JP6872502B2 - Image processing equipment, image processing methods, and programs - Google Patents

Description

The present invention relates to an image processing apparatus, an image processing method, and a program, and more particularly to an image processing apparatus, an image processing method, and a program that perform segmentation using machine learning.

Conventionally, segmentation has been performed in the technical field of image processing. Here, the segmentation is a process of distinguishing between the area composed of the pixels of the subject image to be segmented and the area of other pixels.

For example, in Patent Document 1, a technique for generating a concrete region divided image divided into an aggregate region representing an aggregate and a residual component region containing a cement component in an input concrete image is provided. Are listed. Patent Document 1 describes that, based on a shade monochrome image and a contour monochrome image generated from a concrete image, an image in which a region is divided into an aggregate region and a residual component region in the concrete image is generated. .. In the technique described in Patent Document 1, when the aggregate region and the residual component region cannot be accurately distinguished by the color density, the aggregate is accurately produced by a composite monochrome image obtained by synthesizing a grayscale monochrome image and a contour monochrome image. We are trying to divide the region and the residual component region. Here, the shading monochrome image is composed of a dark region and a light region, and the contour monochrome image is composed of a contour region and a background region.

On the other hand, in recent years, it has been proposed to perform segmentation using machine learning technology. For example, Non-Patent Document 1 aims to perform segmentation more accurately by using a convolutional neural network (CNN) that calculates features from images by learning and performs image recognition processing. Technology has been proposed.

JP-A-2010-230421 Adam Paszke, Abhishek Chaurasia, Sangpil Kim, Eugenio Culurciello “ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation” [online], Submitted on June 7, 2016, [Search on January 18, 2018], Internet <URL: https://arxiv.org/abs/1606.02147>

Here, even in segmentation using machine learning, it may be difficult to accurately detect boundaries between regions. This is because there are various types and appearances of the subject image to be processed by the segmentation, so that it may be difficult for the segmentation device subjected to machine learning to detect the boundary with high accuracy.

The present invention has been made in view of such circumstances, and an object of the present invention is to provide an image processing device, an image processing method, and a program capable of accurately detecting boundaries between regions in segmentation using machine learning. To provide.

In order to achieve the above object, the image processing apparatus according to one aspect of the present invention has an image acquisition unit that acquires a processing target image having a segmentation target subject image, and features of the subject image learned by the first machine learning. Corresponds to the subject image by the image feature detector that generates the emphasized image emphasized by the aspect learned by the first machine learning and the aspect learned by the second machine learning based on the emphasized image and the image to be processed. It is provided with a segmentation device for segmenting the area to be processed.

According to this aspect, the image feature detector generates an enhanced image in which the features of the subject image learned by the first machine learning are emphasized by the aspect learned by the first machine learning. Then, in this aspect, the region corresponding to the subject image is segmented by the segmentation device according to the aspect learned by the second machine learning based on the emphasized image and the image to be processed. That is, using the second machine learning, the region corresponding to the subject image is segmented based on the emphasized image and the image to be processed. Thereby, in this aspect, the boundary between regions can be detected with high accuracy.

Preferably, the segmentation device generates a segmentation image based on the emphasized image and the R image, the G image, and the B image of the image to be processed.

According to this aspect, the segmentation device generates a segmentation image based on the emphasized image and the R image, the G image, and the B image of the image to be processed. That is, according to this aspect, segmentation is performed by a 4-channel image in which a 1-channel emphasized image and a 3-channel processed image are combined. Thereby, in this aspect, it is possible to obtain a segmentation image in which the boundary between regions is detected with high accuracy.

Preferably, the segmentation device generates a segmentation image based on the emphasized image and the monochrome image of the image to be processed.

According to this aspect, the segmentation device generates a segmentation image based on the emphasized image and the monochrome image of the image to be processed. Thereby, in this aspect, it is possible to obtain a segmentation image in which the boundary between the regions is accurately detected.

