes multiple neurons (units / nodes) ne, and each neuron ne in one layer is connected to each neuron ne in the next layer by an edge (line) ed, forming a network structure.

[0016] The machine learning model M1 may be a deep neural network (DNN) with multiple hidden layers HL, for a total depth of four or more layers. Deep learning is a multi-layered version of the machine learning process using neural networks. A typical example is a forward propagation neural network, which sends information from front to back and makes a judgment. In its simplest form, this model requires three layers: an input layer consisting of m1 neurons, a middle layer consisting of m2 neurons determined by parameters, and an output layer consisting of m3 neurons corresponding to the number of classes to be discriminated. The neurons in the input layer, middle layer, and middle layer and output layer are connected by connection weights, and bias values ​​are added between the middle layer and output layer, making it easy to form logic gates. While three layers are sufficient for simple discrimination, adding multiple middle layers makes it possible to learn how to combine multiple features during the machine learning process.

[0017] Various known networks may be used for machine learning. For example, a convolutional neural network (CNN) may be used. This involves a process called "convolution," which compresses image features, and is strong in pattern recognition. Furthermore, a "recurrent neural network" (fully connected recurrent neural network), which can handle more complex information and allows information analysis whose meaning changes depending on the order or sequence, may be used.

[0018] To realize these technologies, conventional general-purpose arithmetic processing circuits such as CPUs and FPGAs can be used, but because much of the processing in neural networks involves matrix multiplication, GPUs and Tensor Processing Units (TPUs), which are specialized for matrix calculations, may also be used.In recent years, such dedicated artificial intelligence (AI) hardware, called "neural network processing units (NPUs)," have been designed to be integrated and embeddable with CPUs and other circuits, and may even become part of the processing circuit.

[0019] The machine learning model M1 performs machine learning using training data LDS in a learning process LP that applies a machine learning method. The training data LDS is a training image dataset that pairs a teacher image I1 with a student image I2. The machine learning model M1 inputs the student image I2 and propagates the data of the student image I2 forward from the input layer IL to the hidden layer HL to the output layer OL to generate an output image I3.

[0020] Then, the machine learning model M1 calculates the difference (loss value) between the generated output image I3 and the teacher image I1 of the learning data LDS using a loss function, and in backpropagation, adjusts parameters using an optimization algorithm to make the loss value as small as possible and performs learning.

[0021] The trained model M2 that has been trained through the learning process LP performs inference on, for example, a low-resolution input image I4 in the inference process IP, and outputs a high-resolution inferred image I5 as the output image.

[0022] The teacher image I1 of the learning data LDS is a high-resolution image, and the student image I2 of the learning data LDS is a low-resolution image created by blurring the teacher image I1.

[0023] In this embodiment, the blurring process is changed depending on whether or not the teacher image I1 contains bright spots. Specifically, if there are no bright spots in the region R1 (see FIG. 6 ) where the teacher image I1 is filtered, blurring is applied. On the other hand, if there are bright spots in the region R1 where the teacher image I1 is filtered, blurring is not applied. Examples of processes that do not blur bright spots include excluding bright spots when filtering image data, and applying blur to the entire image and then replacing the blurred portion with an unblurred image. This replacement can be performed digitally or by printing and pasting the image onto paper. This process creates a student image I2 from the high-resolution teacher image I1 without blurring the bright spots. Bluring can be achieved, for example, by filtering the image data or re-imaging the image output to a monitor or paper using a lower-resolution imaging device.

[0024] FIG. 3 is a flowchart illustrating an example of the flow of the image creation process according to the first embodiment.

[0025] First, a region R1 to be subjected to filtering is detected from the teacher image I1 (S1), and it is then determined whether or not the region R1 to be subjected to filtering contains one or more pixels that are determined to be bright spots (S2).

[0026] If it is determined that there is one or more pixels that are determined to be bright spots in the region R1 to be subjected to the filtering process (S2: YES), the filtering process is not performed (S3).

[0027] On the other hand, if it is determined that there is not one or more pixels that are determined to be bright points in the region R1 to be subjected to the filtering process (S2: NO), the filtering process is performed (S4).

[0028] Next, it is determined whether or not filtering has been performed on the entire teacher image I1 (S5). If it is determined that filtering has not been performed on the entire teacher image I1 (S5: NO), the position of the pixel of interest is shifted (S6), and the process returns to S1. On the other hand, if it is determined that filtering has been performed on the entire teacher image I1 (S5: YES), the filtering process ends.

[0029] Through the above filtering process, a predetermined amount of blur is added to pixels that are not determined to be bright spots, and a student image I2 is created in which no blur is added to pixels that are determined to be bright spots.

