Apparatus, method, and system for extracting target area

Abstract

A target region extraction device according to the present invention comprises: a model storage unit which stores a model of a neural network trained using a plurality of image datasets for learning each obtained by adding a target region as annotation to the image data on a product, the target region being a region to be inspected in the image data; a data acquisition unit which acquires the image data acquired by imaging an exterior of the product; a target region extraction unit which uses the model stored in the model storage unit to extract the target region from the image data, the target region being a region to be inspected in said image data; and an output unit which outputs the target region which is to be inspected in the image data and is extracted by the target region extraction unit.

Description

[Technical Field]

[0001] The present invention relates to an apparatus, method and system for extracting an object region. [Background technology]

[0002] In manufacturing sites such as factories, products manufactured using industrial machines such as machining centers and robots are inspected (see, for example, Patent Document 1). When inspecting products for defects such as scratches, it is extremely important as pre-inspection processing to correctly determine and detect the target inspection area (target pixel area) in an image captured by an imaging device such as a camera. If this area detection is incorrect, a defect may be determined based on image information outside the area, or conversely, a defect may be omitted from the determination area, resulting in an oversight.

[0003] Area detection methods used include defining an area as a region where a set of consecutive pixels with a certain brightness or higher exists, or defining an area by pattern matching after defining the shape pattern of the inspection range in advance. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2018-194542 Summary of the Invention [Problem to be solved by the invention]

[0005] However, this method had problems such as large defects being excluded from the inspection area, and areas being misdetected when the shape of the target object is varied, such as castings. Therefore, a technology for detecting the inspection range with higher accuracy is desired. [Means for solving the problem]

[0006] The target area extraction device according to the present disclosure solves the above problem by extracting areas from the target inspection image using a deep learning device that has been trained in advance using a neural network, thereby extracting the inspection target area with greater accuracy.

[0007] One aspect of the present disclosure is a target area extraction device that extracts a target area for inspection from image data of an external appearance of a product, the device including: a model storage unit that stores a neural network model trained using a plurality of learning image data sets in which target areas for inspection in the image data are annotated for the image data of the product; a data acquisition unit that acquires image data of the external appearance of the product; and a target area extraction unit that uses the model stored in the model storage unit to extract a target area for inspection in the image data from the image data. a target area narrowing unit that narrows down the target area of ​​inspection by performing classical image processing on the image data and the target area of ​​inspection in the image data extracted by the target area extraction unit; The aforementioned The target area narrowing section narrowed down the search results. and an output unit that outputs the inspection target area in the image data.

[0009] Another aspect of the present disclosure is a computer-executed target area extraction method for extracting an inspection target area from image data of an external appearance of a product, the method including the steps of: acquiring image data of the external appearance of the product by a computer; and extracting an inspection target area from the image data of the product using a model stored in a model storage unit that stores a neural network model trained using a plurality of training image data sets to which inspection target areas in the image data are annotated; narrowing down the inspection target area by performing classical image processing on the image data and the extracted inspection target area in the image data; and outputting a target area for inspection in the image data.

[0011] Another aspect of the present disclosure is an appearance inspection system that extracts an inspection target area from image data capturing the appearance of a product and detects defects in the extracted inspection target area, the appearance inspection system including a robot that holds an image sensor, a target area extraction device, and a defect detection unit that detects defective areas in the inspection target area in the image data extracted by the target area extraction device. [Effects of the Invention]

[0012] According to one aspect of the present disclosure, it is possible to detect an inspection range from a target image with higher accuracy without incurring large costs. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a hardware configuration diagram of a target region extraction device according to an embodiment. [Figure 2] 1 is a block diagram showing functions of a target region extraction device according to a first embodiment. [Figure 3] FIG. 10 is a block diagram showing the functions of a target region extraction device according to a second embodiment. [Figure 4] FIG. 10 is a diagram showing a flow of processing for extracting and narrowing down an inspection target area from image data. [Figure 5] FIG. 11 is a block diagram showing the functions of a target region extraction device according to a third embodiment. [Figure 6] FIG. 10 is a block diagram showing the functions of a target region extraction device according to a fourth embodiment. [Figure 7] FIG. 11 is a block diagram showing the functions of a target region extraction device according to a fifth embodiment. [Figure 8] FIG. 1 is a diagram illustrating an example of a robot used for visual inspection. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a schematic hardware configuration diagram showing the main parts of a target region extraction device according to one embodiment of the present invention. The target region extraction device 1 of the present invention can be implemented as a control device that controls industrial machinery, such as a robot for visual inspection. The target region extraction device 1 of the present invention can also be implemented on a computer such as a personal computer attached to a control device that controls industrial machinery for visual inspection, or a personal computer, cell computer, fog computer 6, or cloud server 7 connected to the control device via a wired or wireless network. This embodiment shows an example in which the target region extraction device 1 is implemented on a personal computer connected via a network to a control device that controls industrial machinery.

