Image inspection device, image dividing method, and image dividing program

Abstract

To properly perform appearance inspection using deep learning.SOLUTION: An image inspection device according to one aspect of the present disclosure is an image inspection device that performs image inspection of an object to be inspected based on image data of an image captured by an image capturing unit that captures an appearance of the object to be inspected, and includes a compression unit, a determination unit, and a generation unit. The compression unit compresses the image data by replacing the object to be inspected depicted in the image data with a different representation method. The determination unit determines division positions from a compressed image, which is the image data compressed by the compression unit. The generation unit generates multiple divided images obtained by dividing the image data at the division positions determined by the determination unit.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an image inspection device, an image segmentation method, and an image segmentation program. [Background technology]

[0002] Conventionally, there is known a technique for determining whether a product is good or defective based on image data of the product obtained by photographing the manufactured product in an inspection process. Patent Document 1 discloses a technique for determining defects in a product from image data using artificial intelligence and deep learning.

[0003] The technology described in Patent Document 1 divides image data into multiple resolutions and uses the divided image data to identify defects in products, thereby making it possible to identify defects of different sizes. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-90633 Summary of the Invention [Problem to be solved by the invention]

[0005] In the technology described in Patent Document 1, as the number of resolution changes increases, the amount of image data after division increases, resulting in an increased number of deep learning processes. As a result, the technology described in Patent Document 1 has the problem of delayed processing time. Furthermore, when high-resolution image data is input into a model created by deep learning, there is a possibility that the memory capacity for temporarily storing the image data will be insufficient, causing the processing to stop.

[0006] The present disclosure provides a technology that enables appropriate visual inspection using deep learning. [Means for solving the problem]

[0007] An image inspection device according to one aspect of the embodiment is an image inspection device that performs image inspection of an object to be inspected based on image data of an image captured by an imaging unit that captures the appearance of the object to be inspected, and includes a compression unit, a determination unit, and a generation unit. The compression unit compresses the image data by replacing the object to be inspected depicted in the image data with a different representation method. The determination unit determines division positions from a compressed image, which is the image data compressed by the compression unit. The generation unit generates a plurality of divided images by dividing the image data at the division positions determined by the determination unit.

[0008] According to one aspect of the embodiment, the division position can be determined using the result of replacing image data with a different representation and compressing it. When the division position is determined in this manner, the inspection object is less likely to be divided at an intermediate position, thereby reducing false detections. Furthermore, since the image data is divided at the position determined in this manner, the size of each divided image data is smaller than the image data before division. If the size of each image data is reduced, for example, when inputting the image data into a model created by deep learning, the learning / classification process can be completed without increasing the memory size for temporarily storing the image data beyond the normal size. Thus, according to one aspect of the embodiment, visual inspection using deep learning in which divided image data is input can be performed appropriately. [Effects of the Invention]

[0009] According to the present disclosure, visual inspection using deep learning can be performed appropriately. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a schematic explanatory diagram of a decision model used in image inspection processing according to an embodiment. [Figure 2] FIG. 2 is a diagram illustrating an outline of the image inspection process according to the embodiment. [Figure 3]FIG. 3 is a block diagram showing an example of the configuration of an image inspection device according to the embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of a feature point image according to the embodiment. [Figure 5] FIG. 5 is a diagram showing an example of the split image display mode according to the embodiment. [Figure 6] FIG. 6 is a diagram showing an example of the whole image display mode according to the embodiment. [Figure 7] FIG. 7 is a flowchart showing an example of information processing by the processing unit of the image inspection device according to the embodiment. [Figure 8] FIG. 8 is a diagram showing another example of a compressed image. [Figure 9] FIG. 9 is a diagram showing an example of the relationship between the frequencies included in the frequency image and the processing. [Figure 10] FIG. 10 is a diagram showing an example of the relationship between the feature point density of a feature point image and processing. [Figure 11] FIG. 11 is a diagram illustrating an example of the hardware configuration of a processing unit in an image inspection device according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] The image inspection device, image segmentation method, and image segmentation program disclosed in the present application will be described below with reference to the accompanying drawings. Note that the present invention is not limited to the following embodiments.

