JP2022186079A - Detection method and program - Google Patents

Abstract

To provide a detection method capable of easily and accurately detecting an elliptical shaped object included in an object.SOLUTION: A detection method includes the steps of: acquiring an image in which at least a part of an object including an elliptical shaped object is reflected (S11); dividing the acquired image into blocks in blocks of the same size as the size of the block used for teacher data at the time of learning a machine learning model (S12); and detecting an abnormal place of the object on the basis of the number of pixels representing the elliptical shaped object in each block obtained by inputting an image divided into the blocks in the already-learned model (S13).SELECTED DRAWING: Figure 3

Description

The present disclosure relates to detection methods and programs.

Conventionally, a technique for detecting an object in an image by image recognition using machine learning is known. Such technology is used, for example, to detect a product from an image captured on a production line and to manage the quality of the product.

JP 2018-022484 A

However, the conventional techniques described above have room for further improvement in terms of simply and accurately performing object detection using machine learning. In particular, it is known that detection accuracy of an object having rounded edges such as an elliptical object is low.

Accordingly, the present disclosure provides a detection method that can easily and accurately detect an elliptical object included in an object.

A detection method according to one aspect of the present disclosure acquires an image in which at least a part of an object including an elliptical object is captured, and uses the acquired image as training data for learning a machine learning model. Pixels representing the elliptical object in each of the blocks obtained by dividing the image into blocks of the same size and inputting the block-divided image into a trained model that is the machine learning model that has been trained. based on the number of, detecting an abnormal location of the object, in the detection of the abnormal location, the number of pixels output from the learned model is greater than the first threshold and more than the second threshold If the number of pixels output from the learned model is less than or equal to the first threshold value, or the second If it is equal to or greater than the threshold, it is determined that the location of the object corresponding to the block is the abnormal location.

A program according to an aspect of the present disclosure is a program for causing a computer to execute the detection method.

According to the present disclosure, there is provided a detection method capable of detecting an abnormal portion of an object simply and accurately.

FIG. 1 is a diagram for explaining an overview of a detection system according to an embodiment. FIG. 2 is a block diagram showing an example of the functional configuration of the detection system according to the embodiment. FIG. 3 is a flow chart showing an example of the operation of the detection device according to the embodiment. FIG. 4 is a flow chart showing the detailed flow of step S13 in FIG. FIG. 5 is a diagram showing an example of an image acquired in step S11 of FIG. FIG. 6 is a diagram showing an example of processing the image acquired in step S11 of FIG. FIG. 7 is a diagram for explaining the process of detecting an abnormal location. FIG. 8 is a diagram showing an example of a pixel histogram of detection results. FIG. 9 is a diagram illustrating an example of determination results. FIG. 10 is a diagram illustrating an example of filtering processing. 11 is a diagram showing the results of Comparative Example 1 and Example 1. FIG. 12 is a diagram showing the results of Comparative Example 2 and Example 2. FIG.

Hereinafter, embodiments will be specifically described with reference to the drawings. It should be noted that the embodiments described below are all comprehensive or specific examples. Numerical values, shapes, materials, components, arrangement positions and connection forms of components, steps, order of steps, and the like shown in the following embodiments are examples, and are not intended to limit the present disclosure. Further, among the constituent elements in the following embodiments, constituent elements not described in independent claims will be described as optional constituent elements.

Each figure is a schematic diagram and is not necessarily strictly illustrated. Moreover, in each figure, the same code|symbol is attached|subjected with respect to substantially the same structure, and the overlapping description may be abbreviate|omitted or simplified.

(Embodiment)
[Constitution]
First, the configuration of the detection system according to the embodiment will be described. FIG. 1 is a diagram for explaining an overview of a detection system according to an embodiment. FIG. 2 is a block diagram showing an example of the functional configuration of the detection system according to the embodiment.