[0015] The CPU 11 provided in the object region extraction device 1 according to this embodiment is a processor that provides overall control of the object region extraction device 1. The CPU 11 reads a system program stored in the ROM 12 via the bus 22, and controls the entire object region extraction device 1 in accordance with the system program. The RAM 13 temporarily stores temporary calculation data, display data, and various data input from outside.

[0016] The nonvolatile memory 14 is configured, for example, by a battery-backed memory or SSD (Solid State Drive) (not shown), and maintains its stored state even when the power to the target area extraction device 1 is turned off. The nonvolatile memory 14 stores programs and data read from an external device 72 via the interface 15, programs and data input via the input device 71, programs and data acquired from the industrial machine 3, and the like. The data stored in the nonvolatile memory 14 may be expanded into the RAM 13 during execution / use. In addition, various system programs, such as known analysis programs, are written in the ROM 12 in advance.

[0017] The interface 15 is an interface for connecting the CPU 11 of the target area extraction device 1 to an external device 72 such as a USB device. For example, pre-stored control programs and data related to the operation of each industrial machine 3 can be read from the external device 72. Furthermore, the control programs and setting data edited within the target area extraction device 1 can be stored in external storage means via the external device 72.

[0018] The interface 20 is an interface for connecting the CPU 11 of the target area extraction device 1 to a wired or wireless network 5. The network 5 is connected to industrial machines 3, fog computers 6, cloud servers 7, etc., and exchanges data with the target area extraction device 1.

[0019] The display device 70 displays various data loaded into memory, data obtained as a result of executing programs, etc., output via the interface 17. An input device 71, which is comprised of a keyboard, pointing device, etc., passes instructions, data, etc. based on operations by an operator to the CPU 11 via the interface 18.

[0020] The industrial machine 3 is a machine used to perform visual inspections of products processed by machine tools, injection molding machines, etc. The industrial machine 3 is equipped with an image sensor 4 that captures at least an image of the product's appearance. An example of the industrial machine 3 is a robot holding the image sensor 4, as shown in FIG. 8. The target area extraction device 1 acquires image data showing the appearance of the product captured by the image sensor 4 from the industrial machine 3 via the network 5 and the interface 20. The acquired image data is stored in the RAM 13 through the non-volatile memory 14 and processed by the CPU 11.

[0021] Interface 21 is an interface for connecting CPU 11 and machine learning device 100. Machine learning device 100 includes processor 101 that controls the entire machine learning device 100, ROM 102 that stores system programs and the like, RAM 103 that provides temporary storage for each process related to machine learning, and non-volatile memory 104 that stores models and the like. Machine learning device 100 can observe each piece of information that can be obtained by target area extraction device 1 via interface 21. In addition, target area extraction device 1 obtains processing results output from machine learning device 100 via interface 21, and stores or displays the obtained results, or transmits them to another device via network 5 or the like.

[0022] 2 is a schematic block diagram showing functions provided in the object region extraction device 1 according to the first embodiment of the present invention. Each function provided in the object region extraction device 1 according to this embodiment is realized by the CPU 11 of the object region extraction device 1 and the processor 101 of the machine learning device 100 shown in FIG. 1 executing a system program and controlling the operation of each part of the object region extraction device 1 and the machine learning device 100.

[0023] The target area extraction device 1 of this embodiment includes a data acquisition unit 110, a target area extraction unit 120, and an output unit 140. In addition, a model storage unit 210 is provided in advance on the RAM 103 to the nonvolatile memory 104 of the machine learning device 100 as an area for storing a model that has learned the relationship between image data of an image of a product and data indicating a target area for inspection in the image data.