Preferably, the image feature detector is an edge detector that has learned to detect the contour of the subject image by the first machine learning, generates an edge image in which the contour is detected as a emphasized image, and the segmentation device is a segmentation device. Learning is performed by the second machine learning, and a region corresponding to the subject image is extracted based on the edge image and the image to be processed.

According to this aspect, the image feature detector is an edge detector that has learned to detect the contour of a subject image by the first machine learning, and the edge image whose contour is detected as an emphasized image by the edge detector. Is generated. Then, according to this aspect, the segmentation device is learned by the second machine learning, and the region corresponding to the subject image is extracted based on the edge image and the image to be processed. Thereby, in this aspect, the segmentation that accurately extracts the boundary between the regions can be performed.

Preferably, the segmentation device produces a segmentation image that emphasizes the area corresponding to the segmented subject image.

According to this aspect, since the segmentation device generates a segmentation image in which the region corresponding to the segmented subject image is emphasized, it is possible to obtain a segmentation image in which the boundary of the region corresponding to the subject image is detected with high accuracy. it can.

Preferably, the image acquisition unit acquires a processing target image having a crack image as the subject image to be segmented, and the image feature detector is a crack detector that has learned to detect the crack image by the first machine learning. Yes, a crack image in which a crack image is detected is generated as an emphasized image, and the segmentation device learns by a second machine learning and extracts a region corresponding to the crack image based on the crack image and the image to be processed.

According to this aspect, the image acquisition unit acquires a processing target image having a crack image as a subject image to be segmented, and the image feature detector learns to detect the crack image by the first machine learning. It is a vessel, and a crack image in which a crack image is detected is generated as a emphasized image. Then, in this aspect, a region corresponding to the crack image is extracted based on the crack image and the image to be processed by learning by the second machine learning by the segmentation device. Thereby, in this aspect, it is possible to perform segmentation in which the boundary of the region corresponding to the extracted crack image is detected with high accuracy.

Preferably, the segmentation device segmentes the inspection area containing the cracked image.

According to this aspect, since the inspection area including the crack image is segmented by the segmentation device, the boundary of the inspection area including the crack image can be detected with high accuracy.

Preferably, the segmentation device segmentes the thick line inspection area containing the cracked image.

According to this aspect, since the inspection area of the thick line including the crack image is segmented by the segmentation device, the boundary of the inspection area of the thick line including the crack image can be detected with high accuracy.

Preferably, the segmentation device segments the rectangular inspection area containing the cracked image.

According to this aspect, since the rectangular inspection area including the crack image is segmented by the segmentation device, the boundary of the rectangular inspection area including the crack image can be detected with high accuracy.

Preferably, the segmentation device produces a segmentation image that emphasizes the area corresponding to the segmented crack image.

According to this aspect, since the segmentation device generates a segmentation image in which the region corresponding to the segmented crack image is emphasized, it is possible to obtain a segmentation image in which the boundary of the region is detected with high accuracy.

Preferably, the segmentation device produces a segmentation image that emphasizes the segmented examination area.

According to this aspect, the segmentation device produces a segmentation image in which the segmented examination area is emphasized. As a result, in this aspect, a segmentation image in which the boundary of the inspection area is detected with high accuracy is generated.

In the image processing method according to one aspect of the present invention, a first machine learning is performed on an image acquisition step of acquiring a processing target image having a subject image to be segmented, and a feature of the subject image learned by the first machine learning. An image feature detection step that generates an emphasized image emphasized by the aspect learned by the above, and a segmentation step that segmentes a region corresponding to the subject image by the aspect learned by the second machine learning based on the emphasized image and the image to be processed. And, including.

The program according to one aspect of the present invention learns the features of the subject image learned by the first machine learning and the image acquisition step of acquiring the image to be processed having the subject image to be segmented by the first machine learning. An image feature detection step for generating an emphasized image emphasized by the above-mentioned aspect, and a segmentation step for segmenting a region corresponding to a subject image by an aspect learned by a second machine learning based on the emphasized image and the image to be processed. Have the computer perform the image processing process including.

