WO2020158630A1

WO2020158630A1 - Detecting device, learner, computer program, detecting method, and method for generating learner

Info

Publication number: WO2020158630A1
Application number: PCT/JP2020/002630
Authority: WO
Inventors: 中村　聡
Original assignee: 株式会社カネカ
Priority date: 2019-01-31
Filing date: 2020-01-24
Publication date: 2020-08-06
Also published as: JPWO2020158630A1

Abstract

Provided are a detecting device, a learner, a computer program, a detecting method, and a method for generating a learner, with which it is possible for a minute detection target to be detected accurately. This detecting device is provided with: a candidate region extracting unit which subjects a first image to image processing to extract one or a plurality of candidate regions each containing a detection target candidate; an image generating unit which generates one or a plurality of second images each containing at least one of the extracted candidate regions; a first learner; and a detecting unit which inputs the generated second image into the first learner to detect the presence or absence of the detection target.

Description

Detecting device, learning device, computer program, detecting method, and learning device generating method

The present invention relates to a detection device, a learning device, a computer program, a detection method, and a learning device generation method.

An EL inspection using electroluminescence (EL) is used as one of the inspection methods for the presence/absence of defects in solar cell modules. In the EL inspection, an electric current is supplied to each cell constituting the solar cell module to cause the cell itself to emit light, and cracks, disconnections, and the like are confirmed based on an image captured by a camera on the surface of the cell (see Patent Document 1).

Japanese Patent No. 6208843

However, the brightness distribution and distortion of the image differ from cell to cell, and there are minute detection targets (for example, defects) that are difficult to confirm visually. Further, there are various types of shapes even for minute detection targets, and the quality determination of the cell also differs depending on the type of detection target. Therefore, it is desired to accurately detect a minute detection target.

The present invention has been made in view of such circumstances, and provides a detection device, a learning device, a computer program, a detection method, and a learning device generation method capable of accurately detecting a minute detection target. To aim.

A detection device according to an embodiment of the present invention includes a candidate area extraction unit that performs image processing on a first image to extract one or a plurality of candidate areas including a detection target candidate, and the candidate area extraction unit. An image generation unit that generates one or a plurality of second images including at least one of the extracted candidate regions, a first learning device, and the second image generated by the image generation unit are input to the first learning device. And a detection unit that detects the presence or absence of a detection target.

A learning device according to an embodiment of the present invention includes a first learning device generated by using a first label image including a detection target and a second label image including no detection target as learning data, A second learning device generated by using, as learning data, a plurality of types of label images divided into different types of detection targets included in one label image.

A computer program according to an embodiment of the present invention causes a computer to perform a process of acquiring a first image and one or a plurality of candidate regions including a detection target candidate by performing image processing on the acquired first image. A process of extracting, a process of generating one or a plurality of second images including at least one of the extracted candidate regions, and a process of inputting the generated second images to a first learning device and detecting the presence or absence of a detection target And execute.

A detection method according to an embodiment of the present invention acquires a first image, performs image processing on the acquired first image, extracts one or a plurality of candidate regions including a detection target candidate, and extracts the extracted candidate region. One or a plurality of second images including at least one of the generated candidate areas is generated, and the generated second image is input to the first learning device to detect the presence or absence of the detection target.

A learning device generation method according to an embodiment of the present invention acquires a first label image that includes a detection target and a second label image that does not include a detection target, and acquires the first label image and the second label image. A first learning device is generated by using the label image as learning data, and a second learning device is generated by using a plurality of types of label images divided into different types of detection targets included in the first label image as learning data. To generate.

According to the present invention, a minute detection target can be accurately detected.

It is a block diagram which shows an example of a structure of the detection apparatus of this Embodiment. It is a schematic diagram which shows an example of a defect target. It is a schematic diagram which shows an example of the detection target which the detection apparatus of this Embodiment detects. It is a schematic diagram which shows an example of generation of a second image. It is a schematic diagram which shows an example of a structure of a 1st learning device and a 2nd learning device. It is a schematic diagram which shows an example of the generation method of the 1st learning device using a 1st label image. It is a schematic diagram which shows an example of the generation method of the 1st learning device using a 2nd label image. It is a schematic diagram which shows an example of the generation method of the 2nd learning device using a classification label image. It is a schematic diagram which shows an example of the detection method of the detection target by the detection apparatus of this Embodiment. It is a schematic diagram which shows the other example of a structure of a learning device. It is a schematic diagram which shows an example of a structure of a detection system. It is a flow chart which shows an example of a procedure of processing by a detecting device.

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an example of the configuration of the detection device 50 according to the present embodiment. The detection device 50 includes a control unit 51 that controls the entire device, an input unit 52, a defect determination unit 53, a candidate area extraction unit 54, a storage unit 55, an image generation unit 56, a learning device 57, and a detection unit 58. The learning device 57 includes a first learning device 571 and a second learning device 572. The storage unit 55 can store required data such as an input image described later and a result of detection processing.

