Appearance screening system, appearance screening method, and appearance screening program

Abstract

An appearance screening system 100 includes an imaging device 10 configured to capture an inspection object W and generate an inspection image Wi, a local computer 1 having a first discrimination unit 11 configured to perform a first good / bad product discrimination on the inspection image Wi locally based on a first inspection scheme 11R, a server computer 2 having a second discrimination unit 21 configured to perform a second good / bad product discrimination on the inspection image Wi in the cloud based on a second inspection scheme 21R, and a parallel processing unit 24 configured to perform the first good / bad product discrimination and the second good / bad product discrimination in parallel.

Description

Technical Field

[0001] This invention relates to an appearance screening system, appearance screening method, and appearance screening procedure. Background Technology

[0002] Previously, there was an appearance screening machine that determined OK and NG products (good / defective product discrimination) based on images of the product and screened out good products. The machine would classify the images of NG products into more detailed NG patterns and label them (multi-label classification), and use the labeling results to improve the performance of the inspection scheme (e.g., Patent Document 1).

[0003] Traditional appearance sorting machines require setting thresholds to keep inspection results within a safe range, meeting the requirements of low-cost, high-speed inspection while avoiding missed defects. Furthermore, besides being limited by the specifications of the image processing PC and the image processing time constraints imposed by the production equipment's cycle time, there are also limitations due to over-detection (misclassification). Therefore, it is difficult to accurately implement multi-label classification during product manufacturing, and the sorting results obtained by users from traditional appearance sorting machines may be inaccurate.

[0004] Existing technical documents

[0005] Patent documents

[0006] Patent Document 1: Japanese Patent No. 6999150 Summary of the Invention

[0007] The problem the invention aims to solve

[0008] The purpose of this invention is to provide an appearance screening system, appearance screening method, and appearance screening procedure that can more accurately screen out good products.

[0009] Solution for solving the problem

[0010] (1) An appearance screening system according to one aspect of the present invention includes: an imaging device for capturing an inspection object and generating an inspection image; a local computer having a first discrimination unit for performing a first good / defective product discrimination on the inspection image locally based on a first inspection scheme; a server computer having a second discrimination unit for performing a second good / defective product discrimination on the inspection image in the cloud based on a second inspection scheme; and a parallel processing unit for performing the first good / defective product discrimination and the second good / defective product discrimination in parallel.

[0011] (2) In (1) above, when the first good / defective product discrimination result is a defective product, the first discrimination unit performs a first multi-label classification on the inspection image, classifies it as a defective product mode and marks it; when the second good / defective product discrimination result is a defective product, the second discrimination unit performs a second multi-label classification on the inspection image, classifies it as the defective product mode and marks it; it may also include a comparison and extraction unit, which is used to compare the first classification result of the first multi-label classification and the second classification result of the second multi-label classification, and extract the inspection images that are inconsistent between the two.

[0012] (3) In (1) or (2) above, the first inspection scheme can be constructed by synthesizing rule base discrimination logic based on predefined rules and local AI logic implemented by trained discrimination AI.

[0013] (4) In (3) above, when the first good / defective product judgment result is different from the second good / defective product judgment result, the second good / defective product judgment result can be output.

[0014] (5) An appearance screening method according to one aspect of the present invention includes: an imaging step, capturing an inspection object and generating an inspection image; a first discrimination step, based on a first inspection scheme, performing a first good / defective product discrimination on the inspection image locally via a local computer; a second discrimination step, based on a second inspection scheme, performing a second good / defective product discrimination on the inspection image via a server computer in the cloud; and performing the first good / defective product discrimination and the second good / defective product discrimination in parallel.

[0015] (6) According to one aspect of the present invention, an appearance screening procedure is used to enable a computer to perform: an imaging function to capture an inspection object and generate an inspection image; a first discrimination function to perform a first good / defective product discrimination on the inspection image locally based on a first inspection scheme; a second discrimination function to perform a second good / defective product discrimination on the inspection image in the cloud based on a second inspection scheme; and a parallel processing function to execute the first good / defective product discrimination and the second good / defective product discrimination in parallel.

[0016] The effects of the invention

[0017] According to the present invention, an appearance screening system, appearance screening method, and appearance screening procedure that can more accurately screen good products can be provided. Attached Figure Description

[0018] Figure 1 This is a diagram showing an outline of the appearance screening system according to the first embodiment.

[0019] Figure 2 This is a diagram illustrating the process of the appearance screening method.

[0020] Figure 3 This is an illustrative diagram comparing the first and second multi-label classification methods.

