Workpiece inspection device and method

Abstract

To reduce over-detection of workpiece defects of a given defect type.SOLUTION: An inspection device disclosed herein inputs multiple fragment images extracted from an inspection target image of a workpiece into a learning model designed to take an image as input to output a type (of a defect) so as to determine the type of a defect for each of the multiple fragment images. The inspection device determines whether defects of a given defect type are in the inspection target image or not according to whether or not a type of defect determined for each of the multiple fragment images is the given defect type.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention generally relates to techniques for inspecting workpieces. [Background technology]

[0002] One example of this type of technology is the inspection device disclosed in Patent Document 1. Patent Document 1 discloses the following: The inspection device makes a primary judgment as to whether the quality of the glass tube is good or bad based on a comparison between a test image of the glass tube and a threshold value. If the glass tube is judged to be defective in the primary judgment, the inspection device extracts a defect image (an image of the area containing the defect) from the test image and inputs the defect image into a learning model to make a secondary judgment in which the type of defect is classified. The inspection device then re-judges the quality of the glass tube based on a comparison between the test image and a threshold value corresponding to the classified defect type. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2020-85774 Summary of the Invention [Problem to be solved by the invention]

[0004] Cracks are one type of defect in workpieces, and it is desirable to reduce overdetection of cracks in workpiece inspection.

[0005] However, it is difficult to reduce the overdetection of cracks with the technique disclosed in Patent Document 1. One of the reasons is as follows.

[0006] In other words, the image input to the learning model is an image of the part where the defect is captured, i.e., an image of the entire defect. Therefore, if the defect is a crack, an image of the entire crack will be input to the learning model.

[0007] However, it is difficult to train a learning model to the point where it can accurately determine that a defect type is a crack. One reason for this is that cracks vary in length, and depending on the crack, the image showing the crack may be a wide-area image. In a wide-area image, the area without the crack may be wider than the area with the crack. Therefore, the crack's features are small, resulting in a decrease in the accuracy of the learning model. Furthermore, the area without the crack may contain defects other than cracks (such as powder adhesion). If both cracks and other defects are captured in a single image, the crack's features will be inaccurate, resulting in a decrease in the accuracy of the learning model.

[0008] For these reasons, it is difficult to train the learning model with high accuracy, and therefore, even if images of defects other than cracks are input to the learning model, there is a risk that the defect type will be over-detected as being a crack. Furthermore, such a problem can also occur with other types of defects other than cracks (predetermined types of one-dimensional defects on a workpiece), such as burrs (typically scratches), which are predetermined types of two-dimensional defects on a workpiece. [Means for solving the problem]

[0009] The inspection device inputs a plurality of image fragments extracted from a test image of a workpiece into a learning model that takes an image as input and a type as output, and determines the type of each of the plurality of image fragments. The inspection device determines whether a defect of the predetermined defect type is captured in the test image based on whether the determined type of each of the plurality of image fragments is a predetermined defect type. [Effects of the Invention]

[0010] According to the present invention, since the images input to the learning model are fragmentary images, a learning model with high judgment accuracy can be prepared, thereby reducing over-detection of defects of a specified defect type in a workpiece. [Brief explanation of the drawings]

[0011] [Figure 1] 1 schematically illustrates the configuration of an inspection system according to a first embodiment of the present invention. [Figure 2] The configuration of the control device is shown. [Figure 3] The process performed by the image processing unit is shown below. [Figure 4] 10 shows the processing performed by the individual determination unit. [Figure 5] Schematic diagram of deep learning model training. [Figure 6] 10 shows the processing performed by the overall determination unit. [Figure 7] 10 shows an example of a test result screen. [Figure 8] 10A and 10B show schematic diagrams of processing performed by a model management unit. [Figure 9A] A first example of extracting a fragment image is shown. [Figure 9B] A second example of image fragment extraction is shown. [Figure 9C] A third example of image fragment extraction is shown. [Figure 10] 10 shows an outline of the flow of processing performed by an inspection device according to a second embodiment of the present invention. [Figure 11] The process of determining the entire roughness is shown below. [Figure 12] The small hole overall determination process is shown. [Figure 13] 10 shows the flow of a learning process performed by a model management unit according to a third embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0012] In the following description, an "interface apparatus" may refer to one or more interface devices. The one or more interface devices may be at least one of the following: One or more I / O (Input / Output) interface devices. The I / O (Input / Output) interface devices are interface devices to at least one of the I / O device and a remote display computer. The I / O interface device to the display computer may be a communications interface device. The at least one I / O device may be a user interface device, for example, either an input device such as a keyboard and a pointing device, or an output device such as a display device. One or more communication interface devices. The one or more communication interface devices may be one or more homogeneous communication interface devices (e.g., one or more NICs (Network Interface Cards)) or two or more heterogeneous communication interface devices (e.g., an NIC and an HBA (Host Bus Adapter)).

[0013] In the following description, "memory" refers to one or more memory devices, which are an example of one or more storage devices, and may typically be a primary storage device. At least one memory device in the memory may be a volatile memory device or a non-volatile memory device.

[0014] In the following description, a "persistent storage device" may refer to one or more persistent storage devices, which are an example of one or more storage devices. A persistent storage device may typically be a non-volatile storage device (e.g., an auxiliary storage device), and more specifically, may be, for example, a hard disk drive (HDD), a solid state drive (SSD), a non-volatile memory express (NVME) drive, or a storage class memory (SCM).

[0015] In the following description, the term "storage device" may refer to at least one of memory and persistent storage device.

[0016] Furthermore, in the following description, a "processor" may refer to one or more processor devices. The at least one processor device may typically be a microprocessor device such as a CPU (Central Processing Unit), but may also be another type of processor device such as a GPU (Graphics Processing Unit). The at least one processor device may be a single-core or multi-core. The at least one processor device may also be a processor core. The at least one processor device may also be a processor device in a broader sense, such as a circuit that is a collection of gate arrays written in a hardware description language that performs part or all of the processing (for example, an FPGA (Field-Programmable Gate Array), a CPLD (Complex Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit)).

[0017] Furthermore, in the following description, functions may be described using the expression "yyy unit." However, the functions may be realized by one or more computer programs executed by a processor, by one or more hardware circuits (e.g., FPGAs or ASICs), or by a combination thereof. When a function is realized by a program executed by a processor, the specified processing is performed using a storage device and / or an interface device, etc., as appropriate, and therefore the function may be considered to be at least a part of the processor. Processing described using a function as the subject may be processing performed by a processor or a device having the processor. A program may be installed from a program source. The program source may be, for example, a computer from which the program is distributed or a computer-readable recording medium (e.g., a non-transitory recording medium). The description of each function is an example, and multiple functions may be combined into one function, or one function may be divided into multiple functions.