The input unit 52 has a function as an acquisition unit, and acquires the input image in which the imaging target is captured. The imaging target is an inspection target, and for example, in the case of a solar cell module, each cell that constitutes the module can be used, but the present invention is not limited to this.

The defect judgment unit 53 has a function as a judgment unit and judges whether or not a predetermined defect target is included in the input image. The defect determination unit 53 has an image processing function and can determine the presence or absence of a predetermined defect target based on the input image. When the imaging target is a cell, the predetermined defect target includes, but is not limited to, single cell abnormality, disconnection (also referred to as finger disconnection), cleaning mark (also referred to as water mark), breakage coexisting crack, and the like.

2 is a schematic diagram showing an example of a defect target. FIG. 2 shows an image (also referred to as a cell image) obtained by picking up an image of a cell constituting a solar cell module. Horizontal black lines in the cell image indicate metal lines in the cell. 2A schematically shows a case where there is a finger disconnection in the cell. The finger disconnection is a relatively large width disconnection. For example, if there is one location in the cell, the cell is determined to be defective. FIG. 2B schematically shows the case where water marks are present in the cell. The water trace is a cleaning trace remaining after the cleaning step, and if there is one in the cell, the cell is determined to be defective. FIG. 2C schematically shows a case where a crack having a wire break exists in the cell. If there is one location in the cell, the cell is determined to be defective. 2D schematically shows a cell abnormality. The cell abnormality is when the cell does not emit light or when the amount of emitted light is extremely low, and the cell is determined to be defective.

The defect determination unit 53 outputs the input image determined to have no predetermined defect target to the candidate region extraction unit 54 as the first image. As a result, it is possible to exclude in advance defect targets that are relatively easy to detect, narrow down only the images that may include minute detection targets that are difficult to detect, and perform processing after the candidate region extraction unit 54.

When the defect determining unit 53 determines that there is a predetermined defect target, the detecting unit 58 detects the cell as a defective product.

FIG. 3 is a schematic diagram showing an example of a detection target detected by the detection device 50 of the present embodiment. The detection target is a minute defect target that is difficult to determine by the defect determination unit 53, and various types such as shape and size may exist. In the present specification, as shown in FIG. 3, the detection target is divided into four types A, B, C, and D for convenience according to the shape and the type, for example. The types are not limited to four types. In the case of cells that make up a solar cell module, the detection target can be cracks of different types such as shape and size, and although the detection target is described as "a target called a crack" in this specification, the detection target is a detection target. Is not limited to cracks. Further, the shape and size of the crack shown in FIG. 3 are examples, and the invention is not limited to the example of FIG.

The candidate area extraction unit 54 performs image processing on the first image output by the defect determination unit 53 to extract one or more candidate areas including the detection target candidate. The image processing includes normal image processing such as density correction, noise removal, edge detection, and feature amount extraction. The detection target candidate is, for example, a region including a minute defect that is difficult to detect by visual inspection or the determination by the defect determination unit 53, and is a region including a minute detection target candidate.

The detection unit 58 can determine that the cell corresponding to the first image from which the detection target candidate cannot be extracted by the candidate area extraction unit 54 is a good product.

The image generation unit 56 generates one or more second images including at least one of the candidate regions when the candidate region extraction unit 54 extracts a plurality of candidate regions in one first image. At least one means that a large number of candidate regions are not included, and can be, for example, one, two, or three (a small number). In this specification, the case where the second image includes one candidate region will be described, but the number of candidate regions included in one second image is not limited to one.

FIG. 4 is a schematic diagram showing an example of generation of the second image. As shown in FIG. 4, it is assumed that the candidate area extraction unit 54 extracts three candidate areas of types A, C, and D as detection targets in one first image. In this case, the image generation unit 56 generates three second images including the second image including the type A candidate area, the second image including the type C, and the second image including the type D.

The image generation unit 56 can generate at least one of the candidate regions extracted by the candidate region extraction unit 54 at a predetermined position of the predetermined image to generate the second image. The predetermined image can be, for example, an image having a constant brightness value (an image having a high brightness value, a bright image). The candidate area can be a rectangular area surrounding the detection target candidate according to the shape and size of the detection target candidate. The predetermined position can be, for example, the center (midpoint) of the predetermined image. For example, the candidate areas can be arranged so that the center of the candidate area matches the center of the predetermined image.

Since the position of the candidate area on the second image can be set to the same position for each second image, it is possible to detect a minute detection target with higher accuracy than when the position of the candidate area varies from image to image.