According to the present invention, the image feature detector generates an enhanced image in which the features of the subject image learned by the first machine learning are emphasized by the mode learned by the first machine learning, and the segmentation device generates the enhanced image. Since the region corresponding to the subject image is segmented according to the mode learned by the second machine learning based on the processed image, the boundary between the regions can be detected with high accuracy.

FIG. 1 is a diagram showing the appearance of a computer provided with an image processing device. FIG. 2 is a block diagram showing a functional configuration example of the image processing device. FIG. 3 is a schematic view showing a typical configuration example of CNN. FIG. 4 is a diagram showing an example of segmentation. FIG. 5 is a diagram showing an example of segmentation. FIG. 6 is a flow chart for explaining the process of the image processing method. FIG. 7 is a diagram showing an input image. FIG. 8 is a diagram showing an edge image. FIG. 9 is a diagram showing an object area image. FIG. 10 is a diagram showing an example of segmentation. FIG. 11 is a diagram showing an example of segmentation.

Hereinafter, preferred embodiments of the image processing apparatus, image processing method, and program according to the present invention will be described with reference to the accompanying drawings.

FIG. 1 is a diagram showing the appearance of a computer provided with the image processing apparatus of the present invention.

The computer 3 includes an image processing device 11 (FIG. 2), which is one aspect of the present invention. An input image 23 is input to the computer 3, and a display unit composed of a monitor 9 and an input unit composed of a keyboard 5 and a mouse 7 are connected to the computer 3. The illustrated form of the computer 3 is an example, and the device having the same function as the computer 3 can include the image processing device 11 of the present invention. For example, the image processing device 11 can be mounted on the tablet terminal.

The computer 3 displays, for example, a segmentation image on the monitor 9 as a result of the segmentation performed on the input image 23. In addition, the keyboard 5 and the mouse 7 receive the segmentation command from the user.

FIG. 2 is a block diagram showing a functional configuration example of the image processing device 11. The image processing device 11 mainly includes an image acquisition unit 13, an image feature detector 15, a segmentation device 17, a display control unit 19, and a storage unit 21.

The image acquisition unit 13 acquires an input image 23 having a subject image to be segmented as a processing target image. The image acquisition unit 13 acquires the input image 23 by using an image input / output interface (not shown) provided in the computer 3 by wire or wirelessly. The input image 23 may be a single color image, a binarized R image, a G image, a B image, or a binarized black and white image.

The image feature detector 15 generates an enhanced image in which the features of the subject image learned by the first machine learning are emphasized by the mode learned by the first machine learning. For example, the image feature detector 15 is an edge detector that has learned to detect the contour of a subject image by first machine learning. In this case, the image feature detector 15 generates an edge image in which the outline of the subject image is detected as an enhanced image. When there are a plurality of subject images in the image, the image feature detector 15 performs edge detection on the main subject image and generates an edge image. Further, for example, the image feature detector 15 is a crack detector that has learned to detect a crack image by the first machine learning. In this case, the image feature detector 15 generates a crack image in which a crack image is detected as an enhanced image. A known technique is applied to the first machine learning of the image feature detector 15.

The image feature detector 15 is a portion that emphasizes the features of the subject image of the input image 23. In this example, the convolutional neural network (CNN:) that calculates the feature amount from the image by learning and performs the enhancement processing of the features of the image. Convolutional Neural Network), and the feature amount is calculated and emphasized by the color information in the image, the gradient of the pixel value, etc. Here, the image feature enhancement process is, for example, an edge detection process or a crack detection process.

FIG. 3 is a schematic diagram showing a typical configuration example of a CNN applied to the image feature detector 15.

As shown in FIG. 3, the CNN includes an input layer 15A, a plurality of sets composed of a convolution layer and a pooling layer, an intermediate layer 15B having a fully connected layer, and an output layer 15C, and each layer has a plurality of "". The structure is such that "nodes" are connected by "edges".