[0018] Hereinafter, several embodiments of the present invention will be described with reference to the drawings. [First embodiment]

[0019] FIG. 1 schematically shows the configuration of an inspection system 500 according to a first embodiment of the present invention.

[0020] Inspection system 500 includes a rotating stage 510 , a pair of line illuminators 520 , a line sensor camera 530 , an area camera 560 , a ring illuminator 580 , and inspection device 450 .

[0021] The rotation stage 510 is a stage for placing a cylindrical ceramic honeycomb structure 550 thereon. The cylindrical honeycomb structure 550 has an upper surface 551 (first bottom surface), a bottom surface 552 (second bottom surface), and a side surface 553. The cylindrical honeycomb structure 550 (particularly the side surface 553) is an example of a workpiece. The rotation stage 510 is movable in the X, Y, and Z directions, and is also capable of rotating the cylindrical honeycomb structure 550 around a rotation axis (parallel to the height direction (Z direction) of the cylindrical honeycomb structure 550).

[0022] The line illuminators 520 are illuminators for irradiating light onto the side surface 553 (outer peripheral surface) of the cylindrical honeycomb structure 550. They are arranged on the left and right sides along the Y direction, sandwiching a linear imaging range on the side surface 553. The length (height) of the linear imaging range may be the same as the height of the side surface 553.

[0023] The line sensor camera 530 captures the light reflected from the side surface 553 of the cylindrical honeycomb structure 550 .

[0024] The area camera 560 photographs the upper surface 551 of the cylindrical honeycomb structure 550 placed on the rotary stage 510 from above.

[0025] The ring illuminator 580 is an illuminator capable of irradiating the upper surface 551 of the cylindrical honeycomb structure 550 placed on the rotary stage 510 with light from above.

[0026] The inspection device 450 may be a computer such as a personal computer, and includes an input device 572, a display device 540, and a control device 570 connected thereto. The input device 572 and the display device 540 may be integrated into one device such as a touch panel.

[0027] A captured image from the line sensor camera 530 capturing an image of the side surface 553 of the rotating cylindrical honeycomb structure 550 is input to the control device 570. A two-dimensional captured image is obtained by arranging the linear captured images along the Y direction (a direction perpendicular to the capturing direction of the line sensor camera 530 and the height direction (Z direction)). If there is a crack on the side surface 553 of the cylindrical honeycomb structure 550, the control device 570 detects the crack.

[0028] Although the light emitted by the pair of line illuminators 520 may be different colors, in this embodiment, the pair of line illuminators 520 emits light of the same color onto a linear imaging range. Furthermore, the line sensor camera 530 may be a color line sensor camera, but in this embodiment, it is a monochrome line sensor camera. The line sensor camera 530 has high resolution and sensitivity. For this reason, there is a concern that cracks may be overdetected, but the inspection device 450 according to this embodiment reduces the overdetection of cracks.

[0029] FIG. 2 shows the configuration of the control device 570.

[0030] The control device 570 includes an interface unit 10, a storage unit 20, and a processor 30 connected thereto.

[0031] The interface device 10 is communicatively connected to an input device 572 , a display device 540 and a line sensor camera 530 .

[0032] The storage device 20 stores computer programs and information. For example, the storage device 20 stores a deep learning model 260, workpiece specification information 270, and inspection result information 280.

[0033] The deep learning model 260 is an example of a learning model, and takes an image as input and a defect type as output. The deep learning model 260 is typically a neural network. The deep learning model 260 is used for learning and inference by the individual determination unit 220, which will be described later.

[0034] The work specification information 270 is information that represents the work specifications for each customer. A "customer" is a recipient of the cylindrical honeycomb structure 550. For each customer, the "work specification" is the specification of the cylindrical honeycomb structure 550, and includes the conditions that are considered to be cracks (particularly, the conditions for the length of the cracks).

[0035] The inspection result information 280 is information that indicates the inspection result for each cylindrical honeycomb structure 550. The inspection result may include the presence or absence of defects, the type of defects detected, and a test image showing the defects. The inspection result may also include detailed results such as the position of the defects (for example, coordinates when a predetermined position on the side surface 553 of the cylindrical honeycomb structure 550 is set as the reference coordinate (origin)).

[0036] The processor 30 executes a computer program stored in the storage device 20 to realize functions such as an image processing unit 210, an individual determination unit 220, an overall determination unit 230, a display control unit 240, and a model management unit 250. In addition, a control unit (not shown) that controls various devices such as the rotating stage 510 and the line sensor camera 530 may also be realized.

[0037] The following describes the processing performed by functions 210, 220, 230, 240, and 250. In the following description, the vertical direction is synonymous with the longitudinal direction (Z direction) of the line photographed image, and the horizontal direction is synonymous with the arrangement direction (Y direction) of the line photographed image.

[0038] FIG. 3 shows the processing performed by the image processing unit 210.

[0039] Image processing unit 210 performs the pre-processing of the processing performed by individual determination unit 220. The pre-processing is processing to generate a test image to be input to individual determination unit 220. Specifically, for example, image processing unit 210 may extract two-dimensional defect images from a two-dimensional captured image (an image composed of line captured images arranged horizontally) and perform the following steps S301 to S304 for each defect image. The "two-dimensional captured image" may be, for example, a strip image of the entire circumference of side surface 553. The "defect image" may be the entire two-dimensional captured image, or an image extracted from the two-dimensional captured image and depicting one entire defect. In other words, there may be a defect image for each defect, and the vertical and horizontal sizes of the defect image may vary depending on the size of the defect (for example, the vertical and horizontal sizes of the defect). In the following description, it is assumed that one defect image depicts one defect.

[0040] The image processing unit 210 performs a filtering process (smoothing process) on the defect image to remove noise components from the defect image (S301). For example, for a long vertical range of the defect, low frequency components are removed in the horizontal direction and high frequency components are removed in the vertical direction, resulting in a clearer image of the long vertical range of the defect.

[0041] Next, the image processing unit 210 performs binarization processing on the defect image that has been subjected to the filtering processing (S302), thereby extracting pixels with low brightness in the defect image.