By setting the background color of the specified image to white, it is possible to increase the contrast between the minute detection target and the background.

The first learning device 571 and the second learning device 572 can be configured by, for example, a multilayer neural network (deep learning), and for example, a convolutional neural network (Convolutional Neural Network) can be used. Learning may be used. In the present specification, the first learning device 571 and the second learning device 572 are described as being configured by a convolutional neural network, but the present invention is not limited to this.

FIG. 5 is a schematic diagram showing an example of the configuration of the first learning device 571 and the second learning device 572. The convolutional neural network comprises a hidden layer between the input layer and the output layer. The hidden layer includes a convolutional layer and a pooling layer having a plurality of stages, and a fully-bonded layer in the final stage. The number of convolutional layers, pooling layers, and total bonding layers can be appropriately determined.

Next, the generation method (learning method) of the first learning device 571 and the second learning device 572 will be described. The control unit 51 can perform the learning process of the first learning device 571 and the second learning device 572. The control unit 51 and the learning device 57 are, for example, a CPU (for example, a multi-processor having a plurality of processor cores mounted), a GPU (Graphics Processing Units), a DSP (Digital Signal Processors), and an FPGA (Field-Programmable Gate Arrays). It can be configured by combining hardware such as.

The first learning device 571 can generate the first label image including the detection target and the second label image including no detection target as the learning data. First, the case of using the first label image will be described.

FIG. 6 is a schematic diagram showing an example of a generation method of the first learning device 571 using the first label image. The first label image can be a set of images containing only type A, a set of images containing only type B, a set of images containing only type C, and a set of images containing only type D. .. The first learning device 571 can generate the set of the first label images by giving it to the input layer and the label with the detection target (for example, level 11) to the output layer.

FIG. 7 is a schematic diagram showing an example of a method of generating the first learning device 571 using the second label image. The second label image can be a set of images that do not include any of types A, B, C, and D. The first learning device 571 can generate the set of the second label images by giving the set of the second label images to the input layer and the label (for example, label 12) having no detection target to the output layer.

As a result, when the second image generated by the image generation unit 56 includes any one of types A, B, C, and D as the detection target, the first learning device 571 detects that the detection target exists. Can be detected. If the second image does not include any of types A, B, C, and D as detection targets, the first learning device 571 can detect that there is no detection target.

Next, the generation method (learning method) of the second learning device 572 will be described. The second learning device 572 can generate a plurality of types of label images, which are included in the first label image and are classified according to different types of detection targets, as learning data.

FIG. 8 is a schematic diagram showing an example of a method of generating the second learning device 572 using the type label image. For example, when the set of first label images includes four types of detection targets of types A, B, C, and D, the type label images include a type A level image including type A and a type B level including type B. An image, a type C level image including type C, and a type D level image including type D are divided into four sets.

The second learner 572 gives a set of type A level images to the input layer and gives a label (for example, label 21) indicating that the type is type A to the output layer to generate a set of type B level images. Is given to the input layer, a label (for example, label 23) indicating that the type is type B is given to the output layer, a set of type C level images is given to the input layer, and the type is type C Is given to the output layer, a set of type D level images is given to the input layer, and a label showing that the type is type D (for example, label 27) is given to the output layer. Can be given and generated. When the label 21 is applied to the output layer, the types B, C, and D do not exist, so the labels 24, 26, and 28 may be used together. The same applies when the labels 23, 25 and 27 are given to the output layer.

With this, the second learning device 572 can detect the type of the detection target.

Among types A to D as detection targets, for example, for type D, the quality of the cell is determined according to the number of cells existing in the cell. The second learning device 572 can also perform learning so that the number of predetermined detection targets (for example, type D) can be detected. In this case, a type label image including one or more type D label images is given to the input layer of the second learning device 572, and the number of type D label images (for example, labels 0 to N) is given to the output layer. The second learning device 572 can be generated.

Next, a method of detecting the presence/absence of a detection target by the first learning device 571, and a method of detecting the type of detection target and the number of detection targets by the second learning device 572 will be described.

FIG. 9 is a schematic diagram showing an example of a method of detecting a detection target by the detection device 50 of the present embodiment. The defect determination unit 53 determines whether or not the input image includes a predetermined defect target. When there is a defect target, the detection unit 58 detects that there is a defect, that is, a defective product. The defect determination unit 53 outputs the input image having no defect target to the candidate region extraction unit 54 as the first image.

The candidate area extraction unit 54 performs image processing on the first image and extracts one or more candidate areas including the detection target candidate. When there is no candidate area, the detection unit 58 detects that it is a good product. When there is a candidate area, the candidate area extraction unit 54 outputs the first image including the candidate area to the image generation unit 56.

The image generation unit 56 generates the second images in which the candidate regions are arranged at the predetermined positions by the number of the extracted candidate regions, and outputs the generated second images to the first learning device 571 in order.