[0012] <1. Overview of image inspection processing> First, a decision model used in image inspection processing according to an embodiment will be described with reference to Figures 1 and 2. Figure 1 is a schematic explanatory diagram of the decision model used in image inspection processing according to an embodiment. Figure 2 is a schematic explanatory diagram of image inspection processing according to an embodiment.

[0013] The image inspection process according to the embodiment is a process for inspecting the appearance of an object to be inspected from image data, which is an image of the object to be inspected taken during an inspection process at a manufacturing site. In the example shown in Fig. 1, the object to be inspected is a substrate, and the image inspection process according to the embodiment detects foreign matter such as metal pieces from the image data of the substrate.

[0014] The image inspection process according to the embodiment detects foreign objects such as metal fragments from image data of an image of a substrate by image processing using a detection model created by deep learning such as deep learning. In image processing using a detection model, the larger the amount of image data input to the detection model, the greater the possibility that the image data cannot be temporarily stored in memory and the processing will stop.

[0015] Therefore, by inputting the divided images obtained by dividing the original image data (hereinafter also referred to as "original image") of the object to be inspected into the detection model, it is possible to reduce the amount of image data input each time, thereby reducing the processing load.

[0016] On the other hand, depending on the division position of the original image, electronic components placed on the board may not be visible in the divided image. If such a divided image is used, there is a risk that the electronic components will be detected as foreign objects. Therefore, when using a divided image, it is necessary to optimize the division position of the original image.

[0017] The process of optimizing the division positions of the original image can be determined by deep learning. However, if the division positions are determined using the original image, there is a risk that the process will stop at the stage of determining the division positions. In other words, if the division positions of the original image are determined by deep learning, there is a risk that the image data will not be able to be temporarily stored in memory at the stage of preprocessing when the inspection object is actually inspected, causing the process to stop.

[0018] Therefore, in the image inspection process according to the embodiment, when determining the division position using deep learning, a compressed image is used in which the inspection object shown in the original image is replaced with a different representation and the original image is compressed. Hereinafter, as a method of replacing with a different representation, a case in which the inspection object shown in the original image is replaced with a feature point will be described. The method of replacing with a different representation is not limited to the method of replacing with a feature point, but may also be a method of replacing with a frequency, which will be described later, or other methods.

[0019] Here, the learning process of a determination model that determines division positions of an original image will be described with reference to Fig. 1. As shown in Fig. 1, the learning process according to the embodiment first compresses the original image (step S1) to generate a compressed image from the original image. Note that the original image here is image data for learning, and is an original image that includes foreign objects such as metal pieces.

[0020] The compressed image is an image in which the inspection object shown in the original image is replaced with a different representation method, and is an image with a smaller data size than the original image. More specifically, the compressed image is, for example, a feature point image that shows the distribution of feature points of the inspection object shown in the original image.

[0021] The compressed image can be generated by subjecting the original image to various image recognition processes such as corner detection and statistical processing.

[0022] Next, in the learning process according to the embodiment, the compressed image is input to a determination model that determines division positions. The determination model may be, for example, a convolutional neural network (CNN). The determination model calculates division positions of the original image from the input compressed image and outputs the calculation results regarding the division positions. In the learning process according to the embodiment, the division positions are determined based on the output results of the determination model (step S2).

[0023] Next, in the learning process according to the embodiment, the original image is divided based on the division positions determined using the compressed image to generate divided images (step S3). Next, in the learning process according to the embodiment, object detection is performed on each divided image using an object detection model configured by CNN or the like (step S4).

[0024] Then, in the learning process according to the embodiment, the result of the object detection is fed back to the determination model (step S5). The information fed back to the determination model is information indicating whether the result of the foreign object detection is correct or incorrect.

[0025] In other words, in the learning process according to the embodiment, if a foreign object is not properly detected, or if an object that is not a foreign object is detected as a foreign object, the parameters of the determination model are appropriately corrected.

[0026] In addition, in the present disclosure, the determination model can determine the resolution of each of the multiple divided images from the compressed image in addition to determining the division position. Therefore, the learning process according to the embodiment can generate a determination model that can optimize the division position of the original image from the compressed image and optimize the resolution of each of the multiple divided images by repeatedly performing these learning processes.

[0027] Then, in the image inspection process according to the embodiment, inspection is performed in an inspection process at an actual manufacturing site by the process shown in Fig. 2. Specifically, in the image inspection process according to the embodiment, similar to the learning process shown in Fig. 1, first, an original image is compressed (step S11) to generate a compressed image from the original image.