An image to be recognized is input to the input layer 15A.

The intermediate layer 15B has a plurality of sets including a convolution layer and a pooling layer as one set, and a fully connected layer, and features are extracted from an image input from the input layer. The convolution layer filters nearby nodes in the previous layer (performs a convolution operation using the filter) and acquires a "feature map". The pooling layer reduces the feature map output from the convolution layer to a new feature map. The "convolution layer" plays a role of feature extraction such as edge extraction from an image, and the "pooling layer" plays a role of imparting robustness so that the extracted features are not affected by translation or the like.

The intermediate layer 15B is not limited to the case where the convolution layer and the pooling layer are set as one set, but also includes the case where the convolution layers are continuous and the normalization layer. Further, the parameters of the filter used in each convolution layer are automatically learned in advance from a large amount of learning data (in this example, the correct edge image of the image and the subject image to be segmented in the image).

The output layer 15C outputs an image in which the characteristics of the subject image are emphasized based on the characteristics extracted by the intermediate layer 15B.

The segmentation device 17 segmentes the region corresponding to the subject image based on the emphasized image and the image to be processed according to the mode learned by the second machine learning. For example, the segmentation device 17 generates a segmentation image based on the emphasized image and the R image, the G image, and the B image of the image to be processed. Further, for example, the segmentation device 17 generates a segmentation image based on the emphasized image and the monochrome image of the image to be processed. Here, the monochrome image is a monochromatic binarized image. Further, the segmentation is a process of distinguishing between an area composed of pixels of a subject image to be segmented and an area composed of other pixels. Further, the segmentation device 17 generates a segmentation image based on the information obtained by performing the segmentation. For example, the segmentation device 17 generates a segmentation image in which the region corresponding to the segmented subject image is emphasized, a segmentation image in which the region corresponding to the segmented crack image is emphasized, or a segmentation image in which the segmented inspection region is emphasized. Here, emphasizing an area means, for example, expressing a segmented area and another area with different colors.

The segmentation device 17 is subjected to the second machine learning in the same manner as the image feature detector 15. In the segmentation device 17, FCN (Fully Convolution Network) may be applied. Further, the second machine learning to the segmentation device 17 is learned in advance by a large number of learning data (in this example, the correct segmentation result of the image and the correct edge image and the subject image of the segmentation image in the image).

The display control unit 19 controls the display on the monitor 9. For example, the display control unit 19 controls the display of the segmentation image generated by the segmentation device 17 on the monitor 9.

The storage unit 21 stores the input image to be processed, the emphasized image generated by the image feature detector 15, the information about the segmentation performed by the segmentation device 17, and the segmentation image generated by the segmentation device 17. Further, the storage unit 21 stores information and the like related to various controls of the image processing device 11.

<First Embodiment>
4 and 5 are diagrams showing an example of segmentation in the image processing apparatus 11 of the first embodiment. In the first embodiment, the image feature detector 15 functions as an edge detector, and the segmentation device 17 functions as an object detector.

FIG. 4 conceptually shows the flow of processing in the image processing apparatus 11, and FIG. 5 shows an input image (processed image), an edge image 27 as an emphasized image, and an object area image 35 as a segmentation image. Has been done.

As shown in FIG. 4, the image acquisition unit 13 acquires the input image 23 (R image 23A, G image 23B, and B image 23C of the input image 23). Here, the R image 23A of the input image 23 is a red binarized image, the G image 23B is a green binarized image, and the B image 23C is a blue binarized image. To do. Then, the input image 23 composed of the three channels of the R image 23A, the G image 23B, and the B image 23C is input to the edge detector constructed by the edge detection Net (neural network) 25. Then, the edge detection Net 25 generates an edge image 27. That is, in this case, the image feature detector 15 is an edge detection Net 25 (edge detector) that has learned to detect the contour of the subject image by the first machine learning, and the edge image 27 in which the contour is detected as the emphasized image. Is generated.