[0024] The data acquisition unit 110 may acquire image data of a product captured by the imaging sensor 4 provided in the industrial machine 3. The data acquisition unit 110 may also acquire image data acquired and stored by an external device 72, the fog computer 6, the cloud server 7, etc. The data acquisition unit 110 outputs the acquired image data to the target area extraction unit 120.

[0025] The target region extraction unit 120 uses a deep learning model to extract a target region to be inspected in the image data input from the data acquisition unit 110. The deep learning model uses a multilayer neural network model trained on the image data using a plurality of training image data sets that annotate the image data with data indicating the target region to be inspected in the image data. The target region extraction unit 120 outputs data indicating the target region to be inspected in the extracted image data to the output unit 140.

[0026] The output unit 140 displays data indicating the target area of ​​the image data input by the target area extraction unit 120 on the display device 70. The output unit 140 may display and output the data indicating the target area alongside the image data. The output unit 140 may also transmit and output the data indicating the target area to the industrial machine 3 from which the image data was acquired. In this case, the data indicating the target area is used by the industrial machine 3 to determine an inspection range in the image data. Furthermore, the output unit 140 may transmit and output the data indicating the target area to a higher-level computer such as the fog computer 6 or the cloud server 7 via the network 5. The data may also be output to a log recording area previously provided on the non-volatile memory 14 or the like.

[0027] By using the object area extraction device 1 according to this embodiment, which has the above-described configuration, it is possible to detect the inspection area from the target image with higher accuracy without incurring significant costs. Conventional methods have devised image data acquisition methods, such as adjusting exposure time in advance to facilitate extraction of the inspection area. This is done to reduce the visibility of machined surface features (e.g., cutter marks) that tend to be erroneously detected as outside the inspection area. This has resulted in extra work for the operator during inspection. The object area extraction device 1 according to this embodiment uses a multilayer neural network, which can learn to include machined surface features that tend to be erroneously detected by conventional methods within the inspection area. Therefore, compared to conventional methods, this method requires less effort in acquiring image data and is expected to achieve greater accuracy.

[0028] 3 is a schematic block diagram showing functions of an object region extraction device 1 according to a second embodiment of the present invention. Each function of the object region extraction device 1 according to this embodiment is realized by the CPU 11 of the object region extraction device 1 and the processor 101 of the machine learning device 100 shown in FIG. 1 executing a system program and controlling the operation of each part of the object region extraction device 1 and the machine learning device 100.

[0029] The target area extraction device 1 of this embodiment further includes a target area narrowing-down unit 130 in addition to a data acquisition unit 110, a target area extraction unit 120, and an output unit 140. Furthermore, a model storage unit 210 is provided in advance on the RAM 103 to the non-volatile memory 104 of the machine learning device 100 as an area for storing a model that has learned the relationship between image data of an image of a product and data indicating a target area for inspection in the image data.

[0030] The data acquisition unit 110 and the target region extraction unit 120 according to this embodiment have the same functions as those described in the first embodiment.

[0031] The target area narrowing unit 130 further narrows down the scope of the target area in the image data extracted by the target area extraction unit 120 to a target area for inspection by using a target area detection method based on classical image processing techniques such as focus determination and area determination. In this specification, classical image processing refers to image processing methods using techniques such as edge detection by binarization, shape pattern matching, and Fourier transform, which were used before the introduction of machine learning. The target area narrowing unit 130 may, for example, perform focus determination on the scope of the target area in the image data using known Laplacian differentiation or fast Fourier transform, and detect the in-focus portion as the target area for inspection. Furthermore, areas of the extracted target area that are smaller than a predetermined area may be excluded from the target area. The target area narrowing unit 130 outputs data indicating the detected target area to the output unit 140.

[0032] The output unit 140 displays data indicating the target area of ​​the image data input by the target area narrowing unit 130 on the display device 70. The output unit 140 may display and output the data indicating the target area alongside the image data. The output unit 140 may also transmit and output the data indicating the target area to the industrial machine 3 from which the image data was acquired. In this case, the data indicating the target area is used by the industrial machine 3 to determine the inspection range of the image data. Furthermore, the output unit 140 may transmit and output the data indicating the target area to a higher-level computer such as the fog computer 6 or the cloud server 7 via the network 5. The data may also be output to a log recording area previously provided on the non-volatile memory 14 or the like.