As described above, the first learning device 571 is generated so as to be able to detect the presence or absence of the detection target by using the teacher data. When the first learning device 571 determines that there is no detection target, the detection unit 58 detects that the cell is a good product. When the first learning device 571 determines that there is a detection target, the detection unit 58 detects that the cell is defective. In this way, the detection unit 58 can input the second image to the first learning device 571 and detect the presence or absence of the detection target. Since the second image input to the first learning device 571 includes only a small number of candidate regions, the detection accuracy of the first learning device 571 can be increased. In addition, even if it is difficult to detect a minute defect due to the influence of the luminance distribution and distortion of the image in the normal image processing, by using the first learning device 571, learning is performed until the detection accuracy reaches the required accuracy. Therefore, a minute detection target can be accurately detected.

Further, the detection unit 58 can sequentially input each of the second images to the first learning device 571 to detect the presence or absence of the detection target. For example, when five candidate regions including the detection candidate target are extracted, the five second images are sequentially input to the first learning device 571, and the first learning device 571 determines whether there is a detection target for each second image. To detect. As a result, it is possible to detect whether or not all the detection candidate targets are detection targets without omission.

As described above, the second learning device 572 is generated so as to be able to detect the type of the detection target using the teacher data.

The detection unit 58 has a function as a type detection unit, and of the second images input to the first learning device 571, the second image in which the detection target is detected is used as the third image by the second learning device 572. input. The second learning device 572 detects the type of the detection target. When a plurality of second images including candidate regions are input to the first learning device 571, the third images are the first images except for the second image which is determined by the first learning device 571 to have no detection target. It includes only the second image determined by the learning device 571 that the detection target is present. By using the second learning device 572, learning can be performed until the detection accuracy of the type of the detection target reaches the required accuracy, and a minute type of the detection target can be accurately detected.

As shown in FIG. 9, when the second learning device 572 determines the presence or absence of types A to C, the detection unit 58 can output the detection result for each type of types A to C. For example, when any one of types A to C is included, the detection unit 58 can detect the cell as a defective product.

In this way, the detection unit 58 can input the third image to the second learning device 572 and detect detection targets of different types. For example, when four third images including types A, B, C, and D are input to the second learning device 572, the detection unit 58 detects four different types of types A, B, C, and D. can do. Thus, even if there are various types of detection targets and different types, the types of detection targets can be detected.

Further, when the second learning device 572 determines the number of type D, the detection unit 58 can output the determination result of non-defective product/defective product according to the number of type D. For example, when four or more type D cracks are present in one cell, the cell can be determined as a defective product.

In this way, the detection unit 58 can sequentially input each of the third images to the second learning device 572 and detect the number of detection targets of the same type. For example, when three third images including type D are sequentially input to the second learning device 572, the second learning device 572 detects the type D three times, so the number of type D is detected as three. can do. With this, even when the quality of the imaging target is determined according to the number of minute detection targets existing in the imaging target, the number of detection targets can be detected, and thus the quality of the imaging target can be determined.

FIG. 10 is a schematic diagram showing another example of the configuration of the learning device 57. The learning device 57 includes a plurality of

second learning devices

572a, 572b, 572c,..., 572n. The

second learning devices

572a, 572b, 572c,..., 572n are learned so that detection targets of types A, B, C,... N can be detected.

For example, when there are types A and B as different types, it is possible to provide two second learning devices, a second learning device that detects type A and a second learning device that detects type B. The detection unit 58 inputs the fourth image excluding the third image in which one type is detected among the third images input to the one second learning device to another second learning device, and then the one type Different types can be detected. For example, when the third image including type A and the third image including type B are sequentially input to one second learning device, and the one second learning device detects type A, type A The third image including type B except for is input to the next second learning device as the fourth image. The next second learning device can detect type B. The same applies when the detection type is 3 or more. Thus, even if there are various types of detection targets and different types, the types of detection targets can be detected.

FIG. 11 is a schematic diagram showing an example of the configuration of the detection system. In the detection system of the present embodiment, the server 100 including the detection device 50 is connected to the communication network 1. Although only one server 100 is shown in FIG. 11, the number of servers 100 is not limited to one. A plurality of client devices 10 are connected to the communication network 1. The server 100 and the client device 10 have a communication function capable of transmitting and receiving required data.

When the client device 10 transmits the input image in which the imaging target is captured via the communication network 1 to the server 100, the server 100 receives the input image. The server 100 performs each process by the detection device 50 on the received input image. The server 100 transmits the detection result (for example, the quality of the imaging target, the presence or absence of the detection target, the type of the detection target, etc.) to the client device 10. Accordingly, the client device 10 side can obtain the detection result only by transmitting the input image, and the client device 10 side does not need to include the detection device 50.