FIG. 5 shows an input image 23 (R image 23A, G image 23B, and B image 23C) input to the edge detection Net 25, and an edge image 27 generated by the edge detection Net 25. The input image 23 has a person 101 as a main subject image, a house 103 as a background, and a tree 105. Since the edge detection Net 25 performs the first machine learning so as to detect the edge of the contour of the subject image (emphasize the region of the main subject image), the edge 107 of the contour of the person 101 is detected in the edge image 27. Has been done.

Returning to FIG. 4, the edge image 27 generated by the edge detection Net 25 and the input input images 23 (R image 23A, G image 23B, and B image 23C) are input to the segmentation Net 29. The segmentation device 17 is composed of a segmentation Net 29. The segmentation Net29 learns by the second machine learning and extracts the region corresponding to the subject image based on the edge image 27 and the input image 23. Here, extracting the area corresponding to the subject image means classifying each pixel constituting the image into an extraction area and a background area. The segmentation Net29 outputs the result of extracting the contour region as the result of the segmentation 31. Since the segmentation Net29 receives the edge image 27 and the input image 23 of the three channels (R image 23A, G image 23B, and B image 23C), the boundary of the region can be detected with high accuracy.

The segmentation Net29 generates a segmentation image in which the region corresponding to the segmented subject image is emphasized. FIG. 5 shows an object region image 35 (segmentation image) which is one aspect of the segmentation result 31 output from the segmentation Net 29. The object area image 35 is an image that reflects the segmentation result of the segmentation Net29. Specifically, in the object region image 35, the person 101, which is the main subject image, is segmented and shown in the region 109. As another aspect of the segmentation result 31, information on the segmentation can be mentioned. Specifically, the information regarding segmentation is information on which region each pixel belongs to.

FIG. 6 is a flow chart illustrating a process of the image processing method of the present invention.

First, the image acquisition unit 13 acquires the R image 23A, the G image 23B, and the B image 23C of the input image 23 (step S10, step S11, and step S12: image acquisition step). After that, the R image 23A, the G image 23B, and the B image 23C are input to the edge detection Net 25, and edge detection is performed (step S13). After that, the image feature detector 15 outputs the edge image 27 (step S14: image feature detection step).

After that, the R image 23A, the G image 23B, the B image 23C, and the edge image 27 are input to the segmentation Net29, and segmentation is performed (step S15: segmentation step). Then, the segmentation Net29 outputs a segmentation image (step S16).

In the above embodiment, the hardware structure of the processing unit that executes various processes is various processors as shown below. For various processors, the circuit configuration can be changed after manufacturing the CPU (Central Processing Unit), FPGA (Field Programmable Gate Array), etc., which are general-purpose processors that execute software (programs) and function as various processing units. Programmable Logic Device (PLD), a dedicated electric circuit, which is a processor having a circuit configuration specially designed to execute a specific process such as an ASIC (Application Specific Integrated Circuit). Is done.

One processing unit may be composed of one of these various processors, or may be composed of two or more processors of the same type or different types (for example, a plurality of FPGAs or a combination of a CPU and an FPGA). You may. Further, a plurality of processing units may be configured by one processor. As an example of configuring a plurality of processing units with one processor, first, one processor is configured by a combination of one or more CPUs and software, as represented by a computer such as a client or a server. There is a form in which the processor functions as a plurality of processing units. Secondly, as typified by System On Chip (SoC), there is a form in which a processor that realizes the functions of the entire system including a plurality of processing units with one IC (Integrated Circuit) chip is used. is there. As described above, the various processing units are configured by using one or more of the above-mentioned various processors as a hardware-like structure.

Further, the hardware structure of these various processors is, more specifically, an electric circuit (circuitry) in which circuit elements such as semiconductor elements are combined.