[0030] By performing machine learning using a machine learning model M1 with a teacher image I1 and a student image I2 in which no blurring has been added to pixels determined to be bright spots, and generating a trained model M2, it is possible to obtain an inferred image I5 in which the inference results around the bright spots are natural, even when an input image I4 containing a bright spot is input to the trained model M2.

[0031] 4 is a flowchart for explaining another example of the flow of the image creation process according to the first embodiment. In FIG. 4, the same processes as those in FIG. 3 are denoted by the same reference numerals, and the description thereof will be omitted.

[0032] In the process of S2, if it is determined that there is one or more pixels judged to be bright spots in the region R1 to be subjected to the filtering process (S2: YES), the level of the filtering process is changed and filtering process is performed (S11). For example, if the amount of blurring added to pixels judged not to be bright spots is set to 100%, the level of the filtering process is changed so that the amount of blurring added to pixels judged to be bright spots is 0 to 60%.

[0033] On the other hand, if it is determined in the process of S2 that there are not at least one pixel determined to be a bright spot in the region R1 to be subjected to the filtering process (S2: NO), the filtering level is not changed and filtering is performed (S12). That is, for pixels determined not to be bright spots, the filtering level is not changed and a 100% blur amount is added.

[0034] Next, in the process of S5, it is determined whether or not filtering has been performed on the entire teacher image I1. If it is determined that filtering has not been performed on the entire teacher image I1, the position of the pixel of interest is shifted in the process of S6, and the process returns to the process of S1. On the other hand, if it is determined that filtering has been performed on the entire teacher image I1, the filtering process ends.

[0035] Through the above filtering process, a predetermined amount of blur is added to pixels that are not determined to be bright spots, and a student image I2 can be created in which the blur is reduced for pixels that are determined to be bright spots.

[0036] As a result, by performing machine learning using a machine learning model M1 using a teacher image I1 and a student image I2 in which blurring has been reduced for pixels determined to be bright spots, and generating a trained model M2, it is possible to obtain an inferred image I5 in which the inference results around the bright spots are natural, even when an input image I4 containing a bright spot is input to the trained model M2.

[0037] Next, the configuration of an image creation device that creates student images I2 from teacher images I1 will be described. Fig. 5 is a block diagram showing an example of the configuration of an image creation device according to the first embodiment of the present invention. Note that the method of creating student images I2 from teacher images I1 is not limited to automatic creation by the image creation device shown in Fig. 5, and student images I2 may also be created manually from teacher images I1.

[0038] The image creation device 10 shown in Fig. 5 includes a bright spot detection unit 11 and an image quality processing unit 12. A high-resolution teacher image I1 is input to the image creation device 10. The teacher image I1 is supplied to the bright spot detection unit 11 and the image quality processing unit 12. The teacher image I1 may be either an image that includes bright spots or an image that does not include bright spots.

[0039] The bright spot detection unit 11 detects whether or not there is a bright spot in the teacher image I1, and outputs the detection result to the image quality processing unit 12. More specifically, the bright spot detection unit 11 detects whether or not there is a bright spot in the area of ​​the teacher image I1 where filtering processing is performed, and outputs the detection result to the image quality processing unit 12.

[0040] The image quality processing unit 12 performs a filter process to blur the teacher image I1 and outputs a low-resolution student image I2. The image quality processing unit 12 changes the filter process depending on the detection result of the bright point detection unit 11.

[0041] Specifically, if the bright spot detection unit 11 determines that there is no bright spot in the region of the teacher image I1 where filtering is to be performed, the image quality processing unit 12 performs filtering and adds blur. On the other hand, if the bright spot detection unit 11 determines that there is a bright spot in the region of the teacher image I1 where filtering is to be performed, the image quality processing unit 12 does not perform filtering and does not add blur.

[0042] The image quality processing unit 12 may change the level of filtering depending on the detection result of the bright spot detection unit 11. Specifically, if the bright spot detection unit 11 determines that there is no bright spot in the area of ​​the teacher image I1 where filtering is performed, the image quality processing unit 12 adds blur without changing the level of filtering. On the other hand, if the bright spot detection unit 11 determines that there is a bright spot in the area of ​​the teacher image I1 where filtering is performed, the image quality processing unit 12 changes the level of filtering to reduce the amount of blur added. For example, if the amount of blur outside a predetermined range from the bright spot is set to 100%, the level of filtering is changed so that the amount of blur within a predetermined range from the bright spot is 0 to 60%.