As shown in FIG. 1, the detection system 100 is a system for detecting an abnormal portion of an object 1 including an elliptical object. An abnormal portion is, for example, a portion where a shape abnormality such as chipping occurs in an elliptical object. The detection system 100 is used, for example, for quality inspection of the object 1 . The object 1 is an object including an elliptical object, for example, an object having an elliptical object on its surface, an object composed of an elliptical object, or an object including an elliptical object as part of its configuration. The object may include an elliptical object and may be made of various materials such as fiber, metal, resin, wood, carbon, gel, or clay. The elliptical shape may be, for example, elliptical, polygonal with rounded corners, or circular in top view. Although cloth will be described below as an example of the object 1 including an elliptical object, it is only an example and is not limited to this.

As shown in FIGS. 1 and 2, the detection system 100 includes, for example, a detection device 10 and an imaging device 30. FIG. Also shown in FIG. 1 are a lighting device 20 and a transport device 40 . Detection system 100 may comprise illumination device 20 and transport device 40 . The detection system 100 may detect the presence or absence of an elliptical object included in the object 1 transported by the transport device 40 . The conveying device 40 is, for example, a conveyor.

Each configuration of the detection system 100 will be described below.

[Lighting device]
The illumination device 20 is, for example, a light emitting device that illuminates the object 1 within the angle of view of the imaging device 30 . Two lighting devices 20 may be provided, or three lighting devices 20 may be provided. In this case, the two or more illumination devices 20 irradiate light from different directions to the area captured by the imaging device 30 . The type of light source is not particularly limited, but may be, for example, a light source that emits white light based on an LED (Light Emitting Diode). Also, when two or more illumination devices 20 are provided, each illumination device 20 may include light sources that emit light of the same wavelength or the same wavelength range, or may include light sources that emit light of different wavelengths. good too. It may be appropriately selected according to the type of the object 1, the color or surface shape of the elliptical object included in the object 1, and the like.

[Imaging device]
The imaging device 30 captures an image including at least part of the object 1 . The image may be a still image or a moving image. The image may be a monochrome image or a color image. The imaging device 30 captures an image including at least a portion of the object 1 including an elliptical object by detecting light with which the illumination device 20 illuminates the object 1 . The imaging device 30 has an imaging element for detecting light, such as a CCD (Charge Coupled Device) image sensor, a CMOS (Complementary Metal Oxide Semiconductor) image sensor, or the like.

[Detection device]
The detection device 10 is a device that detects an abnormality in an elliptical object included in the object 1 based on an image captured by the imaging device 30 . The detection device 10 is, for example, a stationary information terminal such as a personal computer, but may be a portable information terminal such as a smart phone or a tablet terminal. 1 and 2, the detection device 10 and the imaging device 30 are separate bodies, but the imaging device 30 may be a camera provided in the detection device 10. FIG. The detection device 10 includes, for example, a communication unit 11, a control unit 12, a storage unit 13, a learning unit 14, a display unit, and an operation reception unit.

[Communication]
The communication unit 11 is a communication module (communication circuit) for the detection device 10 to communicate with the illumination device 20, the imaging device 30, and the transport device 40 via the local communication network. The communication unit 11 is, for example, a wireless communication circuit that performs wireless communication, but may be a wired communication circuit that performs wired communication. A communication standard for communication performed by the communication unit 11 is not particularly limited.

[Control part]
The control unit 12 performs information processing for controlling the operation of the detection device 10 . The control unit 12 is realized by, for example, a microcomputer, but may be realized by a processor or a dedicated circuit. Specifically, the control unit 12 includes an acquisition unit 12a, an image processing unit 12b, a determination unit 12c, and an output unit 12d. The acquisition unit 12a, the image processing unit 12b, the determination unit 12c, and the output unit 12d are all implemented by a processor executing a program for performing the above information processing.

The acquisition unit 12a acquires an image including at least a part of the object 1 including an elliptical object.