[0033] FIG. 4 shows an example in which the object region extraction device 1 according to this embodiment detects an inspection target region from image data. When the object region extraction device 1 according to this embodiment acquires product image data 301, the object region extraction unit 120 first extracts the inspection target region using a trained model of a multilayer neural network. Data 303 is the inspection target region extracted from the image data by the object region extraction unit 120 using a machine learning device. In data 303, the portion extracted as the inspection target region is shown in white. Then, the object region narrowing unit 130 performs image processing such as focus determination and area determination on the range of the image data 301 that is identified as the inspection target region in data 303, thereby further narrowing down the inspection target region. Data 305 is data indicating the inspection target region in the narrowed-down image data detected by the object region narrowing unit 130 through image processing. The object region extraction device 1 performs output processing on this data 305, regarding it as the inspection target region in the final image data.

[0034] By using the object area extraction device 1 according to this embodiment, which has the above-described configuration, it is possible to detect the inspection area from the object image with high accuracy without incurring significant costs. Objects such as products often have complex three-dimensional shapes, and it is necessary to extract a specific area from within the object as the inspection area. The characteristics of the desired area vary widely, and extraction of the detailed condition of the object's surface may be particularly necessary. When using area extraction using machine learning techniques, increasing the diversity of input data can result in a decrease in accuracy and increased training and data collection costs. Furthermore, given the individual variations in workpieces and the need for moving cameras for imaging using robots, it is necessary to consider environments in which imaging conditions fluctuate. For example, an image may be captured in which the entire image is out of focus. In such cases, the entire image must be correctly excluded from flaw detection. This focus detection can also be trained using a deep learning device such as a neural network if sufficient diverse data is available. However, considering realistic data collection costs, it is necessary to deal with cases where training data is limited, and it is difficult to handle everything with a neural network. Therefore, in the object area extraction device 1 according to this embodiment, the results of the inspection object area extraction by the machine learning device are further processed by a classical method to target only the area that is in focus, which has the effects of stabilizing area detection, improving accuracy, and reducing data collection costs.

[0035] 5 is a schematic block diagram showing the functions of a target area extraction device 1 according to a third embodiment of the present invention. In addition to the functions of the target area extraction device 1 according to the first and second embodiments, the target area extraction device 1 according to this embodiment further includes a function for detecting defective portions from image data during visual inspection. The functions of the target area extraction device 1 according to this embodiment are realized by the CPU 11 of the target area extraction device 1 and the processor 101 of the machine learning device 100 shown in FIG. 1 executing a system program and controlling the operation of each part of the target area extraction device 1 and the machine learning device 100.

[0036] The functions of the data acquisition unit 110, the target area extraction unit 120, and the target area narrowing unit 130 according to this embodiment are similar to those of the first and second embodiments. The defect detection unit 150 detects defects in the product image data within the range included in the inspection target area detected by the target area narrowing unit 130. The defect detection unit 150 may use a known, classical image processing defect detection method. For example, a method of detecting scratches or stains on the product surface by binarizing or edge detecting the inspection target area of ​​the image data can be used. Defects can also be detected by applying a blur filter, blob analysis, or contrast conversion. Furthermore, known techniques for detecting defects in image data using a multilayer neural network may be used. These defect detection methods are already known, so a description of them will be omitted here. The defect detection unit 150 outputs the presence or absence of defects in the image data to the output unit 140. Furthermore, if a defect is detected, the defect detection unit 150 also outputs the area of ​​the defect in the image data to the output unit 140.

[0037] The output unit 140 according to this embodiment displays on the display device 70 the presence or absence of a defect in the image data input by the defect detection unit 150. The output unit 140 may also display and output the image data, data indicating the target area, and data indicating the area of ​​the detected defect. The output unit 140 may also transmit and output these data to the industrial machine 3 from which the image data was acquired. Furthermore, the output unit 140 may also transmit and output the data indicating the target area to a higher-level computer such as the fog computer 6 or the cloud server 7 via the network 5. The output unit 140 may also output the data to a log recording area provided in advance on the non-volatile memory 14 or the like.

[0038] By using the object area extraction device 1 according to this embodiment having the above configuration, it is possible to detect an inspection range from an object image with higher accuracy without incurring large costs. Furthermore, it is possible to detect defective parts from an appropriate inspection object area in image data. When a defective part is detected, it is possible to cause the industrial machine 2 to perform an operation such as removing the defective product from the line in accordance with the detection result. In this way, by linking the object area extraction device 1 and the industrial machine 2, it is also possible to construct an appearance inspection system that inspects products with higher accuracy.