[0021] Figure 4 This is a diagram showing an outline of the appearance screening system according to the second embodiment. Detailed Implementation

[0022] (First Implementation)

[0023] The first embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Figure 1 This is a diagram showing an outline of the appearance screening system 100 according to the first embodiment. Figure 2 This is a diagram illustrating the process of the appearance screening method. Figure 3 This is an illustrative diagram comparing the first and second multi-label classification methods. Furthermore, in the following text, parts with common functions may be assigned the same reference numbers or symbols.

[0025] (Appearance screening system)

[0026] The appearance screening system 100 according to the first embodiment can be used, for example, to inspect the appearance of products such as electronic components such as capacitors. The appearance screening system 100 can screen products as good (also known as "OK products") or bad (also known as "NG products") based on their appearance.

[0027] like Figure 1 As shown, the appearance screening system 100 includes: an imaging device 10 for capturing an inspection object W and generating an inspection image Wi; a local computer 1 having a first discrimination unit 11 for performing a first good / defective product discrimination on the inspection image Wi locally based on a first inspection scheme 11R; and a server computer 2 having a second discrimination unit 21 for performing a second good / defective product discrimination on the inspection image Wi in the cloud based on a second inspection scheme.

[0028] Local computer 1 is, for example, built into an inspection device installed along the production line of the object to be inspected, W. Local computer 1 is connected to server computer 2 via network N through communication unit 12 of local computer 1.

[0029] Imaging device 10 is an optical device for taking photographs. Imaging device 10 can be a so-called digital camera, which uses an imaging component such as a CCD or CMOS sensor as a semiconductor sensor to receive light transmitted through a lens and convert it into image data. Imaging device 10 is installed, for example, in an inspection apparatus installed along a production line of the object to be inspected, W. Imaging device 10 photographs the product (part) that is the object to be inspected, W. Imaging device 10 appropriately photographs the same product from multiple directions (e.g., six directions) and acquires multiple images. The captured inspection images Wi are sent to the first discrimination unit 11 of the local computer 1 or to a memory (not shown).

[0030] The first discrimination unit 11 can determine whether the inspected object W is a good product or a defective product through a rule-based discrimination logic R1 based on predefined rules, thereby performing a first good / defective product discrimination. This rule-based discrimination logic R1 is designed based on rules predefined according to the specifications or type of the inspected object W. For example, the rule-based discrimination logic R1 can be an integration of algorithms for detecting defective parts, set for each defect item (stains, notches, foreign objects, pinholes, etc.). Figure 1 As shown, the first inspection scheme 11R may include rule base discrimination logic R1. This reduces the load on the inspection processing (image processing, good / defective product discrimination processing, multi-label classification processing) performed by the first discrimination unit 11.

[0031] The first discrimination unit 11 can perform a first good / defective product classification on the inspected image Wi using local AI logic R2. The local AI logic R2 can be, for example, logistic regression, support vector machine, k-nearest neighbor algorithm, decision tree, or neural network deep learning. This allows for more accurate inspection processing by the first discrimination unit 11, especially for multi-label classification which has a heavy processing load.

[0032] The first good / defective product determination of the first discrimination unit 11 may include multi-label classification, which involves dividing images determined to be defective into more specific defect items (also known as "defective product patterns") and labeling them. That is, in the first good / defective product determination of the first discrimination unit 11, such as... Figure 3 As shown, the first good / defective product discrimination can not only determine whether a product is good or defective, but also classify defective products into more specific defect items and mark them.

[0033] The first inspection scheme 11R can be constructed by synthesizing a rule-based discrimination logic R1 based on predefined rules and an AI logic R2 implemented by a trained discrimination AI. For example, if an image that should be judged as a good product is judged as a defective product by the rule-based discrimination logic R1, and if the local AI logic R2 performs a good / defective product judgment and judges it as a good product, then the first inspection scheme 11R will judge the image as a good product. This can suppress over-detection of defective products generated in the first good / defective product judgment performed by the first discrimination unit.

[0034] Server computer 2 responds to resource requests from local computer 1. Server computer 2 is connected to local computer 1 via network N through its communication unit 22.

[0035] The second discrimination unit 21 can use cloud-based AI logic G to classify the inspected image Wi as good or bad. The cloud-based AI logic G can be, for example, logistic regression, support vector machine, k-nearest neighbor algorithm, decision tree, or neural network deep learning. This enables more accurate multi-label classification.