The image processing unit 12b generates learning image data or detection image data by performing image processing on the image (image data) acquired by the acquisition unit 12a. Specifically, the image processing unit 12b generates learning image data or detection image data by dividing the image data acquired by the acquisition unit 12a into blocks of a predetermined size.

For example, when generating learning image data, an image (referred to as a first image) showing at least a portion of the object 1 is divided into a plurality of blocks (referred to as first blocks) of a predetermined size, and the first A plurality of pieces of learning image data are generated by repeatedly redividing the first image into a predetermined size by shifting the first block vertically, horizontally, or diagonally within a range not exceeding the blocks. The teacher data for machine learning is composed of a plurality of image data for learning and annotations indicating the regions of each of the plurality of elliptical objects associated with the plurality of image data for learning as correct data.

Further, for example, when generating image data for detection, the image processing unit 12b converts an image (image data) acquired by the acquisition unit 12a into a block having the same size as the block size used for the teacher data when learning the machine learning model. Generates a block-divided image with blocks of size. At this time, prior to dividing the image into blocks, the image processing unit 12b excludes an area other than at least a part of the object 1 shown in the image (in other words, an area where the object 1 is not shown) from the detection target, A region to be detected in the image may be divided into blocks of the same size as that used as teacher data when learning the machine learning model.

Further, when the determination unit 12c determines that an abnormal portion has been detected in the object 1, the image processing unit 12b outputs a marking in which the abnormal portion in the image is marked as a result of the determination (hereinafter referred to as determination result). An image may be generated.

The determination unit 12c uses a learned machine learning model (hereinafter referred to as a learned model) stored in the storage unit 13 to divide an image (so-called detection image data) into blocks by the image processing unit 12b. , based on the number of pixels representing the elliptical object in each block obtained by inputting to the trained model, an abnormal portion of the object 1 is detected. More specifically, when the number of pixels output from the trained model is greater than the first threshold and less than the second threshold in detecting an abnormal location, the determining unit 12c determines that the block corresponds to If it is determined that the part of the object 1 is not an abnormal part, and the number of pixels output from the learned model is equal to or less than the first threshold, or if it is equal to or more than the second threshold, the object corresponding to the block 1 is determined to be an abnormal location. The threshold may be appropriately set according to the required detection accuracy and the type of detection target.

The output unit 12d outputs the result determined by the determination unit 12c (hereinafter referred to as determination result). The determination result is output in at least one of images, characters, symbols, and voices, for example. For example, when an abnormal portion is detected in the object 1, the output unit 12d outputs the marking image generated by the image processing unit 12b as the detection result.

[Memory part]
The storage unit 13 is a storage device that stores control programs and the like executed by the control unit 12 . The storage unit 13 may temporarily store the teacher data and the detection image data. The storage unit 13 updates the stored learned model to a machine learning model (so-called learned model) generated by the learning unit 14 . The storage unit 13 is implemented by, for example, a semiconductor memory.

[Study Department]
The learning unit 14 performs machine learning using teacher data. The learning unit 14 receives an image divided into blocks of a predetermined size by machine learning, and generates a machine learning model that outputs the number of pixels representing an elliptical object in each of a plurality of blocks included in the image. do. A machine learning model is, for example, a convolutional neural network (CNN). The machine learning model is not particularly limited as long as it is a CNN. For example, it may be a DeepCrack (A Deep Hierarchical Feature Learning Architecture for Crack Segmentation) Network. A trained machine learning model (a so-called trained model) includes learned parameters adjusted by machine learning. The generated learned model is stored in the storage unit 13 . The learning unit 14 is implemented, for example, by a processor executing a program stored in the storage unit 13 . The teacher data used for learning the machine learning model includes a plurality of first blocks obtained by dividing a first image showing at least part of the object 1 into predetermined sizes, and a plurality of elliptical objects included in the first blocks. a plurality of second blocks obtained by re-dividing the first block into a predetermined size by shifting the first block within a range not exceeding the first block in the first image; and second data composed of annotations indicating respective regions of a plurality of elliptical objects included in the two blocks. The direction in which the first block is shifted when dividing the first image is any of vertical, horizontal, and diagonal directions, and may be shifted by an arbitrary distance as long as the range of the first block is not exceeded. In this way, it is possible to create a plurality of teacher data from the first image in which at least part of the object 1 is shown.