[0039] Fig. 6 is a schematic block diagram showing functions provided in an object region extraction device 1 according to a fourth embodiment of the present invention. The object region extraction device 1 according to this embodiment has functions for learning a model used by the object region extraction devices 1 according to the first to third embodiments to extract an object region. The functions provided in the object region extraction device 1 according to this embodiment are realized by the CPU 11 of the object region extraction device 1 and the processor 101 of the machine learning device 100 shown in Fig. 1 executing a system program and controlling the operation of each part of the object region extraction device 1 and the machine learning device 100.

[0040] The target area extraction device 1 of this embodiment includes a data acquisition unit 110, an annotation assignment unit 160, and a learning unit 180. Furthermore, a data storage unit 220, which is an area for storing image data of an image of a product acquired by the data acquisition unit 110, is prepared in advance on the RAM 13 to the nonvolatile memory 14 of the target area extraction device 1. Furthermore, a model storage unit 210, which is an area for storing a learned model relating to the relationship between the image data of an image of a product and data indicating an inspection target area in the image data, is prepared in advance on the RAM 103 to the nonvolatile memory 104 of the machine learning device 100.

[0041] The data acquisition unit 110 according to this embodiment may acquire image data of a product captured by the imaging sensor 4 provided in the industrial machine 3. The data acquisition unit 110 may also acquire image data acquired and stored by an external device 72, the fog computer 6, the cloud server 7, or the like. The data acquisition unit 110 stores the acquired image data in the data storage unit 220.

[0042] The annotation assigning unit 160 assigns annotations to the product image data acquired by the data acquiring unit 110. In the present invention, the annotation assigned to the product image data is data indicating an area to be inspected within the image data. This data may be, for example, mask data that masks the area outside the area to be inspected within the image data. The annotation assigning unit 160 may acquire annotations manually, for example, by an operator. In this case, the annotation assigning unit 160 displays the image data on the display device 70. The operator then designates the area to be inspected within the image data while referring to the image data. The annotation assigning unit 160 creates data indicating the area to be inspected that has been input. The annotation assigning unit 160 then associates the created data with the image data to create a learning image dataset, which is stored in the data storage unit 220.

[0043] The learning unit 180 generates a model that learns features of an inspection target area in image data for image data based on a learning image dataset in which image data acquired by the data acquisition unit 110 is associated with annotations assigned to the image data. The generated model is then stored in the model storage unit 210. The machine learning performed by the learning unit 180 is a well-known supervised learning method. An example of the model is a multilayer neural network. The model generated by the learning unit 180 can be used to estimate an inspection target area in image data of an image of a product from the image data. Note that a method for learning features of image data using a multilayer neural network as a model is already well known, and therefore a detailed description thereof will be omitted in this specification.

[0044] The target area extraction device 1 according to this embodiment having the above configuration can generate a model to be used in the target area extraction device 1 according to the first to third embodiments. By using the model generated by the target area extraction device 1, it becomes possible to detect the inspection area from the target image with higher accuracy without incurring large costs.

[0045] 7 is a schematic block diagram showing functions provided in an object region extraction device 1 according to a fifth embodiment of the present invention. The object region extraction device 1 according to this embodiment has functions for learning a model used by the object region extraction devices 1 according to the first to third embodiments to extract an object region. The functions provided in the object region extraction device 1 according to this embodiment are realized by the CPU 11 of the object region extraction device 1 and the processor 101 of the machine learning device 100 shown in FIG. 1 executing a system program and controlling the operation of each part of the object region extraction device 1 and the machine learning device 100.

[0046] The target area extraction device 1 of this embodiment further includes a data augmentation unit 190 in addition to the data acquisition unit 110, annotation assignment unit 160, and learning unit 180. A data storage unit 220, which is an area for storing image data of an image of a product acquired by the data acquisition unit 110, is provided in advance on the RAM 13 to nonvolatile memory 14 of the target area extraction device 1. A model storage unit 210, which is an area for storing a learned model relating to the relationship between the image data of an image of a product and data indicating an inspection target area in the image data, is provided in advance on the RAM 103 to nonvolatile memory 104 of the machine learning device 100.