[0042] Next, the image processing unit 210 performs morphology processing on the binarized defect image (S303). In this morphology processing, the pixels extracted in the binarization processing of S302 are expanded and connected in the longitudinal direction of the defect that overlaps the pixels. As a result, even if the pixels extracted in the binarization processing of S302 are intermittent in the longitudinal direction of the defect, they are treated as a single continuous defect. Therefore, even if the defect is a crack and the binarization processing results in the crack being intermittent, the morphology processing extracts the defect as a single continuous crack.

[0043] Finally, the image processing unit 210 calculates features for various items (e.g., circularity, coordinates, aspect ratio) from the defect image that has been subjected to morphological processing, and classifies the type of defect shown in the defect image into either a crack or a non-crack (other than a crack) based on the calculated features.

[0044] The entire or part of a defect image classified as a crack defect type is processed as a test image by the individual determination unit 220. The defect image classified as a non-crack defect type may or may not be processed by the individual determination unit 220.

[0045] Hereinafter, the test image input to the individual determination unit 220 is assumed to be a defect image or a part thereof classified by the image processing unit 210 as a defect type that is a crack. For this reason, in the following description, the test image is assumed to be an image resulting from processing (processing including morphology processing, etc.) performed on the defect image by the image processing unit 210, that is, an image that shows a defect (object) detected as a crack. For this reason, the defect shown in the test image is assumed to be a crack. Note that the crack shown in the test image may actually be a continuous crack even if it is intermittent due to the above-mentioned morphology processing.

[0046] FIG. 4 shows the process performed by the individual determination unit 220.

[0047] The individual determination unit 220 extracts a fragment image from a range of the crack range that has not been acquired as a fragment image (S401). The "crack range" is an image range in the test image in which a crack appears, and may be the entire area or a part of the test image. The "fragment image" is an image in which a crack appears (strictly speaking, a part of a crack), and is typically a rectangular image (see reference numeral 41 in FIG. 4). The fragment image may be, for example, an image of 200 pixels in length and width. The fragment images may also be extracted in a predetermined order (for example, sequentially from the beginning to the end of the crack range).

[0048] The individual determination unit 220 performs inference to determine whether the type of defect shown in the fragmentary image is a crack or not. Specifically, the individual determination unit 220 inputs the fragmentary image extracted in S401 to the deep learning model 260 (S402), and acquires the determination result (determined defect type (class)) and reliability (reliability of determination) for the fragmentary image from the deep learning model 260 (S403).

[0049] If the fragment image acquired in S401 is the last fragment image that can be extracted from the crack range (S404: YES), in other words, if all fragment images have been acquired from the crack range, the processing ends. On the other hand, if there is a range in the crack range that has not yet been acquired as a fragment image (S404: YES), the processing returns to S401.

[0050] In this way, for each fragment image extracted from the crack range, the fragment image is input into deep learning model 260, whereby the defect type is determined. As shown in FIG. 5 , deep learning model 260 has been trained using training data for each defect type (e.g., training data 700A as a plurality of fragment images whose defect type should be determined as a crack, and training data 700B as a plurality of fragment images whose defect type should be determined as a fiber). Defect types other than cracks for which training data is prepared may be at least one of powder adhesion, molding streaks, and dirt, instead of or in addition to fibers. In other words, the output of deep learning model 260 may be a crack or any of a plurality of defect types other than cracks. Training data may be prepared for defect types specific to ceramic workpieces, and deep learning model 260 may be trained by individual judgment unit 220 to be able to determine the defect type.

[0051] 6 shows the processing performed by the overall determination unit 230. In the following description, a fragment image whose defect type is determined to be a crack will be referred to as a "crack fragment image," and a fragment image whose defect type is determined to be any type other than a crack will be referred to as a "non-crack fragment image."

[0052] The overall determination unit 230 selects one unselected fragment image (S601) from among the fragment image fragments whose defect types have been determined by the individual determination unit 220. The fragment image fragments may be selected in a predetermined order (for example, sequentially from the beginning to the end of the crack range).

[0053] The whole determination unit 230 determines whether the fragment image selected in S601 is a crack fragment image (S602). If the determination result in S602 is false (S602: NO), the process proceeds to S605.

[0054] If the determination result of S602 is true (S602: YES), the overall determination unit 230 determines whether the distance between the crack fragment image selected in S601 and the previously selected crack fragment image (the crack fragment image selected in the previous S601) that is closest to the selected crack fragment image satisfies a connection condition (S603). The "connection condition" represents the allowable distance between crack fragment images. The distance may be expressed as the number of pixels or the number of fragment images. If the distance between the crack fragment images is zero, the crack fragment images are adjacent to each other.

[0055] If the determination result of S603 is true (S603: YES), the overall determination unit 230 connects the crack fragment images together (S604). Specifically, for example, as shown in the figure, if there are one or more non-crack fragment images between crack fragment images, the overall determination unit 230 changes the determination result (fragment image attribute) of each of the one or more non-crack fragment images from non-crack (other than crack) to crack. As a result, all the images between the crack fragment images become crack fragment images, and therefore, multiple crack fragment images become continuous without any non-crack fragment images in between.

[0056] After S602: NO or S604, the whole determination unit 230 determines whether or not all the fragment images in the crack range have been selected (S605). If the determination result in S605 is false (S605: NO), the process returns to S601.

[0057] If the determination result of S605 is true (S605: YES), the overall determination unit 230 determines whether the crack length satisfies the condition (S606). Specifically, the overall determination unit 230 identifies work specifications corresponding to the customer of the cylindrical honeycomb structure 550 corresponding to the test image from the work specification information 270. The overall determination unit 230 determines whether the crack length (the length following consecutive crack fragment images without separating non-crack fragment images) satisfies the crack length condition expressed in the identified work specification. The crack length condition is a length condition for being recognized as a crack. The crack length may be expressed in SI units (e.g., mm (millimeters)) or may be expressed as the number of fragment images or pixels.

[0058] If the determination result in S606 is true (S606: YES), the overall determination unit 230 determines that the type of defect in the test image is a crack (S607). That is, it is determined that the detection result (classification result) by the image processing unit 210 is correct.

[0059] On the other hand, if the determination result in S606 is false (S606: NO), the overall determination unit 230 determines that the type of defect in the test image is a non-crack (S608). In other words, the detection result by the image processing unit 210 is determined to be an error (overdetection). Note that if S606: NO, the determination that the defect is a non-crack may include a determination of what type of defect other than a crack the defect is. For example, if the number of fragment images determined to be "powder attached" is the largest among all the fragment images extracted from the crack range (and if the proportion of fragment images determined to be "powder attached" to all the fragment images extracted from the crack range is equal to or greater than a predetermined proportion), the type of defect may be determined to be "powder attached."