[Display part]
The display unit 15 is a display device that displays images under the control of the control unit 12 . The display unit 15 is implemented by, for example, a liquid crystal panel or an organic EL (Electro Luminescence) panel.

[Operation reception part]
The operation reception unit 16 receives a user's operation. The operation reception unit 16 is specifically realized by a mouse, a microphone, a touch panel, or the like.

Note that the operation reception unit 16 may include a microphone (not shown) or a speaker (not shown). A microphone acquires sound and outputs an audio signal according to the acquired sound. A microphone is, specifically, a condenser microphone, a dynamic microphone, a MEMS microphone, or the like. The speaker outputs voice (machine voice), for example, as a response to the spoken voice captured by the microphone. This allows the user to interactively input a scene control execution instruction.

Note that the operation reception unit 16 may include a camera (not shown). The camera captures images of the user operating the detection device 10 . Specifically, the camera captures movements of the user's mouth, eyes, fingers, or the like. In this case, the operation accepting unit 16 accepts the user's operation based on the user's image captured by the camera. A camera is implemented by, for example, a CMOS image sensor.

Note that, as described above, the detection device 10 may include a camera. In this case, the camera may take an image showing at least part of the object 1 including the elliptical object.

[motion]
Next, operation of the detection system 100 according to this embodiment will be described. FIG. 3 is a flow chart showing an example of the operation of the detection device 10 according to the embodiment.

First, the imaging device 30 of the detection system 100 takes an image in which at least part of the object 1 is captured. At this time, the illumination device 20 may illuminate the inside of the angle of view of the imaging device 30 . The imaging device 30 transmits the captured image to the detection device 10 . Note that although an example in which the imaging device 30 is an external device of the detection device 10 is described here, the imaging device 30 may be provided in the detection device 10 .

Next, as shown in FIG. 3, the detection device 10 acquires an image of at least a part of the object captured by the imaging device 30 (S11). FIG. 5 is a diagram showing an example of an image acquired in step S11 of FIG. FIG. 6 is a diagram showing an example of processing the image acquired in step S11 of FIG. For example, as shown in FIG. 5, when the acquired image is a moving image, the detecting device 10 may extract a plurality of frame images (so-called still images) from the moving image. The detection device 10 performs the following processing on each still image.

For example, as shown in (a) of FIG. 6, when the detection device 10 acquires an image (still image), as shown in (b) of FIG. Sometimes, the blocks are divided into blocks of the same size as the blocks used for the teacher data (S12).

Next, the detection device 10 inputs the block-divided image into the trained model, and detects an abnormal portion of the object based on the number of pixels representing the elliptical object in each block (S13). For example, as shown in (c) of FIG. 6, the detection device 10 outputs an image in which pixels representing elliptical objects in each block in the image are shown in black as the output result of the learned model.

FIG. 4 is a flow chart showing the detailed flow of step S13 in FIG.

The detection device 10 inputs the block-divided image to the trained model, and outputs the number of pixels representing an elliptical object in each block (S21). More specifically, the detection device 10 derives the number of pixels representing the elliptical shape in each block in the image from the output result of the trained model ((c) in FIG. 6). FIG. 7 is a diagram for explaining the process of detecting an abnormal location. In FIG. 7, blocks determined to be normal and blocks determined to be abnormal are arbitrarily extracted from the output result shown in (c) of FIG.