[0043] As a result, when the teacher image I1 contains a bright spot, a student image I2 can be obtained in which the image quality within a predetermined range from the bright spot is not degraded compared to the area outside the predetermined range from the bright spot (no blurring is added, or the amount of blurring is reduced). Also, when the teacher image I1 does not contain a bright spot, a student image I2 in which blurring is added to the entire teacher image I1 can be obtained.

[0044] A pair of teacher image I1 and student image I2 with blurring added to reduce image quality is input as learning data LDS to model learning processing unit 21. Here, the number of student images I2 that do not contain bright spots is input to model learning processing unit 21 so that the number of student images I2 that do contain bright spots is greater than the number of student images I2 that do contain bright spots.

[0045] The model learning processing unit 21 includes a machine learning model M1. The machine learning model M1 performs machine learning using the learning data LDS to generate a trained model M2.

[0046] 6 is a diagram illustrating the configuration of the bright spot detection unit and image quality processing unit of the first embodiment. The image quality processing unit 12 includes a calculation unit 13 and an image filter 14. A teacher image I1 is input to the bright spot detection unit 11. The bright spot detection unit 11 sets a pixel of interest P1 and a region R1 for performing filtering on the teacher image I1.

[0047] The pixel of interest P1 is a pixel that undergoes filtering to add blur. The region R1 where the filtering is performed has the same filter size as the image filter 14 included in the image quality processing unit 12, and in the example of Figure 6, it is a 3 x 3 pixel region centered on the pixel of interest P1.

[0048] The bright spot detection unit 11 determines whether there is a pixel in the region R1 where filtering is performed that is determined to be a bright spot. That is, the bright spot detection unit 11 scans the pixel values ​​of the pixels in the region R1 where filtering is performed, for example, from the upper left to the lower right, to determine whether there is a pixel that is determined to be a bright spot. Specifically, the bright spot detection unit 11 determines a pixel in the region R1 where filtering is performed to be a bright spot if the pixel value of that pixel is equal to or greater than a certain value. More specifically, a pixel having a pixel value that is equal to or greater than a predetermined percentage of the maximum possible pixel value in the teacher image I1 is determined to be a bright spot. The maximum possible pixel value can be calculated from the number of bits of the image data of the teacher image I1. For example, an 8-bit image can have 256 levels of information, which is 2 to the power of 8. Therefore, the minimum brightness value is 0 and the maximum brightness value is 255. For a 10-bit image, the maximum brightness value is 1023, and for a 12-bit image, the maximum brightness value is 4095. The predetermined percentage can be, for example, 80%.

[0049] The condition for determining a bright spot is not limited to when the pixel value of any pixel in the region R1 on which filtering is performed is equal to or greater than a certain value. For example, the bright spot detection unit 11 may determine a bright spot to be a bright spot when the average, median, minimum, or maximum value of the pixel values ​​around any pixel in the region R1 on which filtering is performed is equal to or greater than a certain value.

[0050] The calculation unit 13 multiplies each coefficient of the image filter 14 by the pixel value of each pixel in the region R1 and adds up the results of each multiplication to perform filtering on the pixel of interest P1, i.e., processing to add blur. The image quality processing unit 12 performs the same processing on the entire teacher image I1 while shifting the position of the pixel of interest, thereby completing filtering processing to add blur to the teacher image I1 and outputting the student image I2.

[0051] With this configuration of the image creation device 10, when the image quality of a teacher image I1 containing a bright spot is reduced to create a student image I2 that is paired with the teacher image I1, it is possible to create a student image I2 that does not have lower image quality within a specified range from the bright spot compared to outside the specified range from the bright spot.

[0052] Second Embodiment Next, a second embodiment will be described. Fig. 7 is a diagram for explaining the configuration of a bright spot detection unit and an image quality processing unit of the second embodiment. In Fig. 7, the same components as those in Fig. 6 are assigned the same reference numerals and their description will be omitted.

[0053] An image creation device 10A of the second embodiment includes an image quality processing unit 12A instead of the image quality processing unit 12 in Fig. 3. The image quality processing unit 12A includes a calculation unit 13 and two image filters 14A and 14B with different filter coefficients. Note that, although the image quality processing unit 12A includes two pixel filters 14A and 14B, the present invention is not limited to this and may include three or more pixel filters.

[0054] When the bright spot detection unit 11 detects a plurality of pixels whose pixel values ​​are equal to or greater than a certain value, the image quality processing unit 12A changes the image filters 14A and 14B to be applied in accordance with the detected pixel values.

[0055] For example, if a bright spot is detected in region R2 where pixel of interest P2 and filter processing are performed, and region R3 where pixel of interest P3 and filter processing are performed, image filter 14A is applied to pixel of interest P2, and image filter 14B is applied to pixel of interest P3.