[0060] The workpiece ID of the cylindrical honeycomb structure 550 may be input to the image processing unit 210 together with the defect image. The workpiece ID may be passed from the image processing unit 210 to the individual judgment unit 220, and from the individual judgment unit 220 to the overall judgment unit 230, along the processing flow. Information representing the detection results (classification results) by the image processing unit 210, information representing the judgment results (and their reliability) for each fragment image by the individual judgment unit 220, and information representing the judgment results by the overall judgment unit 230 may be stored in the inspection result information 280, and associated with the workpiece ID. The display control unit 240 displays an inspection result screen on the display device 540 based on the inspection result information 280.

[0061] FIG. 7 shows an example of a test result screen 700.

[0062] The test result screen 700 is typically a GUI (Graphical User Interface). For example, the test result screen 700 displays the following: (A) The customer's customer ID as specified by the operator (e.g., administrator). (B) Work ID, inspection results and details for each work (cylindrical honeycomb structure 550) for which it has been determined whether the crack length satisfies the conditions based on the work specifications corresponding to the customer ID.

[0063] The inspection result (B) is either the detection result (classification result) by the image processing unit 210 or the judgment result by the overall judgment unit 230.

[0064] The details of (B) include the reason for the inspection result of (B), and if the inspection result of (B) is not a crack, the details of (B) include the type of defect.

[0065] FIG. 8 shows a schematic diagram of the processing performed by the model management unit 250.

[0066] The model management unit 250 determines whether to continue using the deep learning model 260 based on the reliability of each of the multiple fragment images (reliability of the determination result). As a result, if data drift occurs, the data drift is detected and the continuation of the deep learning model 260 is stopped, thereby maintaining the reliability of the inspection device 450. For example, as follows.

[0067] That is, it is assumed that a continuous crack is obtained from the top end (beginning) to the bottom end (end) of the vertically long crack range by morphology processing.

[0068] If no data drift occurs, as shown by the solid line graph, all of the fragmentary images are determined to be crack fragmentary images, and the reliability of the determination results for each fragmentary image is high.

[0069] On the other hand, when data drift occurs, as shown by the dashed line graph, some fragment images that should be determined to be crack fragment images are determined to be non-crack fragment images, or even if a fragment image is determined to be a crack fragment image (i.e., even if the determination result is correct), the reliability of the determination result is low. The model management unit 250 can stop the continued use of such a deep learning model 260.

[0070] The above-described embodiments can be summarized, for example, as follows. The following summary may include supplementary explanations and explanations of modifications to the above explanations.

[0071] In the inspection device 450, the individual determination unit 220 determines the type of each of a plurality of image fragments extracted from the test image by inputting each of the plurality of image fragments into the deep learning model 260. In the inspection device 450, the overall determination unit 230 determines whether or not a crack is present in the test image based on whether or not the determined type of each of the plurality of image fragments is a crack.

[0072] Since the image input to the deep learning model 260 is a fragmented image smaller than the entire test image, a learning model with high judgment accuracy can be prepared, which is expected to reduce overdetection of cracks. For example, in the above-mentioned embodiment, the crack detection by the image processing unit 210 (judgment result based on the result of feature classification) may be overdetection, but the individual judgment unit 220 and overall judgment unit 230 at the subsequent stage can reduce overdetection of cracks.

[0073] One method for reducing overdetection is to employ a computationally intensive filtering process, such as processing the test image using a bilateral filter, instead of using the deep learning model 260. However, due to the high computational load of the process, the inspection takes a long time. In the above-described embodiment, by employing processing using the deep learning model 260 in the latter stage, the filtering process in the former stage can be a computationally intensive process such as smoothing, which is a process with a low computational load. As a result, it is expected that the inspection time can be shortened while maintaining the inspection accuracy.

[0074] Other types of learning models, such as decision trees, may be used instead of the deep learning model 260. However, it is difficult for other types of learning models to perform highly accurate inference using images as input. The deep learning model 260 is suitable for inference using images as input, and is expected to perform highly accurate inference.

[0075] Furthermore, the deep learning model 260 is a so-called black box model. That is, even if the type of an input image fragment is determined (output), the reason for the determination is not output. In other words, there is no explainability. Therefore, if the image input to the deep learning model 260 is a test image, the determination result for the test image cannot be explained. In this embodiment, the individual determination unit 220 determines the type of each image fragment, and then the overall determination unit 230 determines whether or not a crack is present in the test image based on the determination result for each image fragment (specifically, for example, if a defect is present in the test image, the type of the defect). Therefore, it is possible to achieve both high accuracy in inference using an image as input and explainability in the determination result for the test image (for example, an explanation that the crack was determined to be a crack (or not) because the crack length satisfied (or did not satisfy) a condition).

[0076] The test image may be a photographed image of the entire workpiece (for example, the side surface 553 of the cylindrical honeycomb structure 550), or may be a defect image (an image of a range in which a defect is captured) of a portion of the photographed image identified by a predetermined method (for example, rule-based judgment). The defect image may be an image of the crack range described above, or may be a wide image including the crack range (an example of a defect range in which a defect is captured).

[0077] A plurality of image fragments may be extracted from the test image by the individual determination unit 220. Examples of extracted image fragments may be any of the following (note that the image fragments may typically be square or rectangular). As shown in Fig. 9A, the individual determination unit 220 extracts image fragments 901 from the test image 900 so that no gaps are formed between the image fragments 901. The entire area or a part of the test image 900 is the crack range. That is, in the example shown in Fig. 9A, image fragments 901 may also be extracted from areas other than the crack range. Furthermore, it does not matter whether the image processing unit 210 identifies which area of ​​the test image 900 is the crack range. As illustrated in FIG. 9B , the individual judgment unit 220 extracts image fragments 901 only from crack ranges 910 in the test image 900 so that there are no gaps between the image fragments 901. Cracks may be discontinuous in their longitudinal direction, or multiple cracks may be lined up. In such cases, the test image 900 is roughly divided into one or more crack ranges 910 and ranges other than the crack ranges 910. As illustrated in FIG. 9B , the individual judgment unit 220 extracts image fragments 901 only from the crack ranges 910. For each crack range 910, an image covering the entire crack range 910 may be used as the test image 900. 9C, the individual determination unit 220 may extract image fragments 901 from any or predetermined locations of the test image 900, and gaps of any or predetermined length may be generated between the image fragments 901. The length of the gap may be the same as the length of the image fragments (the length along the arrangement direction of the image fragments), or may be shorter or longer than the length of the image fragments (for example, it may be n times the fragment length (n is a natural number)). Furthermore, image fragments may be extracted from ranges other than the crack range, or may be extracted only from the crack range.