Next, the detection device 10 starts loop processing for each block (S22). The detection device 10 determines whether or not the number of pixels output from the trained model in step S21 is greater than the first threshold and less than the second threshold (S23). FIG. 8 is a diagram showing an example of a pixel histogram of detection results. The first threshold and the second threshold are set based on the histogram shown in FIG. 8, for example. The detection device 10 aggregates the number of pixels representing an elliptical object in each of the plurality of output blocks, generates a histogram of the aggregated numbers of pixels in the plurality of blocks, and displays the generated histogram on the display unit 15. good too. The user may set the above threshold according to the desired accuracy from the displayed histogram.

When the detection device 10 determines that the number of pixels is greater than the first threshold and less than the second threshold (Yes in S23), it is determined that the location of the object 1 corresponding to the block is not an abnormal location. Determine (S24). For example, assuming that the first threshold for the number of pixels representing an elliptical object in a block (the number of pixels in FIG. 7) is set to 35000 and the second threshold is set to 37000, (a) to (e) in FIG. ) has more pixels than the first threshold value and less than the second threshold value, the part of the object 1 corresponding to the divided image is determined to be a normal part. .

On the other hand, when the detection device 10 determines that the number of pixels is equal to or less than the first threshold value or equal to or more than the second threshold value (No in S23), the part of the object 1 corresponding to the block is an abnormal part. It is determined that there is (S25). For example, since the number of pixels of the divided image in (f) of FIG. 7 is equal to or less than the first threshold value, the portion of the object 1 corresponding to the divided image is determined to be an abnormal portion.

When the detection device 10 has performed the above processing for all blocks included in the image, the loop processing for each block ends (S26).

Next, the detection device 10 determines whether or not an abnormal portion is detected in the object 1 (more specifically, the portion corresponding to the object 1 shown as the subject in the image) (S27). FIG. 9 is a diagram illustrating an example of determination results. For example, as shown in (a) of FIG. 9, the detection device 10 determines whether or not an abnormal point is detected in the object 1, and if it is determined that an abnormal point is detected, the abnormal point is detected. The divided image determined as is marked.

When the detecting device 10 determines that an abnormal point is detected in the object 1 (Yes in S27), it generates a marking image (see (b) of FIG. 9) in which the abnormal point in the image is marked (S28), The generated marking image is output as a determination result (S29).

On the other hand, when the detecting device 10 determines that no abnormal portion is detected in the object 1 (No in S27), the processing ends. In step S27, when the detection device 10 determines that no abnormal portion is detected in the object 1 (No), it may output "no abnormality" by voice or text, or may output a symbol such as ◯. may be output, or a green lamp may be turned on to inform the user that there is no abnormality.

In step S12, prior to the division of blocks, the detection device 10 excludes areas other than at least a part of the object 1 appearing in the image (in other words, areas in the image in which the object 1 does not appear) from the detection targets. However, the region to be detected in the image (hereinafter also referred to as the target region) may be divided into blocks of the same size as the size used as teacher data when learning the machine learning model. This processing is called filtering processing. The filtering process will be specifically described with reference to FIG.

FIG. 10 is a diagram illustrating an example of filtering processing. As shown in (a) of FIG. 10, the detection device 10 extracts a target region representing a detection target (in other words, target object 1) appearing in an image.

Next, as shown in (b) of FIG. 10 , the detection device 10 paints out a region in which the target object 1 is not shown in the image (in other words, a non-target region), thereby removing the non-target region from the detection target. exclude.

Next, as shown in (c) of FIG. 10, the detection device 10 divides the target region into blocks of the same size as the blocks used as teacher data when learning the machine learning model.

Next, the detection device 10 inputs the block-divided image (for example, (c) in FIG. 10) to the trained model, so that the pixels representing the elliptical object are shown in black in each block in the image. An output result (for example, (d) in FIG. 10) is obtained.

[Effects, etc.]
Next, effects of the detection system 100, the detection device 10, the detection method, and the program according to the present embodiment will be described.