[0047] The functions of the data acquisition unit 110 and the annotation assignment unit 160 according to this embodiment are the same as those described in the fourth embodiment. The data augmentation unit 190 generates multiple image data by performing image processing (image processing) on ​​the image data acquired by the data acquisition unit 110, such as brightness correction (gamma correction), scaling, rotation, and compositing images corresponding to defects. The data augmentation unit 190 also generates a training image dataset by associating the annotations assigned to the original image data by the annotation assignment unit 160 with each of the processed image data. The training image dataset in which the original image data and annotations are associated, as well as the created training image dataset, are then output to the training unit 180. If the processing of the image data involves changing the position or size of image data transferred within the image data, the data augmentation unit 190 also performs similar processing on the annotations assigned to the processed image data. The data augmentation unit 190 augments a single training image dataset to generate multiple training image datasets that take into account environmental changes, etc.

[0048] The learning unit 180 generates a model that learns the features of the inspection target area in the image data for the image data based on the training image data set augmented by the data augmentation unit 190. The generated model is then stored in the model storage unit 210. The machine learning performed by the learning unit 180 is known supervised learning. An example of the model is a multilayer neural network. The model generated by the learning unit 180 can be used to estimate the inspection target area in the image data from the image data of an image of a product. Note that since the method of learning the features of image data using a multilayer neural network as a model is already well known, a detailed description thereof will be omitted in this specification.

[0049] The object region extraction device 1 according to this embodiment having the above configuration can generate a model to be used in the object region extraction device 1 according to the first to third embodiments. By using the model generated by the object region extraction device 1, it becomes possible to detect the inspection area from the object image with higher accuracy without incurring large costs. Furthermore, because the data augmentation unit 190 augments the learning image dataset, it becomes possible to efficiently learn the model using a small amount of image data.

[0050] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described examples of the embodiments, and can be embodied in various forms by making appropriate modifications. [Explanation of symbols]

[0051] 1. Target area extraction device 3. Industrial machinery 4. Image sensor 5. Network 6. Fog Computer 7. Cloud Server 11 CPU 12 ROM 13 RAM 14 Non-volatile memory 15 Interface 17,18,20,21 Interface 22 Bus 70 Display device 71 Input Device 72 External equipment 100 Machine Learning Machines 101 processors 102 ROM 103 RAM 104 Non-volatile memory 110 Data Acquisition Unit 120 Target region extraction unit 130 Target Area Narrowing Section 140 Output section 150 Defective part detection unit 160 Annotation section 180 Learning Department 190 Data Augmentation Department 210 Model Memory Unit 220 Data storage unit

Claims

1. A target area extraction device that extracts a target area for inspection from image data of an external appearance of a product, a model storage unit that stores a neural network model trained using a plurality of training image data sets in which inspection target areas in product image data are annotated; and a data acquisition unit that acquires image data of an external appearance of a product; an object region extraction unit that extracts an object region to be inspected from the image data using the model stored in the model storage unit; a target area narrowing unit that narrows down the target area of ​​inspection by performing classical image processing on the image data and the target area of ​​inspection in the image data extracted by the target area extraction unit; an output unit that outputs the inspection target area in the image data narrowed down by the target area narrowing unit; A target region extraction device comprising:

2. The classical image processing performed by the target area narrowing unit is a focus determination process for an image. The object region extraction device according to claim 1 .

3. The focus determination process is a focus determination process using Laplacian derivatives. The object region extraction device according to claim 2 .

4. A computer-implemented target area extraction method for extracting an inspection target area from image data of an external appearance of a product, comprising: The computer acquiring image data of an external appearance of a product; extracting an inspection target area from image data of a product using a model stored in a model storage unit that stores a neural network model trained using a plurality of training image data sets to which inspection target areas in the image data are annotated; narrowing down the target area of ​​inspection by performing classical image processing on the image data and the extracted target area of ​​inspection in the image data; outputting the narrowed down inspection target area in the image data; A method for extracting regions of interest.

5. An appearance inspection system that extracts an inspection target area from image data of an appearance of a product and detects defects in the extracted inspection target area, a robot that holds an image sensor; A target region extraction device according to any one of claims 1 to 3; a defect detection unit that detects a defect in the inspection target area in the image data extracted by the target area extraction device; A visual inspection system equipped with

Images