[0056] It should be noted that the image quality processing unit 12A does not need to have multiple image filters 14A and 14B, and may, for example, have one image filter and change the filter coefficient of the image filter according to the pixel value detected by the bright spot detection unit 11.

[0057] In this way, in this embodiment, when there are multiple pixels determined to be bright spots, the amount of blur added can be adjusted more finely than in the first embodiment by applying image filter 14A or 14B with different filter coefficients.

[0058] (Modification) Fig. 8 is a diagram illustrating an example of a region in which bright spots are detected. In the above-described embodiment, the region in which bright spots are detected is the same as the region R1 in which filtering is performed, but as shown in Fig. 8, it may also be a region R4 in which a margin is added to the region R1 in which filtering is performed. In other words, the region R4 in which bright spots are detected has a square shape of 4 x 4 pixels with the target pixel P1 at its center.

[0059] 8, the margin width is set to 1 (1 pixel), but this is not limited to this. The margin width may be set to any value as long as it is 1 or greater and the pixel size is such that the filter size plus the margin width is within the image size minus 1.

[0060] The shape of the margined region R4 is not limited to a square. The shape of the margined region R4 may be, for example, a rectangle, a parallelogram (diamond), a trapezoid, a circle, an ellipse, a sector (semicircle), a cross, a polygon, a star, or the like. Furthermore, these shapes of the margined region R4 may be rotated. Generally, bright spots are often circular or elliptical, so it is preferable to set the shape of the margined region R4 to a circle or an ellipse. Setting the shape of the margined region R4 to a circle or an ellipse allows for just the right amount of blurring to be added around the bright spot.

[0061] Furthermore, the shape of the margined region R4 and / or the value of the margin width may be changed depending on the type of endoscope or observation light used to acquire the teacher image I1.

[0062] Third Embodiment Next, a third embodiment will be described. Fig. 9 is a diagram for explaining the configuration of a bright spot detection unit and an image quality processing unit of the third embodiment. In Fig. 9, the same components as those in Fig. 6 are assigned the same reference numerals and will not be described again.

[0063] An image creation device 10B of the third embodiment includes an image quality processing unit 12B instead of the image quality processing unit 12 in Fig. 3. The image quality processing unit 12B includes a calculation unit 13, an image filter 14C for an R image, an image filter 14D for a G image, and an image filter 14E for a B image.

[0064] The filter size of the image filter 14C for R images is 4 x 4 pixels, the filter size of the image filter 14D for G images is 3 x 3 pixels, and the filter size of the image filter 14E for B images is 5 x 5 pixels. In this way, the image filter 14C for R images, the image filter 14D for G images, and the image filter 14E for B images each have a different filter size.

[0065] In this way, when the filter sizes for the R, G, and B images are different, the largest filter size is set as the range in which bright spots are detected by the bright spot detection unit 11. That is, in this embodiment, the bright spot detection unit 11 determines whether or not there are any pixels whose pixel values ​​are equal to or greater than a certain value within a 5 x 5 pixel range centered on the target pixel P1. Note that the range in which bright spots are detected by the bright spot detection unit 11 may be the filter size for the R or G image.

[0066] In the above-described embodiments, the teacher image I1 input to the image creation device 10 is input as the teacher image, and the student image I2 output from the image creation device 10 is input as the student image to the model learning processing unit 21, but this is not limited to this. For example, a plurality of student images I2 output from the image creation device 10 to which no blur has been added (or to which the amount of blur added is reduced) may be input to the model learning processing unit 21 as the teacher image and the student image.

[0067] Alternatively, after blurring the entire teacher image I1, the bright spots or the bright spots and their surrounding areas can be synthesized from the image before blurring. That is, first, a low-quality image of the entire teacher image I1 is created by blurring the entire teacher image I1, and then the image of the bright spots or the bright spots and their surrounding areas is manually replaced with the original image without blurring. This allows the generation of a student image I2 in which only the bright spots or the bright spots and their surrounding areas are not blurred.

[0068] It should be noted that the steps in the flowcharts in this specification may be executed in a different order, may be executed multiple times simultaneously, or may be executed in a different order each time, as long as this does not contradict the nature of the steps.

[0069] The present invention is not limited to the above-described embodiments, and it goes without saying that various modifications, combinations, and applications are possible within the scope of the invention without departing from the spirit of the invention.