FIG. 12 is a flowchart showing an example of a processing procedure by the detection device 50. Hereinafter, for convenience, the main body of processing will be described as the control unit 51. The control unit 51 acquires the input image (S11) and determines the presence/absence of a defect (S12). Here, the defect is, for example, the one illustrated in FIG. When there is no defect (NO in S12), the control unit 51 inputs the first image into the candidate area extraction unit 54 (S13). The first image is an input image determined to have no defect target.

The control unit 51 extracts a candidate area (S14) and determines the presence or absence of the candidate area (S15). If there is no candidate area (NO in S15), there is neither a defect target nor a detection target in the input image, so the imaging target (for example, cell) corresponding to the input image is determined as a non-defective item (S16), and the processing is performed. To finish.

If there is a candidate area (YES in S15), the control unit 51 generates a second image (S17). For example, when there are three candidate areas in the first image, three second images in which the respective candidate areas are arranged at predetermined positions are generated. The control unit 51 sequentially inputs the second image to the first learning device 571 (S18) and determines the presence/absence of a crack as a detection target (S19). When there is no crack (NO in S19), the control unit 51 performs the process of step S16.

If there is a crack (YES in S19), the control unit 51 inputs the third image into the second learning device 572 (S20). The third image is an image in which a crack is detected by the first learning device 571 among the second images input to the first learning device 571. The control unit 51 detects the type of crack (S21).

The control unit 51 determines the presence or absence of another third image (S22), and when there is another third image (YES in S22), repeats the processing from step S20. If there is no other third image (NO in S22), the control unit 51 determines whether there is a crack of types A to C (S23), and if there is no crack of types A to C (NO in S23), type It is determined whether the number of D cracks is equal to or greater than a predetermined value (S24).

When the number of type D cracks is not equal to or greater than the predetermined value (NO in S24), the control unit 51 performs the process of step S16. When there is a defect (YES in S12), there are cracks of types A to C (YES in S23), or when the number of type D cracks is equal to or more than a predetermined value (YES in S24), the control unit 51 The imaging target (eg, cell) corresponding to the input image is determined to be a defective product (S25), and the process ends. When there are a plurality of input images, the process shown in FIG. 12 may be repeated.

The detection device 50 can also be realized by using a computer including a CPU (processor), GPU, RAM (memory), and the like. That is, as shown in FIG. 12, by loading a computer program that defines the procedure of each processing into a RAM (memory) provided in the computer and executing the computer program by a CPU (processor), the detection device on the computer 50 can be realized. The computer program may be recorded in a recording medium and distributed.

According to the present embodiment, it is possible to detect a detection target with high accuracy even if it is a minute detection target that is difficult to visually confirm, or even a minute detection target having various shapes and sizes. Also, various types of detection targets can be detected.

In the above-described embodiment, the cells constituting the solar cell module are used as the imaging target and the "target called a crack" is used as the detection target. However, the imaging target and the detection target are limited to these. It is not something that will be done.

The detection device according to the present embodiment performs image processing on the first image to extract one or more candidate regions including the detection target candidate, and the candidate region extraction unit extracts the candidate regions. An image generation unit that generates one or a plurality of second images including at least one of the candidate regions, a first learning device, and the second image generated by the image generation unit are input to the first learning device and detected. And a detection unit that detects the presence or absence of a target.

The computer program according to the present embodiment causes a computer to perform a process of acquiring a first image and an image process on the acquired first image to extract one or a plurality of candidate regions including a detection target candidate. The process, the process of generating one or a plurality of second images including at least one of the extracted candidate regions, and the process of inputting the generated second image to the first learning device and detecting the presence or absence of the detection target. Let it run.

The detection method according to the present embodiment obtains a first image, performs image processing on the obtained first image, extracts one or a plurality of candidate regions including a detection target candidate, and extracts the candidate region. One or a plurality of second images including at least one of the candidate regions is generated, and the generated second image is input to the first learning device to detect the presence or absence of the detection target.

The candidate area extraction unit performs image processing on the first image and extracts one or more candidate areas including the detection target candidate. The first image is, for example, an image determined to have no visually observable defect in the input image in which the imaging target is captured. The image processing includes normal image processing such as density correction, noise removal, edge detection, and feature amount extraction. The detection target candidate is, for example, a region including a minute defect that is difficult to detect by visual observation, and is a region including a minute detection target candidate. It should be noted that the first image from which the detection target candidate cannot be extracted can be determined to be a non-defective item.

The image generation unit generates one or more second images including at least one of the extracted candidate areas. At least one means that a large number of candidate regions are not included, and can be, for example, one, two, or three (a small number). For example, when three detection target candidates A, B, and C are extracted from one first image, the image generation unit causes the second image including the detection target candidate A and the second image including the detection target candidate B. , Second three images of the second image including the detection target candidate C are generated.