Each of the above configurations and functions can be appropriately realized by any hardware, software, or a combination of both. For example, for a program that causes a computer to perform the above-mentioned processing steps, a computer-readable recording medium (non-temporary recording medium) that records such a program, or a computer on which such a program can be installed. However, it is possible to apply the present invention.

Next, the edge image and the object area image will be described in detail.

FIG. 7 is a diagram showing an input image 201, FIG. 8 is a diagram showing an edge image 221 obtained from the input image 201 shown in FIG. 7, and FIG. 9 is an object region image obtained from the input image 201. It is a figure which shows 231.

The input image 201 shown in FIG. 7 has a person 203, which is the main subject image to be segmented. Further, the input image 201 has images of a hat 205, a shadow 207, a mountain 209, and a sea 211 in addition to the person 203.

The edge image 221 shown in FIG. 8 is generated by the edge detection Net 25 when the input image 201 is input to the edge detection Net 25. The edge image 221 has a human contour 223 on which the contour of the person 203 in the input image 201 is detected. The edge detection Net 25 is subjected to the first machine learning to detect the human contour 223.

In the portion 225 of the edge image 221, a portion other than the human contour 223 is also detected as a contour. Further, the contour of the shadow 207 is detected at the position 229 of the edge image 221. Further, the contour of the hat 205 is detected at the portion 227 of the edge image 221 and there is a portion that is not well detected in the contour 223 of the person.

In the object area image 231 shown in FIG. 9, the input image 201 and the edge image 221 are input to the segmentation Net29 and generated by the segmentation Net29. The segmentation Net29 can accurately detect the boundary of the human region 233 by inputting the input image 201 and the edge image 221.

For example, at location 225 of the object region image 231, the boundary of the human region 233 can be accurately detected. Further, although the contour of the shadow 207 has been detected in the edge image 221, the boundary of the human region 233 can be accurately detected at the location 225 of the object area image 231. Further, in the edge image 221 the contour of the hat 205 was detected or the contour 223 of the person was not detected well, but in the portion 227 and the portion 229 of the object area image 231, the boundary of the human region 233 is accurate. Well detected.

<Second embodiment>
Next, the second embodiment will be described. In the present embodiment, the image feature detector 15 functions as a crack detector, and the segmentation device 17 functions as a crack region detector.

10 and 11 are diagrams showing an example of segmentation in the image processing apparatus 11 of the second embodiment. FIG. 10 conceptually shows the processing flow in the image processing device 11, and in FIG. 11, the input image (processed image) 37, the cracked image 41 as the emphasized image, and the cracked region image 43 as the segmentation image are shown. It is shown.

As shown in FIG. 10, the image acquisition unit 13 acquires the input image 37 (R image 37A, G image 37B, and B image 37C of the input image 37). Then, the input image 37 composed of the three channels of the R image 37A, the G image 37B, and the B image 37C is input to the crack detector constructed by the crack detection Net (neural network) 39. Then, the crack detection Net39 generates a crack image 41.

FIG. 11 shows the input image 23 and the cracked image 41. The input image 23 has a crack image 111 as a subject image to be segmented. The crack detection Net39 learns by the first machine learning, detects the crack image 111, and generates the crack image 41. In the crack image 41, the crack image 113 detected by the crack detection Net 39 is shown.

Returning to FIG. 10, the crack image 41 generated by the crack detection Net 39 and the input input images 37 (R image 37A, G image 37B, and B image 37C) are input to the segmentation Net 43. Then, the segmentation Net43 learns by the second machine learning and extracts the cracked region based on the cracked image 41 image and the input image 37. Here, the cracked region is not necessarily limited to the region corresponding to the cracked image 113. For example, the segmentation Net43 may segment a predetermined range of regions including the crack image 113. Further, the segmentation Net43 may segment the inspection region of the thick line including the crack image 113. Further, the segmentation Net43 may segment a rectangular inspection region including the crack image 113. Further, the segmentation Net 43 may generate a crack region image 43 in which the region of the segmented crack image 113 is emphasized. In addition, the segmentation Net 43 may generate a cracked region image 43 that emphasizes the segmented inspection region.