[0078] The overall determination unit 230 may determine that a crack is present in the test image if the crack length identified from one or more crack fragment images satisfies the crack length condition. The accuracy of the determination result for each fragment image is high, and excessive crack detection can be reduced.

[0079] For example, if the shortest crack length that satisfies the crack length condition is the same as or shorter than the crack length represented by one crack fragment image, it will be determined that a crack is present in the test image if there is at least one crack fragment image.

[0080] The overall determination unit 230 may determine that a crack is present in the test image if the crack length (for example, the longitudinal length of the area formed by two or more consecutive crack fragment images (the length along the arrangement direction of the crack fragment images)) identified from two or more consecutive crack fragment images satisfies the crack length condition. This makes it possible to reduce overdetection of cracks.

[0081] If there is a distance between crack fragment images and the distance is less than the allowable distance, the overall determination unit 230 may consider the distance as part of the crack (crack length). This allows for accurate crack detection. For example, even if one or more non-crack fragment images are present between crack fragment images using the fragment image extraction illustrated in FIG. 9A or 9B , if the distance between the crack fragment images is less than the allowable distance, the overall determination unit 230 may change the fragment image determination result for each of the one or more non-crack fragment images to crack. This allows for crack detection even if some fragment images are erroneously determined to be non-crack due to a decrease in reliability of the deep learning model 260, etc. Furthermore, for example, if there is a gap between two consecutive crack fragment images (between a crack fragment image and the next crack fragment image) using the fragment image extraction illustrated in FIG. 9C and the distance of the gap is less than the allowable distance, the overall determination unit 230 may consider the distance of the gap as part of the crack length.

[0082] The display control unit 240 displays the inspection results based on the inspection result information 280, which includes information indicating the results of the judgment by the overall judgment unit 230. The inspection results may be displayed on a display device 540 of the inspection apparatus 450, or on a remote computer connected to the inspection apparatus 450 (e.g., a server). The displayed inspection results may include (a) a judgment result as to whether or not a crack is present in the test image, and (b) a reason for the judgment result in (a), including whether or not the crack length condition is satisfied. This allows the judgment results using the deep learning model 260, which is a black-box model, to be displayed together with the reason (explanation) for the judgment result.

[0083] The crack length conditions may be the conditions specified in the work specifications corresponding to the customer of the cylindrical honeycomb structure 550, among the work specifications specified for each customer. When the crack length conditions differ from customer to customer, if the image input to the deep learning model 260 is a test image (e.g., an image showing the entire crack), training data and training are required for each customer. According to the above-described embodiment, the crack length conditions compared with the crack length identified from the continuous crack fragment images differ from customer to customer, and the judgment for each fragment image is the same regardless of the customer. Therefore, the training data and training can be common regardless of the customer, which is highly convenient.

[0084] The model management unit 250 may determine whether to continue using the deep learning model 260 based on the reliability obtained from the deep learning model 260 for the determination results of each of the multiple image fragments. As a result, if data drift occurs, the data drift is detected and the continuation of the deep learning model 260 with reduced reliability is stopped, thereby maintaining the reliability of the inspection device 450. [Second embodiment]

[0085] A second embodiment of the present invention will be described below, focusing mainly on the differences from the first embodiment, and explanations of the commonalities with the first embodiment will be omitted or simplified.

[0086] FIG. 10 shows an outline of the flow of processing performed by the inspection device 450 according to the second embodiment.

[0087] As the predetermined defect type, other defect types, such as burrs or small holes, can be used instead of or in addition to cracks. Also, although the cracks are vertical cracks in the first embodiment, horizontal cracks can be used in addition to (or instead of) vertical cracks. In other words, "crack" can refer to vertical cracks, horizontal cracks, or a general term for both.

[0088] In this embodiment, vertical cracks, horizontal cracks, burrs, and small holes are adopted as the types of defects.

[0089] For each of the multiple defect types, there is a learning model corresponding to that defect type. Specifically, there are, for example, a vertical crack model 260A that is a learning model for vertical cracks, a horizontal crack model 260B that is a learning model for horizontal cracks, a burr model 260C that is a learning model for burrs, and a small hole model 260D that is a learning model for small holes. Each of the models 260A to 260D is, for example, a deep learning model (typically a neural network).

[0090] For each of the plurality of defect types, the individual determination unit 220 inputs a plurality of image fragments extracted from the test image into a learning model corresponding to the defect type, and determines the type of each of the plurality of image fragments. The overall determination unit 230 determines whether a defect corresponding to the defect type is captured in the test image based on whether the determined type of each of the plurality of image fragments corresponds to the defect type. This is expected to reduce overdetection of various defects in the workpiece (in this embodiment, the side surface 553 of the cylindrical honeycomb structure 550).

[0091] 10, the image processing unit 210 generates a test image to be input to the individual judgment unit 220, and inputs the test image to the individual judgment unit 220. The test image may be common to the four individual judgment processes described below (vertical crack individual judgment process, horizontal crack individual judgment process, burr individual judgment process, and small hole individual judgment process), or may be prepared separately for each individual judgment process. Alternatively, the test image may be prepared separately for each individual judgment process, and each test image for each individual judgment process may be used in all four individual judgment processes.

[0092] Furthermore, when test images are prepared for each individual determination process, the image processing unit 210 may perform filtering, binarization, or morphology processing corresponding to the individual determination process (defect type) for each individual determination process. The image processing unit 210 may classify the morphology-processed image into one of the defect types using a predetermined method such as a rule base. The test image associated with the defect type may be used in the four individual determination processes.

[0093] The individual judgment unit 220 performs individual judgment processing for each defect type, four individual judgment processing in this embodiment. The four individual judgment processing are performed in parallel, but two or more of the individual judgment processing may be performed sequentially. The test image is common to all four individual judgment processing. Not limited to the vertical crack individual judgment processing, the flow of any other individual judgment processing may be similar to the flow shown in FIG. 4. In other words, fragmentary images may be acquired from the test image and input into the learning model 260, whereby the defect type of the fragmentary images may be judged. The extracted fragmentary images may partially overlap with each other (similar to the first embodiment).