The first learning device can be configured by, for example, a multi-layer neural network (deep learning), for example, a convolutional neural network (Convolutional Neural Network) can be used, but other machine learning may be used. The first learning device is generated so as to be able to detect the presence/absence of a detection target using the teacher data.

The detection unit inputs the second image to the first learning device and detects the presence or absence of the detection target. Since the second image input to the first learning device includes only a small number of candidate regions, the detection accuracy of the first learning device can be improved. In addition, in the normal image processing, even if it is difficult to detect a minute defect due to the influence of the luminance distribution or distortion of the image, by using the first learning device, learning is performed until the detection accuracy reaches the required accuracy. Therefore, it is possible to accurately detect a minute detection target.

In the detection device according to the present embodiment, the image generation unit arranges at least one of the candidate regions extracted by the candidate region extraction unit at a predetermined position of a predetermined image to generate a second image.

The image generation unit arranges at least one of the extracted candidate areas at a predetermined position of the predetermined image to generate the second image. The predetermined image can be, for example, an image having a constant brightness value (an image having a high brightness value, a bright image). The candidate area can be a rectangular area surrounding the detection target candidate according to the shape and size of the detection target candidate. The predetermined position can be, for example, the center (midpoint) of the predetermined image. For example, the candidate areas can be arranged so that the center of the candidate area matches the center of the predetermined image.

In the detection device according to the present embodiment, the detection unit sequentially inputs each of the second images to the first learning device to detect the presence/absence of a detection target.

The detection unit sequentially inputs each of the second images to the first learning device and detects the presence or absence of the detection target. For example, when five candidate regions including the detection candidate target are extracted, the five second images are sequentially input to the first learning device, and the first learning device detects the presence or absence of the detection target for each second image. To do. Thereby, it is possible to detect whether or not all the detection candidate targets are the detection targets.

In the detection device according to the present embodiment, the first learning device is generated using the first label image including the detection target and the second label image including no detection target as the learning data.

The first learning device is generated using the first label image including the detection target and the second label image including no detection target as the learning data. For example, it is assumed that the detection target is divided into types A, B, C, and D according to the shape and the type, for convenience. The first label image can be a set of images that include only type A, images that include only type B, images that include only type C, and images that include only type D. The first learning device can be generated using the first label image and the label with the detection target. The second label image can be a set of images that do not include any of types A, B, C, and D. The first learning device can be generated using the second label image and the label having no detection target.

With this, when the second image includes any one of types A, B, C, and D, it can be detected that the detection target exists. If the second image does not include any of types A, B, C, and D, it can be detected that the detection target does not exist.

The detection device according to the present embodiment inputs a third image in which a detection target is detected among the second images input to the second learning device and the first learning device to the second learning device. And a type detection unit that detects the type of the detection target.

The second learning device can be configured by, for example, a multilayer neural network (deep learning), for example, a convolutional neural network (Convolutional Neural Network) can be used, but other machine learning may be used. The second learning device is generated so that the type of the detection target can be detected using the teacher data.

The type detection unit inputs, to the second learning device, the third image in which the detection target is detected among the second images input to the first learning device, and detects the type of the detection target. When a plurality of second images including candidate regions are input to the first learning device, the third images are the first learning device except for the second image in which the first learning device determines that the detection target does not exist. Only the second image for which it is determined that the detection target exists is included. By using the second learning device, it is possible to learn until the detection accuracy of the type of the detection target reaches the required accuracy, and it is possible to accurately detect the minute type of the detection target.

In the detection device according to the present embodiment, the type detection unit inputs each of the third images to the second learning device in order and detects the number of detection targets of the same type.

The type detection unit sequentially inputs each of the third images to the second learning device and detects the number of detection targets of the same type. For example, if the three third images including the type D are sequentially input to the second learning device, the second learning device detects the type D three times. Therefore, the number of the type D is detected as three. You can With this, even when the quality of the imaging target is determined according to the number of minute detection targets existing in the imaging target, the number of detection targets can be detected, and thus the quality of the imaging target can be determined.

In the detection device according to the present embodiment, the type detection unit inputs the third image to the second learning device and detects detection targets of different types.

The type detection unit inputs the third image to the second learning device and detects a detection target of a different type. For example, when four third images including types A, B, C, and D are input to the second learning device, the type detection unit detects four different types of types A, B, C, and D. be able to. Thus, even if there are various types of detection targets and different types, the types of detection targets can be detected.

The detection device according to the present embodiment includes a plurality of the second learning devices, and the type detection unit detects the one type of the third images input to the one second learning device. The fourth image excluding the three images is input to another second learning device to detect a type different from the one type.