[0036] The second good / defective product determination by the second determination unit 21 can be performed by determining whether a product is good or defective based on a second inspection plan 21R pre-set according to the specifications or type of the inspected object W. The second good / defective product determination by the second determination unit 21 may include multi-label classification, which divides images determined to be defective into more specific defect patterns and marks them. That is, as... Figure 3 As shown, the second good / defective product discrimination can not only determine whether a product is good or defective, but also classify defective products into more specific defect items (defective product patterns) and mark them.

[0037] Here, the appearance screening system 100 includes a parallel processing unit 24 for performing the first good / defective product discrimination and the second good / defective product discrimination in parallel. For example... Figure 1As shown, the parallel processing unit 24 can be installed on the server computer 2, the local computer 1, or another computer (not shown) on the network N. Based on the detection of images captured by the imaging device 10, the parallel processing unit 24 can perform a first good / defective product judgment and a second good / defective product judgment in parallel by the first judgment unit 11 and the second judgment unit 21. Parallel execution of good / defective product judgment means that two judgment processes are performed simultaneously. This allows time-consuming processing to be distributed to the two processors on the local computer 1 and the server computer 2 respectively, efficiently advancing the processing and improving the overall processing speed, while combining the two judgment results for screening. Furthermore, this eliminates the need to rely on the performance-limited local computer 1, allowing the judgment processing to be performed by the high-performance server computer 2. This maintains the device's cycle time while the local computer 1 performs judgment processing, and the server computer 2 performs supplementary judgment processing, enabling high-precision screening by combining the two judgment results. In this way, even with the limitations of the image processing PC's specifications and the image processing time constraints imposed by the production equipment's cycle time, the appearance screening system 100 can more accurately screen out good products.

[0038] Compared to server computer 2, local computer 1 can employ relatively lower specifications in terms of image processing performance. For example, the GPU in local computer 1 can be of lower specifications compared to the GPU in server computer 2. This reduces the allocation rate of the processing load for the inspection image Wi in local computer 1, which suffers from limited performance scalability due to its installation on the manufacturing equipment used to inspect the product W, and increases the allocation rate of the processing load for the inspection image Wi in server computer 2, which has sufficient performance scalability. Consequently, the inspection image Wi can be processed efficiently.

[0039] Furthermore, to further improve the multi-label classification performance of the local AI logic R2 or the cloud AI logic G, it is effective to have the local AI logic R2 or the cloud AI logic G perform machine learning on images with a large number of labels. However, users (mainly customers using appearance screening services) need to visually inspect a large number of collected images one by one and accurately label them (assigning good or bad labels and coloring) for the local AI logic R2 or the cloud AI logic G to perform machine learning. Moreover, the labeling work usually requires a lot of labor.

[0040] Therefore, given that some of the collected images contribute to improving multi-label classification performance while others do not, a feature was designed that extracts only the images that contribute significantly to machine learning, rather than forcing users to label all images. This allows users to label only the extracted images, reducing their workload and facilitating efficient machine learning for local AI logic R2 or cloud AI logic G, thereby improving their multi-label classification performance.

[0041] Specifically, when the first good / defective product discrimination result is a defective product, the first discrimination unit 11 performs a first multi-label classification on the inspection image Wi, classifying it as a defective product and labeling it. When the second good / defective product discrimination result is a defective product, the second discrimination unit 21 performs a second multi-label classification on the inspection image Wi, classifying it as the defective product and labeling it. Furthermore, the appearance screening system 100 also includes a comparison extraction unit 23, used to compare the first classification result of the first multi-label classification and the second classification result of the second multi-label classification, and extract inspection images Wi that are inconsistent between the two. The comparison extraction unit 23 can be installed on the local computer 1, such as... Figure 1 As shown, it can also be set on server computer 2, or on other computers (not shown) on network N. Thus, the first classification result and the second classification result can be compared, and inspection images Wi that are inconsistent between the two can be extracted. This allows for the selection of images that contribute significantly to machine learning, where a threshold is set within a safe range, and which have resulted in over-detection (misclassification) of good products as defective products. Then, for the selected extracted images Wp, annotation is (appropriately by the user) prioritized, and the local AI logic R2 or the cloud AI logic G learns from the annotated images Wa that have been correctly labeled through annotation. Therefore, the accuracy of good / defective product discrimination and multi-label classification by the first discrimination unit 11 or the second discrimination unit 21 can be improved.