[0028] Next, the image inspection process according to the embodiment determines division positions using the determination model (step S12), and divides the original image based on the determined division positions to generate multiple divided images (step S13).Then, the image inspection process according to the embodiment performs object detection using the object detection model for each of the multiple divided images (step S14).

[0029] In this way, the learning process according to the embodiment determines the division positions of the original image using a compressed image obtained by compressing the original image after replacing the inspection object shown in the original image with another representation method.

[0030] Furthermore, the image inspection process according to the embodiment can determine the division position and resolution from a compressed image with a lower data volume than the original image by using a decision model optimized by the learning process according to the embodiment.

[0031] Therefore, according to the image inspection process of the embodiment, it is possible to appropriately perform visual inspection using deep learning.

[0032] <1.2. Example of image inspection device configuration> Fig. 3 is a block diagram showing an example of the configuration of an image inspection device 100 according to an embodiment. The image inspection device 100 shown in Fig. 2 includes a communication unit 10, a storage unit 11, a display unit 12, an operation unit 13, and a processing unit 14. The image inspection device 100 is also connected to a camera 2 and lighting 3 via a communication network N.

[0033] The camera 2 is an imaging unit that captures the appearance of the inspection object. The camera 2 is installed on a production line where the inspection object is produced, and captures the inspection object. The lighting 3 irradiates the inspection object with light when the camera 2 captures the inspection object. Note that the camera 2 may be, for example, a CCD (Charge Coupled Device) camera.

[0034] <1.2.1. Communication Unit 10> The communication unit 10 is realized by, for example, a communication module. The communication module includes, for example, at least one of a cellular module and a LAN (Local Area Network) communication module. The communication unit 10 is connected to a communication network N by wire or wirelessly, and transmits and receives information to and from various devices.

[0035] <1.2.2. Storage section 11> The storage unit 11 is realized by, for example, a semiconductor memory element such as an SSD (Solid State Drive) or a flash memory, or a storage device such as a hard disk or an optical disk.

[0036] The storage unit 11 stores, for example, a determination model that determines the division positions of an original image, a detection model that detects a detection target from image data, an information processing program, etc. The storage unit 11 also stores learning images, which are original images used when learning the determination model.

[0037] <1.2.3.Display section 12> The display unit 12 is, for example, an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence) display.

[0038] <1.2.4.Operation section 13> The operation unit 13 includes, for example, a keyboard, a mouse, a power button, etc. If the display unit 12 is a touch panel display, the operation unit 13 includes a touch panel.

[0039] <1.2.5. Processing Unit 14> 3, the processing unit 14 includes an acquisition unit 20, a compression unit 21, an extraction unit 22, a counting unit 23, a determination unit 24, a generation unit 25, a detection unit 26, a feedback unit 27, and a display control unit 28, and realizes or executes the functions and actions of the information processing described below. Note that the internal configuration of the processing unit 14 is not limited to the configuration shown in FIG. 3, and may be any other configuration that performs the information processing described below.

[0040] <1.2.5.1. Acquisition part 20> The acquisition unit 20 acquires image data, which is data of an image captured by a camera 2 that captures the appearance of an inspection target, via the communication unit 10. The acquisition unit 20 also acquires training images, which are original images used when training a determination model. The training images are prepared in advance by, for example, an administrator or the like.

[0041] <1.2.5.2. Compression section 21> The compression unit 21 compresses the image data by replacing the inspection object shown in the image data with a different representation. For example, the compression unit 21 compresses the image data by extracting feature points of the inspection object from the original image. The feature points can be extracted by image recognition processing such as corner detection.

[0042] Then, the compression unit 21 generates a feature point image showing the distribution of the extracted feature points. Fig. 4 is a diagram showing an example of the feature point image according to the embodiment. As shown in Fig. 4, the compression unit 21 extracts feature points of the inspection target from the original image (step S21). In the example shown in Fig. 4, the extracted feature points are indicated by white circles.

[0043] Next, the compression unit 21 converts the extracted feature points into a feature point image showing the distribution of the feature points (step S22). The feature point image is, for example, a grayscale image, with varying shading depending on the density of the feature points. In this way, the compression unit 21 compresses the original image into an image related to the feature points of the inspection target, thereby reducing the amount of image data to be input to the determination model. Furthermore, as will be described later, the compression unit 21 can also compress the original image by converting the original image into a grayscale image and then performing a discrete cosine transform on the grayscale image.