[0094] In the individual vertical crack determination process, each fragment image is either a vertical crack fragment image (a fragment image classified as a vertical crack) or a non-vertical crack fragment image (a fragment image classified as something other than a vertical crack).In the individual horizontal crack determination process, each fragment image is either a horizontal crack fragment image (a fragment image classified as a horizontal crack) or a non-horizontal crack fragment image (a fragment image classified as something other than a horizontal crack).In the individual burr determination process, each fragment image is either a burr fragment image (a fragment image classified as a burr) or a non-burr fragment image (a fragment image classified as something other than a burr).In the individual small hole determination process, each fragment image is either a small hole fragment image (a fragment image classified as a small hole) or a non-small hole fragment image (a fragment image classified as something other than a small hole).

[0095] The overall determination unit 230 performs an overall determination process for each defect type, and in this embodiment, four overall determination processes (a vertical crack overall determination process, a horizontal crack overall determination process, a burr overall determination process, and a small hole overall determination process). The four overall determination processes are performed in parallel, but two or more of the overall determination processes may be performed sequentially.

[0096] The flow of the overall vertical crack determination process is as shown in Figure 6. The flow of the overall horizontal crack determination process is also as shown in Figure 6. Specifically, for example, by replacing the "vertical direction" and "vertical crack" in the explanation of the overall vertical crack determination process with "horizontal direction" and "horizontal crack," the following can be adopted as an explanation of the overall horizontal crack determination process (the "vertical direction" and "horizontal direction" are examples of "one-dimensional directions"). The overall determination unit 230 determines that a horizontal crack is present in the test image if the horizontal crack length identified from one or more horizontal crack fragment images arranged horizontally satisfies a horizontal crack length condition, which is a condition related to length (S606: YES). For example, the overall determination unit 230 determines that a horizontal crack is present in the test image if the horizontal crack length identified from two or more horizontally consecutive horizontal crack fragment images satisfies the horizontal crack length condition. If there is a distance between horizontal crack fragment images and the distance is less than the allowable distance, the overall determination unit 230 considers the distance to be part of the horizontal crack. For example, even if there are one or more non-horizontal crack fragment images between horizontal crack fragment images, if the distance between the horizontal crack fragment images is less than the allowable distance, the overall determination unit 230 changes the fragment image determination result for each of the one or more non-horizontal crack fragment images to a horizontal crack.

[0097] The flow of the overall burr determination process is as shown in Fig. 11. S1101 to S1108 correspond to S601 to S608 shown in Fig. 6, and the main differences are as follows: That is, in S1102, the overall determination unit 230 S1101 If the result of the determination in S1102 is true, in S1103, the whole determination unit 230 determines whether the selected fragment image is a burr fragment image. S1101The process determines whether the distance between the selected burr fragment image and the previously selected burr fragment image (the burr fragment image previously selected in S1101) closest to the selected burr fragment image satisfies the connection condition. The "connection condition" refers to the allowable two-dimensional distance between burr fragment images (this distance may be expressed in terms of the number of pixels or the number of fragment images, as with the distance between crack fragment images). If the determination result in S1103 is true, in S1104, the overall determination unit 230 connects the burr fragment images. Specifically, for example, as shown in the figure, if there are one or more non-burr fragment images between the burr fragment images, the overall determination unit 230 changes the determination result (fragment image attribute) of each of the one or more non-burr fragment images from non-burr fragment to burr fragment. As a result, all of the burr fragment images between the burr fragment images become burr fragment images, and therefore, multiple burr fragment images are continuous without any gaps between the non-burr fragment images. If there is a distance between the burr fragment images and the distance is less than the allowable distance, the overall determination unit 230 may consider the distance to be part of the burr. Even if there are one or more non-burr fragment images between the burr fragment images, if the distance between the burr fragment images is less than the allowable distance, the overall determination unit 230 may change the determination result of the fragment image for each of the one or more non-burr fragment images to burr. In S1106, the overall determination unit 230 determines whether the burr area and / or burr density identified from one or more burr fragment images arranged in a two-dimensional direction satisfy the area / density condition (condition related to area and / or density). The "burr area" is the area of ​​the burr, and the "burr density" is the density of the burr. The area / density condition may be a condition identified from the work specification information 270 (for example, a condition based on the work specifications corresponding to the customer). If the determination result in S1106 is true, in S1107, the overall determination unit 230 determines that the type of defect in the test image is a burr (determines that a burr is captured in the test image).

[0098] The flow of the small hole overall determination process is as shown in FIG. 12. That is, the overall determination unit 230 selects a fragment image from the test image (S1201) and determines whether the selected fragment image is a small hole fragment image (S1202). If the determination result of S1202 is true (S1202: YES), the overall determination unit 230 temporarily classifies the small hole fragment image selected in S1201 as a small hole (S1203) and determines whether all fragment images have been selected (S1204). If the determination result of S1204 is true (S1204: YES), the overall determination unit 230 determines whether the diameter and / or roundness specified from the small hole fragment image satisfy the diameter / roundness condition (condition related to diameter and / or roundness) (S1205). The diameter / roundness condition may be a condition specified from the work specification information 270 (for example, a condition based on a work specification corresponding to a customer). If the determination result of S1205 is true (S1205: YES), the overall determination unit 230 determines that the type of defect in the test image is a small hole (determines that a small hole is captured in the test image) (S1206).If the determination result of S1205 is false (S1205: NO), the overall determination unit 230 determines that the type of defect in the test image is a non-small hole (S1207).

[0099] The overall judgment unit 230 may output a judgment result based on the results of the four overall judgment processes. The display control unit 240 may display an inspection result information inspection result screen including information representing the judgment result on the display device 540. The inspection result may include a judgment result as to whether or not a defect of a predetermined defect type is captured in the test image, and a reason for the judgment result, including whether or not the conditions for determining that a defect of the predetermined defect type is captured are satisfied. The "predetermined defect type" may be at least one of vertical cracks, horizontal cracks, burrs, and small holes, and in particular, vertical cracks, horizontal cracks, and burrs, in which fragmentary image connections are made in the overall judgment process. [Third embodiment]

[0100] The third embodiment of the present invention will be described below, focusing mainly on the differences from the first or second embodiment, and explanations of the commonalities with the first or second embodiment will be omitted or simplified.

[0101] In any of the first to third embodiments, the model management unit 250 may determine whether to continue using the deep learning model based on the reliability obtained from the deep learning model for the determination result of each of the multiple image fragments. For example, the process described with reference to FIG. 8 may be performed for each of the models 260A to 260D.