As described above, the detection method according to the present embodiment acquires an image showing at least part of an object including an elliptical object, and uses the acquired image as teacher data when learning a machine learning model. The number of pixels representing an elliptical object in each block obtained by dividing the block into blocks of the same size as the block size obtained by dividing the block into blocks and inputting the block-divided image into a trained model, which is a trained machine learning model. Based on the number, an abnormal location of the object is detected, and in detecting the abnormal location, if the number of pixels output from the trained model is greater than the first threshold and less than the second threshold, the block If it is determined that the part of the object corresponding to It is determined that the part of the object is an abnormal part.

A device that executes such a detection method uses a trained model to easily detect pixels representing a plurality of elliptical objects in divided images (so-called blocks) obtained by dividing an image that captures at least a part of an object into blocks. And it can be detected with high accuracy. Therefore, the device that executes the detection method can easily and accurately detect an abnormal portion of an object including an elliptical object.

Further, in the detection method according to the present embodiment, when an abnormal portion is detected in the object, a marking image is generated by marking the abnormal portion in the image, and the generated marking image is output as a determination result. may

An apparatus that executes such a detection method can output an abnormal portion of an object to a user in an easily recognizable form.

Further, in the detection method according to the present embodiment, prior to block division, areas other than at least a part of the object appearing in the image are excluded from the detection target, and the detection target area in the image is used for learning of the machine learning model. It may sometimes be divided into blocks of the same size as the size used as teacher data.

An apparatus that executes such a detection method detects only an area in which an object is captured, so that an elliptical object can be efficiently detected. Therefore, the device that executes the detection method can more efficiently detect the abnormal portion of the object.

Further, in the detection method according to this embodiment, the machine learning model may be a convolutional neural network.

A device that executes such a detection method can efficiently perform a convolution operation using an image as an input.

Further, in the detection method according to the present embodiment, the teacher data used for learning the machine learning model includes a plurality of first blocks obtained by dividing a first image in which at least a part of the object is shown, and the first blocks. First data composed of annotations indicating areas of each of a plurality of elliptical objects included in one block, and the first image is re-divided into the above size by shifting the first block within a range not exceeding the first block. It may include second data composed of a plurality of second blocks and annotations indicating regions of the plurality of elliptical objects included in the second blocks.

A device that executes such a detection method can generate a large amount of teacher data from a small number of images, thereby improving the learning effect of the machine learning model and improving the detection accuracy of an elliptical object by the machine learning model. Therefore, the device that executes the detection method can more easily and accurately detect the elliptical object included in the object.

(Other embodiments)
As described above, the detection method and program according to the present disclosure have been described based on the above embodiments, but the present disclosure is not limited to these embodiments. As long as they do not deviate from the gist of the present disclosure, various modifications that can be conceived by those skilled in the art may be included in the scope of the present disclosure.

Further, each unit included in the detection system, detection device, detection method, and program according to the above embodiments is typically implemented as an LSI, which is an integrated circuit. These may be made into one chip individually, or may be made into one chip so as to include part or all of them.

Further, circuit integration is not limited to LSIs, and may be realized by dedicated circuits or general-purpose processors. An FPGA (Field Programmable Gate Array) that can be programmed after the LSI is manufactured, or a reconfigurable processor that can reconfigure connections and settings of circuit cells inside the LSI may be used.

In each of the above-described embodiments, each component may be configured by dedicated hardware, or realized by executing a software program suitable for each component. Each component may be implemented by a program execution unit such as a CPU or processor reading and executing a software program recorded in a storage medium such as a hard disk or semiconductor memory.

In addition, the numbers used above are all examples for specifically describing the present disclosure, and the embodiments of the present disclosure are not limited to the illustrated numbers.

Also, the division of functional blocks in the block diagram is an example, and a plurality of functional blocks can be realized as one functional block, one functional block can be divided into a plurality of functional blocks, and some functions can be moved to other functional blocks. may Moreover, single hardware or software may process the functions of a plurality of functional blocks having similar functions in parallel or in a time-sharing manner.