[0044] <1.2.5.3.Extraction part 22> The extraction unit 22 extracts feature points of the inspection target from the image data. For example, the extraction unit 22 extracts feature points of the inspection target from the original image by image recognition processing such as corner detection, similar to the compression unit 21. Note that the compression unit 21 may also perform the processing by the extraction unit 22.

[0045] <1.2.5.4. Aggregation Unit 23> The counting unit 23 counts the number of feature points extracted by the extraction unit 22 for each predetermined area. For example, the counting unit 23 counts the number of feature points for each area obtained by dividing the original image into a grid pattern. The size of the areas into which the original image is divided may be set in advance by an administrator.

[0046] <1.2.5.5. Determination Unit 24> The determination unit 24 determines division positions from a compressed image, which is image data compressed by the compression unit 21. The determination unit 24 inputs the compressed image to a determination model and determines division positions of the original image from the output result of the determination model. For example, the determination unit 24 determines the arrangement and size of rectangular areas used when dividing the original image as division positions.

[0047] Furthermore, when the determination model is a model that determines the resolution of each divided image, the determination unit 24 determines the resolution of each divided image from the output result of the determination model. The determination unit 24 determines the resolution for each rectangular area determined as the division position. Note that, as will be described later, the determination unit 24 may determine the resolution for each area based on the counting result by the counting unit 23.

[0048] Furthermore, the determination unit 24 determines a background region in the original image. Specifically, the determination unit 24 determines the background region based on the counting result by the counting unit 23. The background region is a region in the original image where electronic components and the like are not shown and the circuit board itself is shown as the background. In other words, the background region is a region where image inspection does not need to be performed.

[0049] For example, the determining unit 24 determines whether each region is a background region based on the number of feature points for each region tallied by the tallying unit 23, i.e., the density of feature points for each region. For example, the determining unit 24 determines a region in which the density of feature points is equal to or less than a threshold as a background region. Then, the determining unit 24 excludes the background region from the target of image inspection of the inspection target. This allows object detection in the background region by the detecting unit 26, which will be described later, to be skipped, thereby further shortening the processing time.

[0050] <1.2.5.6.Generation part 25> The generation unit 25 generates a plurality of divided images by dividing the image data at the division positions determined by the determination unit 24. The generation unit 25 generates a plurality of divided images by dividing the image data at the division positions determined by the determination unit 24 and dividing them into respective parts.

[0051] Furthermore, when the resolution of each of the plurality of divided images is determined by the determination unit 24, the generation unit 25 generates divided images in which the resolution of each of the plurality of divided images is the resolution determined by the determination unit 24.

[0052] <1.2.5.7.Detection Unit 26> The detection unit 26 is a processing unit that performs image inspection of the inspection target. In this embodiment, the detection unit 26 detects foreign objects appearing in the original image. The detection unit 26 detects foreign objects appearing in the original image using a detection model such as CNN.

[0053] The detection unit 26 detects a foreign object using a detection model from each of the multiple divided images generated by the generation unit 25. When the detection unit 26 detects a foreign object, it notifies an administrator of an alert via the display unit 12, for example.

[0054] Furthermore, the detection unit 26 outputs the foreign object detection result to the feedback unit 27 during the learning stage of the decision model.

[0055] <1.2.5.8.Feedback Section 27> The feedback unit 27 feeds back to the determination model the detection result of the inspection object by the detection unit 26. Specifically, in the learning stage of the determination model, the feedback unit 27 compares the detection result of the detection unit 26 with ground truth data, which is data on foreign objects actually appearing in the original image.

[0056] Then, the feedback section 27 generates feedback information indicating whether the detection result of the detection section 26 matches the correct data or whether the detection result of the detection section 26 does not match the correct data, and feeds this information back to the determination model.

[0057] As a result, if the detection result of the detection unit 26 does not match the correct data, the parameters of the determination model are updated and optimized so that the detection result of the detection unit 26 matches the correct data.