[0070] Furthermore, among the technologies described herein, many of the controls and functions, mainly those described in the flowcharts, can be set by a program, and the above-described controls and functions can be realized by a computer reading and executing the program. The program can be recorded or stored, in whole or in part, as a computer program product on portable media such as nonvolatile memory, such as a flexible disk or CD-ROM, or on storage media such as a hard disk or volatile memory, and can be distributed or provided at the time of product shipment, via portable media, or via a communication line. A user can easily realize the machine learning image creation program of this embodiment by downloading the program via a communication network and installing it on a computer, or by installing it on a computer from a recording medium.

Claims

1. A method for creating images for machine learning, in which when creating images for machine learning by reducing the image quality of a teacher image that includes a bright spot to create a student image that is paired with the teacher image, the image quality within a specified range from the bright spot is not reduced compared to the image quality outside the specified range from the bright spot.

2. A method for creating images for machine learning as described in claim 1, wherein the image quality of the entire teacher image including the bright spot is reduced to generate an overall reduced image, and then a predetermined range from the bright spot in the overall reduced image is combined with the same portion of the teacher image including the bright spot.

3. A method for creating images for machine learning as described in claim 1, wherein a teacher image including the bright spot is processed using an image filter so that, when the amount of blur outside a predetermined range from the bright spot is set to 100%, the amount of blur within a predetermined range from the bright spot is set to less than a predetermined percentage.

4. The method for creating images for machine learning described in claim 3, wherein the predetermined ratio is 0 to 60%.

5. The method for creating an image for machine learning according to claim 1, wherein pixels whose pixel values ​​are equal to or greater than a predetermined value are defined as bright points.

6. A method for creating an image for machine learning as described in claim 1, wherein the amount of blurring when blurring a predetermined range from the bright point is determined according to the pixel value of the bright point.

7. A machine learning method for performing machine learning using a teacher image containing bright points as described in claim 1 and a student image containing bright points created by the method for creating images for machine learning as described in claim 1.

8. The machine learning method according to claim 7, further comprising: performing machine learning using a teacher image that does not contain bright spots; and a student image that does not contain bright spots and is created by reducing the image quality of the teacher image that does not contain bright spots.

9. The machine learning method according to claim 7, wherein the machine learning is performed so that the number of student images that do not contain the bright spots is greater than the number of student images that do contain the bright spots.

10. An image creation device for machine learning, comprising: a bright spot detection unit that detects bright spots from a teacher image; and an image quality processing unit that, when lowering the image quality of the teacher image to create a student image that pairs with the teacher image, does not lower the image quality within a specified range from the bright spot compared to outside the specified range.

11. The image creation device for machine learning described in claim 10, wherein the image quality processing unit includes an image filter, and the image quality is reduced by processing with the image filter so that the amount of blur within a predetermined range from the bright point is less than a predetermined ratio compared to outside the predetermined range from the bright point.

12. An image creation device for machine learning as described in claim 11, wherein the amount of blurring when blurring a predetermined range from the bright point is determined according to the pixel value of the bright point.

13. The image creation device for machine learning described in claim 11, wherein the area in which the bright spot detection unit detects bright spots from the teacher image is the same as the filter size of the image filter.

14. An image creation device for machine learning as described in claim 11, wherein the area in which the bright spot detection unit detects bright spots from the teacher image is an area with a margin added to the filter size of the image filter.

15. The image creation device for machine learning described in claim 14, wherein the width of the margin is 1 or more, and the filter size and the width of the margin are a pixel size that fits within the image size - 1.

16. The image creation device for machine learning described in claim 14, wherein the shape of the area with a margin added to the filter size of the image filter is a square, rectangle, parallelogram, trapezoid, circle, ellipse, sector, cross, polygon, or star.

17. The image creation device for machine learning described in claim 10, wherein the image quality processing unit includes a plurality of image filters with different filter coefficients, and applies one of the plurality of image filters depending on the pixel value of the bright spot.

18. The image creation device for machine learning described in claim 10, wherein the image quality processing unit includes filters for R images, G images, and B images, and the area in which the bright spots are detected from the teacher image in the bright spot detection unit is the same as the filter size of the filter with the largest filter size among the filters for the R images, the G images, and the B images.

19. An image creation device for machine learning that includes a processor that detects bright spots from a teacher image, and when reducing the image quality of the teacher image to create a student image that is a counterpart to the teacher image, does not reduce the image quality within a specified range from the bright spot compared to outside the specified range from the bright spot.

20. An image creation program for machine learning that causes a computer to perform the following processes: a process of detecting bright spots from a teacher image; and a process of not lowering the image quality of a specified range from the bright spot compared to an area outside the specified range from the bright spot when lowering the image quality of the teacher image to create a student image that is a counterpart to the teacher image.

Images