[0102] Furthermore, in this embodiment, if the amount of training data including each fragment image and the type corresponding to the fragment image is less than a certain amount, the training data may classify each fragment image into one of two or more detailed types belonging to the predetermined defect type or one of two or more detailed types belonging to the non-defect type, and the model management unit 250 may use the training data to train a deep learning model. This may be performed for each of the models 260A to 260D, for example.

[0103] Specifically, for example, the process shown in Fig. 13 may be performed for each of models 260A to 260D. Taking one deep learning model as an example, the model management unit 250 determines whether the amount of training data for the deep learning model is sufficient (greater than or equal to a threshold) (S1301).

[0104] If the determination result in S1301 is true (S1301: YES), the model management unit 250 performs small-class learning (S1302). "Small-class learning" means that, in the training data used for learning, for both predetermined defect types and non-defect types (types other than the predetermined defect types), the types prepared are the predetermined defect types themselves, non-defect types themselves, or a small number of types.

[0105] On the other hand, if the determination result in S1301 is false (S1301: NO), the model management unit 250 performs multi-class learning (S1303). "Multi-class learning" means that the number of types prepared in the training data used for learning, both for the predetermined defect type and for non-defect types (types other than the predetermined defect type), is greater than the number of types prepared in the training data used for narrow-class learning. When the amount of data in the training data is insufficient, the deep learning model is expected to be highly accurate considering the amount of data in the training data because the image fragments are classified into more detailed types for both the predetermined defect type and the non-defect type (in other words, for the predetermined defect type, multiple detailed types are associated with each coarse type, such as a non-defect type, and the image fragments are classified into the detailed types).

[0106] Although several embodiments have been described above, these are merely examples for explaining the present invention, and it is not intended that the scope of the present invention be limited to these embodiments. The present invention can be implemented in various other forms. [Explanation of symbols]

[0107] 450...Inspection equipment

Claims

1. An individual determination unit that determines the type of each of a plurality of fragment images extracted from a test image of a workpiece by inputting the fragment image into a learning model that takes an image as input and outputs a type; an overall determination unit that determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images is a predetermined defect type; Equipped with the predetermined defect type is a crack, The learning model is a deep learning model for cracking; The fragment image whose type is determined to be a crack is a crack fragment image, the overall determination unit determines that a crack is present in the test image when a crack length identified from one or more crack fragment images arranged in a one-dimensional direction satisfies a crack length condition, which is a condition related to length; Even if there are one or more non-crack fragment images between crack fragment images, if the distance between the crack fragment images is less than an allowable distance, the overall determination unit changes the determination result of the fragment image for each of the one or more non-crack fragment images to crack. Inspection equipment.

2. An individual determination unit that determines the type of each of a plurality of fragment images extracted from a test image of a workpiece by inputting the fragment images into a learning model that takes an image as input and outputs a type; an overall determination unit that determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images is a predetermined defect type; Equipped with the predetermined defect type is a burr, The learning model is a deep learning model for Sasakure, The image fragment whose type is determined to be ragged is a ragged image fragment, The overall determination unit determines that a burr is present in the test image when the burr area and / or burr density identified from one or more burr fragment images arranged in a two-dimensional direction satisfy an area / density condition, which is a condition related to area and / or density. Inspection equipment.

3. The overall determination unit determines that a burr is present in the test image when the burr area and / or burr density identified from two or more burr fragment images consecutive in a two-dimensional direction satisfy the area / density condition. The inspection device according to claim 2 .

4. If there is a distance between the fragment images of the burr and the distance is less than an allowable distance, the whole determination unit considers the distance to be a part of the burr. The inspection device according to claim 3 .

5. Even if there are one or more non-burr fragment images between the burr fragment images, if the distance between the burr fragment images is less than the allowable distance, the overall determination unit changes the determination result of the fragment image for each of the one or more non-burr fragment images to burr. The inspection device according to claim 4.

6. An individual determination unit that determines the type of each of a plurality of fragment images extracted from a test image of a workpiece by inputting the fragment images into a learning model that takes an image as input and outputs a type; an overall determination unit that determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images is a predetermined defect type; Model Management Department and Equipped with the learning model is a deep learning model, The model management unit determines whether to continue using the deep learning model based on the reliability obtained from the deep learning model for the determination results of each of the plurality of image fragments. Inspection equipment.

7. An individual determination unit that determines the type of each of a plurality of fragment images extracted from a test image of a workpiece by inputting the fragment images into a learning model that takes an image as input and outputs a type; an overall determination unit that determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images is a predetermined defect type; Equipped with The workpiece is a ceramic workpiece. Inspection equipment.

8. An individual determination unit that determines the type of each of a plurality of fragment images extracted from a test image of a workpiece by inputting the fragment images into a learning model that takes an image as input and outputs a type; an overall determination unit that determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images is a predetermined defect type; Equipped with a learning model corresponding to each of a plurality of defect types including the predetermined defect type; For each of the plurality of defect types, the individual determination unit determines the type of each of the plurality of image fragments by inputting the plurality of image fragments extracted from the test image into a learning model corresponding to the defect type; the overall determination unit determines whether or not a defect corresponding to the defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images corresponds to the defect type. Inspection equipment.

9. An individual determination unit that determines the type of each of a plurality of fragment images extracted from a test image of a workpiece by inputting the fragment images into a learning model that takes an image as input and outputs a type; an overall determination unit that determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images is a predetermined defect type; Model Management Department and Equipped with the learning model is a deep learning model, If the amount of training data including each fragment image and the corresponding type of the fragment image is less than a certain amount, In the training data, for each fragmented image in the training data, the fragmented image is classified into one of two or more detailed types belonging to the predetermined defect type, or one of two or more detailed types belonging to a non-defect type, The model management unit uses the training data to train the deep learning model. Inspection equipment.

10. A display control unit is further provided which displays the test results based on test result information including information representing the result of the judgment by the overall judgment unit, The test results are a determination result as to whether or not the defect of the predetermined defect type is captured in the test image; the reason for the determination result, including whether or not a condition for determining that a defect of the predetermined defect type is captured is satisfied; and Including, 10. The inspection device according to claim 1, 2, 6, 7, 8, or 9.

11. The conditions for determining that a defect of the predetermined defect type is captured are conditions defined in work specifications corresponding to the customer of the work, among work specifications defined for each customer to whom the work is provided.