[0058] <1.2.5.9. Display control unit 28> The display control unit 28 switches between the image data and the divided images in response to a user operation and displays them on the display unit 12. Specifically, the display control unit 28 switches between a divided image display mode and a whole image display mode and displays them on the display unit 12.

[0059] Fig. 5 is a diagram showing an example of a split image display mode according to the embodiment. As shown in Fig. 5, the split image display mode is a display mode in which a split image is displayed. In the split image display mode, the display control unit 28 switches and displays the split image to be displayed in response to an operation on the operation unit 13.

[0060] The administrator then checks the divided image displayed on the display unit 12 and selects the start button displayed on the display unit 12, which starts the detection process for the image by the detection unit 26. When the detection process is complete, the portion that has been detected is surrounded by a frame. Note that Figure 5 shows the state after the detection process is completed.

[0061] The detection results are displayed on the indicator in the upper right corner. In Figure 5, there were no problems with the detection results, so the indicator shows "OK." If a foreign object or other problem is detected, the indicator will show "NG." The administrator can switch the displayed split image by operating the GUI that shows the two switching buttons located below the partial image.

[0062] 6 is a diagram showing an example of the whole image display mode according to the embodiment. As shown in FIG. 6, the whole image display mode is a mode for displaying the whole image, that is, the image before the original image is divided.

[0063] For example, in the whole image display mode, the division positions for dividing the original image are displayed superimposed on the original image, as shown in Fig. 6. This allows the administrator to check what divided images will be generated from the original image.

[0064] In this way, display control unit 28 can make it easier for the administrator to check the divided image by switching between the divided image and the entire image and displaying them on display unit 12. Note that the divided image and the entire image can be switched when an operation is performed via operation unit 13 on a GUI indicating a switching button at the top left of the screens shown in FIGS.

[0065] <2. Processing Procedure> Next, the procedure of information processing by the processing unit 14 of the image inspection device 100 according to the embodiment will be described. Fig. 7 is a flowchart showing an example of information processing by the processing unit 14 of the image inspection device 100 according to the embodiment. Note that, here, the processing executed by the processing unit 14 of the image inspection device 100 in the learning stage of the decision model will be described.

[0066] 7, the processing unit 14 acquires an original image to be used for learning a determination model (step S101). Next, the processing unit 14 extracts feature points of the inspection object from the original image and generates a feature point image (step S102).

[0067] Next, the processing unit 14 inputs the feature point image to a CNN, which is a determination model (step S103), and determines division positions of the original image (step S104). Next, the processing unit 14 generates a plurality of divided images by dividing the original image based on the determined division positions (step S105).

[0068] Next, the processing unit 14 performs object detection for each of the plurality of divided images using the detection model (step S106), and feeds back the detection result to the CNN, which is the decision model (step S107).

[0069] Then, the processing unit 14 determines whether the learning of the determination model has been completed (step S108). If the processing unit 14 determines that the learning of the determination model has been completed (step S108: Yes), the processing unit 14 terminates the processing. If the processing unit 14 determines that the learning of the determination model has not been completed (step S108: No), the processing unit 14 proceeds to the processing of step S101.

[0070] <3. Modifications> In the above-described embodiment, the case where the learning of the decision model is performed by the image inspection device 100 has been described, but this is not limited to this. In other words, the device that performs the learning of the decision model and the image inspection device 100 may be configured as separate devices.

[0071] In the above-described embodiment, the division position and resolution are determined using a determination model such as CNN, but the present invention is not limited to this. For example, the division position and resolution may be determined on a rule basis.

[0072] A modified example of the image inspection device 100 according to the embodiment will be described with reference to Figures 8 to 10. Figure 8 is a diagram showing another example of a compressed image. In the example shown in Figure 8, the compression unit 21 of the processing unit 14 compresses the original image into a frequency image in which the object to be inspected shown in the original image is converted into a different representation method.

[0073] 8, the compression unit 21 performs image division on a grayscale image obtained by converting an original image into a grayscale image (step S31). The compression unit 21 divides the grayscale image into a grid pattern at a preset scale. For example, the compression unit 21 divides the grayscale image into 4 vertically and 8 horizontally, thus dividing the grayscale image into a total of 32 parts.

[0074] Next, the compression unit 21 performs image compression on each of the divided grayscale images (step S32), compressing each divided grayscale image into a frequency image. The compression unit 21 performs a discrete cosine transform on each divided grayscale image, thereby generating a plurality of frequency images from each divided grayscale image.