Includes multiple second learners. For example, when there are types A and B as different types, it is possible to provide two second learning devices, a second learning device that detects type A and a second learning device that detects type B. The type detection unit inputs, to the other second learning device, the fourth image excluding the third image in which the one type is detected among the third images input to the one second learning device, and inputs the one type into the second learning device. And a type different from. For example, when the third image including type A and the third image including type B are sequentially input to one second learning device, and the one second learning device detects type A, type A The third image including type B except for is input to the next second learning device as the fourth image. The next second learning device can detect type B. Thus, even if there are various types of detection targets and different types, the types of detection targets can be detected.

In the detection device according to the present embodiment, the second learning device generates using a plurality of types of label images divided into different types of the detection target of the first label image including the detection target as learning data. I am doing it.

The second learning device is generated by using, as learning data, a plurality of types of label images divided into different types of detection targets included in the first label image. For example, when the set of first label images includes four types of detection targets of types A, B, C, and D, a type A level image including type A, a type B level image including type B, and a type C are selected. It is divided into four sets: a type C level image containing the type C and a type D level image containing the type D. The second learning device generates using a set of type A level images and a label indicating that the type is type A, and uses a set of type B level images and a label indicating that the type is type B Generated by using a set of type C level images and a label indicating that the type is type C, and using a set of type D level images and a label indicating that the type is type D can do.

The detection apparatus according to the present embodiment includes an acquisition unit that acquires an input image in which an imaging target is captured, and a determination unit that determines whether or not a predetermined defect target is included in the input image acquired by the acquisition unit. And the input image determined by the determination unit to have no defect target is the first image.

The acquisition unit acquires the input image in which the imaging target is captured. The imaging target is an inspection target, and for example, in the case of a solar cell module, each cell that constitutes the module can be used, but the present invention is not limited to this.

The judgment unit judges whether or not the input image includes a predetermined defect target. The determination unit has an image processing function and can determine the presence or absence of a predetermined defect target based on the input image. When the imaging target is a cell, the predetermined defect target includes, but is not limited to, single cell abnormality, disconnection (also referred to as finger disconnection), cleaning mark (also referred to as water mark), breakage coexisting crack, and the like. The input image determined by the determination unit to have no defect target is the first image. As a result, it is possible to exclude defects that are relatively easy to detect in advance, and narrow down and process only images that may include minute detection targets that are difficult to detect.

The learning device according to the present embodiment includes a first learning device generated by using, as learning data, a first label image including a detection target and a second label image including no detection target, and the first label. The second learning device is generated by using, as learning data, a plurality of types of label images that are divided into different types of detection targets included in the image.

The learning device generation method according to the present embodiment acquires a first label image including a detection target and a second label image including no detection target, and acquires the first label image and the second label image acquired. Is used as learning data to generate a first learning device, and a plurality of types of label images divided into different types of detection targets included in the first label image are used as learning data to generate a second learning device. ..

The learning device includes a first learning device and a second learning device. The first learning device and the second learning device can be configured by, for example, a multilayer neural network (deep learning), for example, a convolutional neural network (Convolutional Neural Network) can be used, but other machine learning can be performed. You may use.

The first learning device is generated using the first label image including the detection target and the second label image including no detection target as the learning data. For example, if the detection target is divided into types A, B, C, and D according to shape and type for convenience, the first label image is an image including only type A, an image including only type B, It can be a set of images including only type C and images including only type D. The first learning device can be generated using the first label image and the label with the detection target. The second label image can be a set of images that do not include any of types A, B, C, and D. The first learning device can be generated using the second label image and the label having no detection target.

With this, when different types of detection targets are included in the third image set, the detection target types can be detected.

In the learning device according to the present embodiment, the first label image includes at least one of candidate regions including detection target candidates.

The first label image includes at least one of the candidate areas including the detection target candidates. At least one means that a large number of candidate regions are not included, and can be, for example, one, two, or three (a small number). For example, when one learning image includes three detection target candidates A, B, and C, the first label image includes the image including the detection target candidate A, the image including the detection target candidate B, and the detection target candidate. There are three images including C. Since only a small number of candidate areas are included, the detection accuracy of the first learning device can be improved.

In the learning device according to the present embodiment, the first label image has a candidate area including a detection target candidate arranged at a predetermined position of a predetermined image.

In the first label image, a candidate area including a detection target candidate is arranged at a predetermined position of a predetermined image. The predetermined image can be, for example, an image having a constant brightness value (an image having a high brightness value, a bright image). The candidate area can be a rectangular area surrounding the detection target candidate according to the shape and size of the detection target candidate. The predetermined position can be, for example, the center (midpoint) of the predetermined image. For example, the candidate areas can be arranged so that the center of the candidate area matches the center of the predetermined image.