Also, the order in which each step in the flowchart is executed is for illustrative purposes in order to specifically describe the present disclosure, and orders other than the above may be used. Also, some of the above steps may be executed concurrently (in parallel) with other steps.

It should be noted that any form obtained by applying various modifications that a person skilled in the art can come up with to the above-described embodiments, or by arbitrarily combining the constituent elements and functions in each embodiment without departing from the scope of the present disclosure. Any form is also included in the present disclosure.

The detection apparatus and detection method according to the present disclosure will be specifically described below in Examples, but the present disclosure is not limited to the following Examples.

In the following, DeepCrack Network was used as a machine learning model to detect an abnormal portion of the texture of a cloth product. An image of a fabric product with no abnormalities was used.

Cloth products are composed of a plurality of warp threads and a plurality of weft threads, and the warp threads are woven above and below the weft threads to create a plurality of textures. Each texture looks like rice grains (so-called elliptical shape). These textures appear on the surface of the fabric. However, when the weave collapses, the texture does not look like rice grains when viewed from above. For example, (1) when the warp threads pass through the back side of the weft threads when they should pass through the front side of the weft threads, the warp threads in that portion do not appear on the surface, resulting in a state of no texture. (2) When the warp thread passes through the front side of the weft thread instead of the back side of the weft thread, the weave pattern appears on the front side of the weft thread, resulting in a large weave pattern. In the following, the detection accuracy of an elliptical object in Comparative Example 1 and Example 1 was verified using a cloth product including textures as an elliptical object. In Comparative Example 2 and Example 2, the detection accuracy of the elliptical object in the above-described abnormal portion of the cloth product was verified.

(Comparative example 1)
In Comparative Example 1, a machine learning model trained using teacher data generated by a conventional method was used. The training data of Comparative Example 1 includes a plurality of blocks (hereinafter referred to as first blocks) obtained by dividing an image showing a cloth product (referred to as a first image) into blocks of a predetermined size (hereinafter referred to as first blocks), and a plurality of blocks included in the first blocks. Annotations indicating areas of each texture (hereinafter referred to as first data). The results detected by the detection method of Comparative Example 1 are shown in FIG. 11 is a diagram showing the results of Comparative Example 1 and Example 1. FIG. In FIG. 11, pixels representing ellipses (here, weaves) are shown in black. As shown in FIG. 11, in Comparative Example 1, black pixels were sparse, and the texture detection accuracy was not so good.

(Example 1)
In Example 1, a machine learning model trained using teacher data generated by the method of the present disclosure was used. In Example 1, the same image as in Comparative Example 1 was used as the image used to generate the teacher data, but the method of generating the teacher data was different. Therefore, more learning image data than Comparative Example 1 could be generated from one image. The training data of Example 1 includes a plurality of blocks (so-called first blocks) obtained by dividing an image showing a cloth product (so-called first image) into blocks of the same size as those of Comparative Example 1 (so-called first blocks), and The first data composed of annotations indicating the areas of each of the plurality of textures, and the first block in the first image is shifted vertically, horizontally, or diagonally within a range not exceeding the first block, and Comparative Example 1 and It includes second data composed of a plurality of blocks subdivided into the same size (referred to as second blocks) and annotations indicating regions of each of the plurality of meshes included in the second blocks. The results detected by the detection method of Example 1 are shown in FIG.

As shown in FIG. 11 , black pixels were seen densely, and it was confirmed that the texture detection accuracy was improved as compared with Comparative Example 1.

(Comparative example 2)
Comparative Example 2 was performed in the same manner as Comparative Example 1, except that an image including an abnormal portion of the cloth product was input to the machine learning model to detect the abnormality. The results of detection by the detection method of Comparative Example 2 are shown in FIG. 12 is a diagram showing the results of Comparative Example 2 and Example 2. FIG. As shown in FIG. 12, the left side of the input image includes an abnormal portion where the yarn is not knitted and protrudes to the surface of the cloth product. In FIG. 12, as in FIG. 11, pixels representing elliptical objects (here, weaves) are shown in black.