The segmentation Net 43 produces a cracked region image 43 that emphasizes the cracked region as a result of segmentation 45. FIG. 11 shows a cracked region image 43. The crack region image 43 is segmented according to the detected crack image 113 and is shown in the inspection region 115. In the cracked region image 43 output by the segmentation Net 43, the cracked image 41 and the input image 37 of the 3 channels (R image 37A, G image 37B, and B image 37C) are input, so that the boundary of the inspection region 115 is accurately defined. It has been detected.

Although the examples of the present invention have been described above, it goes without saying that the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present invention.

3: Computer 5: Keyboard 7: Mouse 9: Monitor 11: Image processing device 13: Image acquisition unit 15: Image feature detector 17: Segmentation device 19: Display control unit 21: Storage unit 23: Input image 27: Edge image 35 : Object area image 37: Input image 41: Crack image 43: Crack area image Step S10-S16: Image processing step

Claims (13)

An image acquisition unit that acquires an image to be processed that has a subject image to be segmented,
An image feature detector that generates an enhanced image in which the features of the subject image learned by the first machine learning are emphasized by the mode learned by the first machine learning.
A segmentation device that segments a region corresponding to the subject image according to the mode learned by the second machine learning based on the emphasized image and the image to be processed.
An image processing device comprising. The image processing apparatus according to claim 1, wherein the segmentation device generates a segmentation image based on the emphasized image and the R image, the G image, and the B image of the processing target image. The image processing apparatus according to claim 1, wherein the segmentation device generates a segmentation image based on the emphasized image and a monochrome image of the processing target image. The image feature detector is an edge detector that has learned to detect the contour of the subject image by the first machine learning, and generates an edge image in which the contour is detected as the emphasized image.
The segmentation device according to any one of claims 1 to 3, wherein the segmentation device learns by the second machine learning and extracts a region corresponding to the subject image based on the edge image and the processing target image. Image processing device. The image processing apparatus according to claim 4, wherein the segmentation device generates a segmentation image in which a region corresponding to the segmented subject image is emphasized. The image acquisition unit acquires the image to be processed having a cracked image as the subject image to be segmented.
The image feature detector is a crack detector that has learned to detect the crack image by the first machine learning, and generates a crack image in which the crack image is detected as the emphasized image.
The segmentation device according to any one of claims 1 to 3, wherein the segmentation device learns by the second machine learning and extracts a region corresponding to the crack image based on the crack image and the processed image. Image processing device. The image processing apparatus according to claim 6, wherein the segmentation device segments an inspection region including the crack image. The image processing apparatus according to claim 7, wherein the segmentation device segments the inspection area of the thick line including the crack image. The image processing apparatus according to claim 8, wherein the segmentation device segments the rectangular inspection area including the crack image. The image processing apparatus according to claim 6, wherein the segmentation device generates a segmentation image in which a region corresponding to the segmented crack image is emphasized. The image processing apparatus according to any one of claims 7 to 9, wherein the segmentation device generates a segmentation image in which the segmented inspection area is emphasized. By the processor executing the program
An image acquisition step of acquiring a processing target image having a subject image to be segmented, and
An image feature detection step of generating an emphasized image in which the features of the subject image learned by the first machine learning are emphasized by the mode learned by the first machine learning, and
A segmentation step for segmenting a region corresponding to the subject image according to the mode learned by the second machine learning based on the emphasized image and the processed image.
Image processing method to perform. An image acquisition step of acquiring a processing target image having a subject image to be segmented, and
An image feature detection step of generating an emphasized image in which the features of the subject image learned by the first machine learning are emphasized by the mode learned by the first machine learning, and
A segmentation step for segmenting a region corresponding to the subject image according to the mode learned by the second machine learning based on the emphasized image and the processed image.
A program that causes a computer to execute an image processing process including.

Images