[0075] Then, the determination unit 24 of the processing unit 14 determines the processing based on the intensity of each frequency included in the frequency image. For example, in this case, the determination unit 24 determines the processing based on a rule. Fig. 9 is a diagram showing an example of the relationship between the frequencies included in the frequency image and the processing.

[0076] In the example shown in FIG. 9, "period" and "rule" are associated with each other. "Period" indicates the distance per one period of frequency. More specifically, "period" is the value obtained by converting the distance per one period into pixels. "Rule" indicates the rule for the corresponding period.

[0077] The greater the change in luminance in the grayscale image, the shorter the frequency of the peak extracted in the frequency image, and the smaller the change in luminance in the grayscale image, the longer the frequency of the peak extracted in the frequency image.

[0078] Therefore, if the peak extracted from the grayscale image is of low frequency, it indicates that there is little difference from the surroundings, and if the frequency extracted from the grayscale image is of high frequency, it indicates that there is a large difference from the surroundings.

[0079] When inspecting a circuit board, areas on the board where no electronic components are placed are imaged with only the board itself, so there is little difference in brightness, and low-frequency peaks are extracted from the grayscale image. On the other hand, areas on the board where electronic components are placed are imaged with a difference in brightness between the board and the electronic components, so high-frequency peaks are extracted from the grayscale image.

[0080] Therefore, there is a high possibility that electronic components are placed in areas with high frequencies, i.e., short periods, and there is a low possibility that electronic components are placed in areas with low frequencies, i.e., long periods.

[0081] 9, the determination unit 24 integrates a grayscale image having a period of less than 8 pixels with a grayscale image of an adjacent region. In this case, the determination unit 24 integrates a plurality of frequency images to determine the division positions of the original image.

[0082] 9, the determination unit 24 determines the resolution according to the number of pixels for a frequency image having a period of 8 pixels or more. In the example shown in Fig. 9, the determination unit 24 determines the resolution for a frequency image having a period of 8 pixels or more and less than 16 pixels to be the original resolution (100% resolution), which is the resolution of the original image.

[0083] Furthermore, the determination unit 24 determines the resolution to be 50% for frequency images with a period of 17 pixels or more and less than 32 pixels, and determines the resolution to be 25% for frequency images with a period of more than 32 pixels.

[0084] That is, the determination unit 24 gradually decreases the resolution as the period becomes longer, thereby enabling the determination unit 24 to optimize the resolution according to the frequency.

[0085] Furthermore, when determining these rules, the feedback unit 27 performs feedback according to the detection results by the detection unit 26, thereby optimizing the rules. Specifically, the feedback unit 27 adjusts the specific value of the period and the specific value of the resolution according to the detection results by the detection unit 26. This makes it possible to optimize the rules related to the frequency image.

[0086] The determination unit 24 may also determine the processing to be performed on a rule basis for the feature point image by the compression unit 21. Fig. 10 is a diagram showing an example of the relationship between the feature point density of the feature point image and the processing.

[0087] As shown on the left side of Fig. 10, the determination unit 24 divides the feature point image into a plurality of regions. Then, the determination unit 24 determines the resolution according to the feature point density of each divided region according to the rule shown on the right side of Fig. 10.

[0088] In the example shown in Fig. 10, the resolution of an area where the feature point density is greater than the first threshold th1 is 100%, i.e., the resolution of the original image. Also, the resolution of an area where the feature point density is greater than or equal to the first threshold th1 and the second threshold th2 (<first threshold th1) is 50% of the resolution of the original image, and the resolution of an area where the feature point density is less than the second threshold th2 and the third threshold th3 (<second threshold th2) is 25% of the resolution of the original image. Also, as shown in Fig. 4, an area where the feature point density is less than the third threshold is determined to be a background area.

[0089] Furthermore, when determining these rules, the feedback unit 27 performs feedback in accordance with the detection results by the detection unit 26, thereby optimizing the rules. Specifically, the feedback unit 27 adjusts the specific values ​​of the first threshold value th1 to the third threshold value th3 and the specific value of the resolution in accordance with the detection results by the detection unit 26.

[0090] <4. Hardware Configuration> FIG. 11 is a diagram showing an example of the hardware configuration of the processing unit 14 in the image inspection device 100 according to the embodiment.