As shown in FIG. 12, in Comparative Example 2, the pixels are shown in black as if a plurality of weaves existed in the abnormal portion where the thread protruded. In other words, in Comparative Example 2, it is erroneously detected that the texture is present in a portion where the yarn is protruded and the texture is not present. As a result, since the change in the total number of pixels is small, it is determined that there is no abnormality in the range shown in the input image.

(Example 2)
In Example 2, as in Comparative Example 2, the same procedure as in Example 1 was performed, except that an image including an abnormal portion of the cloth product was input to the machine learning model to detect an abnormality. The results detected by the detection method of Example 2 are shown in FIG.

As shown in FIG. 12, in Example 2, pixels are shown in black along the shape of the protruding thread at the abnormal location where the thread protrudes. As a result, the change in the total number of pixels increases, and it is determined that there is an abnormality in the range shown in the input image.

(summary)
From the results of Comparative Examples 1 and 2 and Examples 1 and 2, when teacher data is generated by the method of the present disclosure, each of the learning image data (so-called divided images) generated from one image is the first block In order to generate the second block, third block, fourth block, etc. by shifting the first block vertically, horizontally, or diagonally within a range not exceeding It has an area that overlaps with one block. By using the divided images generated by re-dividing one image (first image) by shifting the division mesh in this way for machine learning, the learning effect can be easily enhanced, and as a result , it was confirmed that the detection accuracy was improved.

INDUSTRIAL APPLICABILITY According to the present disclosure, an elliptical object can be detected simply and accurately, and therefore, for example, an abnormal portion of an object including an elliptical object can be easily and accurately detected. Therefore, the present disclosure can be used for quality inspection of products in various fields such as foods, industrial products, or daily necessities.

Reference Signs List 1 target object 10 detection device 11 communication unit 12 control unit 12a acquisition unit 12b image processing unit 12c determination unit 12d output unit 13 storage unit 14 learning unit 15 display unit 16 operation reception unit 20 illumination device 30 imaging device 40 transport device 100 detection system

Claims (6)

Acquiring an image showing at least part of an object including an elliptical object,
dividing the obtained image into blocks of the same size as the block size used for the teacher data when learning the machine learning model;
An abnormality of the object based on the number of pixels representing the elliptical object in each of the blocks obtained by inputting the block-divided image into a learned model that is the machine learning model that has been learned. detect the point,
In detecting the abnormal location,
when the number of pixels output from the trained model is greater than a first threshold and less than a second threshold, determining that the location of the object corresponding to the block is not the abnormal location;
When the number of pixels output from the trained model is equal to or less than the first threshold or equal to or greater than the second threshold, the part of the object corresponding to the block is determined to be the abnormal part. Determine detection method. Further, when the abnormal point is detected in the object, generating a marking image in which the abnormal point in the image is marked,
The detection method according to claim 1, wherein the generated marking image is output as a determination result. prior to dividing the blocks, excluding an area other than at least a part of the object appearing in the image from a detection target;
3. The detection method according to claim 1, wherein the area to be detected in the image is divided into blocks of the same size as the teacher data used when learning the machine learning model. The detection method according to any one of claims 1 to 3, wherein the machine learning model is a convolutional neural network. The teacher data used for learning the machine learning model is
First data composed of a plurality of first blocks obtained by dividing a first image showing at least part of an object into the size, and annotations indicating regions of each of the plurality of elliptical objects included in the first blocks. When,
a plurality of second blocks obtained by shifting the first block in the first image and redividing the first block into the size within a range not exceeding the first block; The detection method according to any one of claims 1 to 4, comprising: second data composed of annotations indicating A program for causing a computer to execute the detection method according to any one of claims 1 to 5.

Images