10. The inspection device according to claim 1, 2, 6, 7, 8, or 9.

12. (A) A computer inputs a plurality of fragment images extracted from a test image of a workpiece into a learning model that takes an image as input and outputs a type, thereby determining the type of each of the plurality of fragment images; (B) the computer determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; the predetermined defect type is a crack, The learning model is a deep learning model for cracking; The fragment image whose type is determined to be a crack is a crack fragment image, In (B), even if there are one or more non-crack fragment images between the crack fragment images, if the distance between the crack fragment images is less than the allowable distance, The computer changes the determination result of the fragment image for each of the one or more non-crack fragment images to a crack, The computer determines that a crack is present in the test image when a crack length identified from one or more crack fragment images arranged in a one-dimensional direction satisfies a crack length condition, which is a condition related to length. Testing method.

13. (A) A computer inputs a plurality of fragment images extracted from a test image of a workpiece into a learning model that takes an image as input and outputs a type, thereby determining the type of each of the plurality of fragment images; (B) the computer determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; the predetermined defect type is a burr, The learning model is a deep learning model for Sasakure, The image fragment whose type is determined to be ragged is a ragged image fragment, In (B), when the burr area and / or burr density identified from one or more burr fragment images arranged in a two-dimensional direction satisfy an area / density condition, which is a condition related to area and / or density, the computer determines that a burr is present in the test image. Testing method.

14. (A) A computer inputs a plurality of fragment images extracted from a test image of a workpiece into a learning model as a deep learning model that takes an image as input and outputs a type, thereby determining the type of each of the plurality of fragment images; (B) the computer determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; (C) the computer determines whether to continue using the deep learning model based on the reliability obtained from the deep learning model for the judgment results of each of the plurality of image fragments. Testing method.

15. (A) A computer inputs a plurality of fragment images extracted from a test image of a ceramic workpiece into a learning model that takes an image as input and outputs a type, thereby determining the type of each of the plurality of fragment images; (B) the computer determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type. Testing method.

16. (A) A computer inputs a plurality of fragment images extracted from a test image of a workpiece into a learning model that takes an image as input and outputs a type, thereby determining the type of each of the plurality of fragment images; (B) the computer determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; a learning model corresponding to each of a plurality of defect types including the predetermined defect type; For each of the plurality of defect types, (A), the computer inputs a plurality of image fragments extracted from the test image into a learning model corresponding to the defect type, and determines the type of each of the plurality of image fragments; (B) the computer determines whether or not a defect corresponding to the defect type is captured in the test image based on whether or not the determined type of each of the plurality of fragment images corresponds to the defect type; Testing method.

17. A computer inputs a plurality of fragment images extracted from a test image of a workpiece into a learning model as a deep learning model that takes an image as input and outputs a type, thereby determining the type of each of the plurality of fragment images; (B) the computer determines whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; (C) When the amount of training data including the fragment image and the type corresponding to the fragment image for each fragment image is less than a certain amount, In the training data, for each fragmented image in the training data, the fragmented image is classified into one of two or more detailed types belonging to the predetermined defect type, or one of two or more detailed types belonging to a non-defect type, The computer uses the training data to train the deep learning model. Testing method.

18. (A) A learning model that takes an image as input and outputs a type, and inputs a plurality of fragment images extracted from a test image of a workpiece, thereby determining the type of each of the plurality of fragment images; (B) determining whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; Let the computer do that, the predetermined defect type is a crack, The learning model is a deep learning model for cracking; The fragment image whose type is determined to be a crack is a crack fragment image, In (B), even if there are one or more non-crack fragment images between the crack fragment images, if the distance between the crack fragment images is less than the allowable distance, changing the determination result of the fragment image for each of the one or more non-crack fragment images to a crack; If a crack length identified from one or more crack fragment images arranged in a one-dimensional direction satisfies a crack length condition, which is a condition related to length, it is determined that a crack is captured in the test image. A computer program for causing the computer to execute the above.

19. (A) A learning model that takes an image as input and outputs a type, and inputs a plurality of fragment images extracted from a test image of a workpiece, thereby determining the type of each of the plurality of fragment images; (B) determining whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; Let the computer do that, the predetermined defect type is a burr, The learning model is a deep learning model for Sasakure, The image fragment whose type is determined to be ragged is a ragged image fragment, In (B), if the burr area and / or burr density identified from one or more burr fragment images arranged in a two-dimensional direction satisfy an area / density condition, which is a condition related to the area and / or density, it is determined that a burr is included in the test image. A computer program for causing the computer to execute the above.

20. (A) A learning model is a deep learning model that takes an image as input and outputs a type, and inputs a plurality of fragment images extracted from a test image of a workpiece, thereby determining the type of each of the plurality of fragment images; (B) determining whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; (C) determining whether to continue using the deep learning model based on the reliability obtained from the deep learning model for the judgment results of each of the plurality of image fragments; A computer program that causes a computer to do something.

21. (A) A learning model that inputs an image and outputs a type receives a plurality of fragment images extracted from a test image of a ceramic workpiece, and determines the type of each of the plurality of fragment images; (B) determining whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; A computer program that causes a computer to do something.

22. (A) A learning model that takes an image as input and outputs a type, and inputs a plurality of fragment images extracted from a test image of a workpiece, thereby determining the type of each of the plurality of fragment images; (B) determining whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; Let the computer do that, a learning model corresponding to each of a plurality of defect types including the predetermined defect type; For each of the plurality of defect types, (A) inputting a plurality of image fragments extracted from the test image into a learning model corresponding to the defect type, and determining the type of each of the plurality of image fragments; (B) determining whether a defect corresponding to the defect type is included in the test image based on whether the determined type of each of the plurality of fragment images corresponds to the defect type; A computer program that causes a computer to do something.

23. (A) A learning model is a deep learning model that takes an image as input and outputs a type, and inputs a plurality of fragment images extracted from a test image of a workpiece, thereby determining the type of each of the plurality of fragment images; (B) determining whether or not a defect of the predetermined defect type is captured in the test image based on whether or not the determined type of each of the plurality of image fragments is a predetermined defect type; (C) When the amount of training data including the fragment image and the type corresponding to the fragment image for each fragment image is less than a certain amount, In the training data, for each fragmented image in the training data, the fragmented image is classified into one of two or more detailed types belonging to the predetermined defect type, or one of two or more detailed types belonging to a non-defect type, The deep learning model is trained using the training data. A computer program that causes a computer to do something.

Images