[0091] 11, each processing unit 14 has a processor 60, a memory 61, an input / output interface (I / F) 62, and a media interface (I / F) 63. The processor 60, the memory 61, the input / output interface 62, and the media interface 63 are connected by a bus 65.

[0092] The processor 60 includes, for example, one or more of a central processing unit (CPU), a micro processing unit (MPU), and a system large scale integration (LSI). The memory 61 is a random access memory (RAM) such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), or a semiconductor memory element.

[0093] The processor 60 transmits and receives information to and from output devices such as the display unit 12 and input devices such as the operation unit 13 via the input / output interface 62. The processor 60 acquires data from the input devices via the input / output interface 62. The processor 60 also outputs generated data to the output devices via the input / output interface 62.

[0094] The media interface 63 reads a program or data stored in a recording medium 64 and provides the read data or program (an example of an information processing program) to the processor 60 via the memory 61. The processor 60 loads the program from the recording medium 64 onto the memory 61 via the media interface 63 and executes the loaded program. The recording medium 64 is, for example, an optical recording medium such as a DVD (Digital Versatile Disc) or a PD (Phase Change Rewritable Disc), a magneto-optical recording medium such as an MO (Magneto-Optical disk), a tape medium, a magnetic recording medium, or a semiconductor memory.

[0095] The processor 60 executes a program using the memory 61 or the like as a working area, thereby realizing the functions of the acquisition unit 20, compression unit 21, extraction unit 22, aggregation unit 23, determination unit 24, generation unit 25, detection unit 26, feedback unit 27, and display control unit 28.

[0096] The processing unit 14 may be partially or entirely realized by an integrated circuit such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or a digital signal processor (DSP).

[0097] <5.Other> Although an embodiment of the present invention has been described above, this embodiment is presented as an example and is not intended to limit the scope of the invention. This embodiment can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, as well as within the scope of the invention and its equivalents as set forth in the claims.

[0098] The present technology can be configured as follows. (1) An image inspection device that performs an image inspection of an object to be inspected based on image data of an image captured by an imaging unit that captures an external appearance of the object to be inspected, a compression unit that replaces the inspection object shown in the image data with another representation method and compresses the image data; a determination unit that determines division positions from a compressed image, which is the image data compressed by the compression unit; a generation unit that generates a plurality of divided images by dividing the image data at the division positions determined by the determination unit; An image inspection device comprising: (2) The determination unit: determining a resolution for each of the plurality of divided images from the compressed image; The generation unit The divided image having the resolution determined by the determination unit is generated. An image inspection device as described in (1). (3) The determination unit: In the learning stage, the division positions are determined based on an output result obtained by inputting the compressed image to a learning model that has learned the relationship between the division positions of the compressed image and the inspection results of the inspection object in the divided images. (2) An image inspection device according to the present invention. (4) The learning model is A model that further learns a relationship between the resolution of each of the plurality of divided images and an inspection result of the inspection object in the divided images, The determination unit Using the learning model, the resolution of each of the plurality of divided images is determined. (3) An image inspection device according to the present invention. (5) The compression section extracting feature points of the inspection object from the image data, and replacing the inspection object shown in the image data with the feature points; The determination unit The dividing position is determined based on the compressed image in which the inspection object shown in the image data has been replaced with feature points. An image inspection device according to any one of (1) to (4). (6) The determination unit A feature amount image indicating the distribution of the feature points is used as the compressed image, and the division positions are determined. (5) An image inspection device according to (5). (7) a counting unit that counts the number of feature points for each predetermined region; Equipped with The determination unit A region where the number of the feature amounts collected by the collecting unit is equal to or less than a threshold is determined as a background region, and is excluded from the target of image inspection of the inspection object. An image inspection device according to (5) or (6). (8) The determination unit: The division position is determined from the compressed image in which the inspection object shown in the image data is converted into frequency information corresponding to a change in luminance. An image inspection device according to any one of (1) to (7). (9) a display processing unit that switches between the image data and the divided images in response to a user operation and displays them on a display unit; The image inspection device according to any one of (1) to (8) is provided with: (10) An image segmentation method for performing an image inspection of an object to be inspected based on image data of an image captured by an imaging unit that captures an appearance of the object to be inspected, comprising: a compression step of replacing the inspection object shown in the image data with another representation method and compressing the image data; a determining step of determining division positions from a compressed image, which is the image data compressed in the compressing step; a generating step of generating a plurality of divided images by dividing the image data at the division positions determined in the determining step; An image segmentation method comprising: (11) An image segmentation program for performing an image inspection of an object to be inspected based on image data of an image captured by an imaging unit that captures an appearance of the object to be inspected, a compression step of compressing the image data by replacing the object of inspection depicted in the image data with another representation method; a determination step of determining division positions from a compressed image, which is the image data compressed by the compression step; a generating step of generating a plurality of divided images by dividing the image data at the division positions determined by the determining step; An image segmentation program that causes a computer to execute the following.