Since the position of the candidate area on the first label image can be set to the same position for each first label image, it is possible to accurately detect a minute detection target as compared with the case where the position of the candidate area varies from image to image. it can.

In the learning device according to the present embodiment, the predetermined image has a white background color.

The background color of the specified image is white. As a result, the contrast between the minute detection target and the background can be increased.

In the learning device according to the present embodiment, the first label image is an image in which it is determined that there is no predetermined defect target among the images captured by the imaging target.

The first label image is an image in which it is determined that there is no predetermined defect target among the images captured by the imaging target. The imaging target is an inspection target, and for example, in the case of a solar cell module, each cell that constitutes the module can be used, but the present invention is not limited to this. When the imaging target is a cell, the predetermined defect target includes, but is not limited to, single cell abnormality, disconnection (also referred to as finger disconnection), cleaning mark (also referred to as water mark), breakage coexisting crack, and the like. As a result, it is possible to exclude defects that are relatively easy to detect in advance, and narrow down and process only images that may include minute detection targets that are difficult to detect.

1 Communication Network 10 Client Device 50 Detection Device 51 Control Unit 52 Input Unit 53 Defect Judgment Unit 54 Candidate Area Extraction Unit 55 Storage Unit 56 Image Generation Unit 57 Learning Device 571

First Learning Device

572, 572a, 572b, 572c, 572n Second Learner 58 Detector 100 Server

Claims

A candidate area extraction unit that performs image processing on the first image to extract one or more candidate areas including detection target candidates;
An image generation unit that generates one or more second images including at least one of the candidate regions extracted by the candidate region extraction unit;
A first learner,
A detection unit that inputs the second image generated by the image generation unit to the first learning device and detects the presence or absence of a detection target.
The image generation unit,
The detection device according to claim 1, wherein at least one of the candidate regions extracted by the candidate region extraction unit is arranged at a predetermined position of a predetermined image to generate the second image.
The detection unit,
The detection device according to claim 1, wherein each of the second images is sequentially input to the first learning device to detect the presence or absence of a detection target.
The first learner is
The detection device according to any one of claims 1 to 3, which is generated by using a first label image including a detection target and a second label image including no detection target as learning data.
A second learner,
A type detection unit configured to input, to the second learning device, a third image in which a detection target has been detected among the second images input to the first learning device, and to detect the type of the detection target. 5. The detection device according to claim 4.
The type detection unit,
The detection device according to claim 5, wherein each of the third images is sequentially input to the second learning device to detect the number of detection targets of the same type.
The type detection unit,
The detection device according to claim 5 or 6, wherein the third image is input to the second learning device to detect different types of detection targets.
A plurality of the second learning device,
The type detection unit,
Of the third images input to one second learning device, a fourth image other than the third image in which one type is detected is input to another second learning device, and a different type from the one type. The detection device according to claim 5 or 6, which detects the.
The second learning device is
The first label image including a detection target is generated by using a plurality of types of label images divided into different types of the detection target as learning data. Detection device.
An acquisition unit that acquires an input image in which the imaging target is captured,
And a determination unit that determines whether or not a predetermined defect target is included in the input image acquired by the acquisition unit,
The detection device according to any one of claims 1 to 9, wherein an input image determined by the determination unit to have no defect target is the first image.
A first learning device generated by using, as learning data, a first label image including a detection target and a second label image including no detection target;
A second learning device, which is generated by using, as learning data, a plurality of types of label images divided into different types of detection targets included in the first label image.
The first label image is
The learning device according to claim 11, comprising at least one of candidate regions in which a detection target candidate is included.
The first label image is
The learning device according to claim 12, wherein a candidate region including a detection target candidate is arranged at a predetermined position of a predetermined image.
The learning device according to claim 13, wherein the predetermined image has a white background color.
The first label image is
The learning device according to any one of claims 12 to 14, wherein the image capturing target is an image determined to have no predetermined defect target among the captured images.
On the computer,
A process of acquiring the first image,
A process of performing image processing on the acquired first image to extract one or a plurality of candidate regions including detection target candidates;
A process of generating one or more second images including at least one of the extracted candidate regions;
A computer program that executes a process of inputting the generated second image to the first learning device and detecting the presence or absence of a detection target.
Take the first image,
Image processing is performed on the acquired first image to extract one or more candidate regions including detection target candidates,
Generating one or more second images including at least one of the extracted candidate regions,
A detection method of inputting the generated second image to a first learning device to detect the presence or absence of a detection target.
Acquiring a first label image including a detection target and a second label image including no detection target,
A first learning device is generated using the acquired first label image and second label image as learning data,
A method for generating a learning device, which generates a second learning device by using a plurality of types of label images divided into different types of detection targets included in the first label image as learning data.