[0099] The disclosed embodiments should be considered in all respects as illustrative and not restrictive. Indeed, the above-described embodiments may be embodied in various forms. Furthermore, the above-described embodiments may be omitted, substituted, or modified in various forms without departing from the scope and spirit of the appended claims. [Explanation of symbols]

[0100] 2 Cameras 3. Lighting 10. Communications Department 11 Storage section 12 Display section 13 Control section 14 Processing section 20 Acquisition Department 21 Compression section 22 Extraction part 23 Counting Department 24 Decision Section 25 Generation part 26 Detector 27 Feedback Section 28 Display control unit 100 Image inspection device

Claims

1. An image inspection device that performs an image inspection of an object to be inspected based on image data of an image captured by an imaging unit that captures an appearance of the object to be inspected, a compression unit that replaces the inspection object shown in the image data with another representation method and compresses the image data; a determination unit that determines division positions from a compressed image, which is the image data compressed by the compression unit; a generation unit that generates a plurality of divided images by dividing the image data at the division positions determined by the determination unit; An image inspection device comprising:

2. The determination unit determining a resolution for each of the plurality of divided images from the compressed image; The generation unit The divided image having the resolution determined by the determination unit is generated. The image inspection device according to claim 1 .

3. The determination unit In the learning stage, the division positions are determined based on an output result obtained by inputting the compressed image to a learning model that has learned the relationship between the division positions of the compressed image and the inspection results of the inspection object in the divided images. The image inspection device according to claim 2 .

4. The learning model is A model that further learns a relationship between the resolution of each of the plurality of divided images and an inspection result of the inspection object in the divided images, The determination unit Using the learning model, the resolution of each of the plurality of divided images is determined. The image inspection device according to claim 3.

5. The compression section extracting feature points of the inspection object from the image data, and replacing the inspection object shown in the image data with the feature points; The determination unit The dividing position is determined based on the compressed image in which the inspection object shown in the image data has been replaced with feature points. The image inspection device according to claim 1 .

6. The determination unit A feature amount image indicating the distribution of the feature points is used as the compressed image, and the division positions are determined. The image inspection device according to claim 5.

7. a counting unit that counts the number of feature points for each predetermined region; Equipped with The determination unit A region in which the number of feature points counted by the counting unit is equal to or less than a threshold is determined as a background region, and is excluded from the target of image inspection of the inspection object. The image inspection device according to claim 5.

8. The determination unit The division position is determined from the compressed image in which the inspection object shown in the image data is converted into frequency information corresponding to a change in luminance. The image inspection device according to claim 1 .

9. a display processing unit that switches between the image data and the divided images in response to a user operation and displays them on a display unit; The image inspection device according to claim 1 , comprising:

10. 1. An image segmentation method for performing image inspection of an inspection object based on image data of an image captured by an imaging unit that captures an appearance of the inspection object, comprising: a compression step of replacing the inspection object shown in the image data with another representation method and compressing the image data; a determining step of determining division positions from a compressed image, which is the image data compressed in the compressing step; a generating step of generating a plurality of divided images by dividing the image data at the division positions determined in the determining step; An image segmentation method comprising:

11. An image segmentation program for performing an image inspection of an inspection object based on image data of an image captured by an imaging unit that captures an appearance of the inspection object, a compression step of compressing the image data by replacing the object of inspection depicted in the image data with another representation method; a determination step of determining division positions from a compressed image, which is the image data compressed by the compression step; a generating step of generating a plurality of divided images by dividing the image data at the division positions determined by the determining step; An image segmentation program that causes a computer to execute the following.

Images