[0042] like Figure 1 As shown by the dashed line, the local AI logic R2 or the cloud AI logic G can learn from the annotated image Wa generated after adding labels to the extracted image Wp extracted by the annotation comparison extraction unit 23. Therefore, without placing excessive burden on the annotation process, the performance of the first discriminant unit 11 (containing the local AI logic R2) or the second discriminant unit 21 (containing the cloud AI logic G) in good / defect product discrimination or multi-label classification can be effectively improved. It should be noted that based on the annotated image Wa, the model parameters or hyperparameters can be modified by searching for parameters through optimization algorithms or by manual adjustment to improve the rule base discriminant logic R1.

[0043] When the first good / defective product judgment result for inspection image Wi differs from the second good / defective product judgment result, the appearance screening system 100 can output the second good / defective product judgment result. The good / defective product judgment result for inspection image Wi is output to an output unit provided by the appearance screening system 100. The output unit can be installed on the local computer 1, on the server computer 2, or on another computer (not shown) on the network N. The output to the output unit can take forms such as displaying on a monitor or printing via a printer. Therefore, based on the premise that the second good / defective product judgment result processed on the relatively high-specification server computer 2 has higher reliability than the first good / defective product judgment result processed on the relatively low-specification local computer 1, a more reliable result can be used.

[0044] (Appearance screening method)

[0045] Next, the process of the appearance screening method will be described. The appearance screening method can be implemented using the appearance screening system 100. Each step of the appearance screening method can be executed by the corresponding component in the appearance screening system 100. Figure 2 This is a diagram illustrating the process of the appearance screening method. Figure 3 This is an illustrative diagram comparing the first and second multi-label classification methods.

[0046] (1) such as Figure 2 As shown, firstly, the object to be inspected, W, is photographed by the imaging device 10 and an inspection image Wi is generated (imaging step S1).

[0047] (2) Next, the captured inspection image Wi is sent to the first discrimination unit 11 of the local computer 1. At the same time, the communication unit 12 of the local computer 1 sends the inspection image Wi to the second discrimination unit 21 of the server computer 2 through the communication unit 22 of the server computer 2.

[0048] (3) Based on the first inspection scheme 11R, the first discrimination unit 11 of the local computer 1 performs a first good / defective product discrimination on the inspection image Wi (first discrimination step S2).

[0049] (3-1) In detail, if the result of the first good / defective product judgment is a defective product, the first judgment unit 11 classifies the inspection image Wi into a defective product pattern and labels it (first multi-label classification step S2-1).

[0050] (4) Based on the second inspection scheme 21R, the second discrimination unit 21 of the server computer 2 performs a second good / defective product discrimination on the inspection image Wi in the cloud (second discrimination step S3).

[0051] (4-1) In detail, if the result of the first good / defective product judgment is a defective product, the second judgment unit 21 classifies the inspection image Wi into a defective product pattern and labels it (second multi-label classification step S3-1).

[0052] (5) Here, the first discrimination step S2 and the second discrimination step S3 are executed in parallel. For example, as follows Figure 3 As shown, in the first discrimination unit 11, a first good / defective product discrimination is performed on the inspection image Wi, such as... Figure 3 As shown in the table, the first classification result, displaying the detection results for each defective item, is obtained. Simultaneously, in the second discrimination unit 21, a second good / defective product discrimination is performed on the inspection image Wi, as shown... Figure 3 As shown in the table, a second classification result is obtained, displaying the detection results for each defective item. This allows for the allocation of time-consuming processing to the two processors on local computer 1 and server computer 2 respectively, efficiently advancing the processing and improving overall processing speed. Furthermore, it eliminates the need to rely on the performance-limited local computer 1, allowing for the use of the high-performance server computer 2 for discrimination processing. This maintains the equipment cycle time while local computer 1 performs discrimination processing, allowing for supplementary discrimination processing by server computer 2, thus enabling high-precision screening by combining the two results. In this way, even with limitations imposed by the image processing PC's specifications and the image processing time constraints imposed by the production equipment's cycle time, the appearance sorting method can more accurately screen out good products.

[0053] (6) Next, compare the first classification result of the first multi-label classification with the second classification result of the second multi-label classification (comparison step S4).

[0054] (7) In comparison step S4, if the comparison result shows that the defect items of the first classification result and the second classification result are consistent, then output any classification result, such as the first classification result (consistent result output step S5).

[0055] (7-1) In comparison step S4, if the comparison result shows that the defect items of the first classification result and the second classification result are inconsistent, then the inspection image Wi, which is the inspection object, is extracted as the extraction image Wp (extraction step S6). For example, as Figure 3 As shown, comparing the first and second classification results obtained from judging good / defective products on the inspection image Wi, as follows: Figure 3 As shown in the table, consistency or inconsistency is determined based on the defect items. Furthermore, if an inspection image Wi contains a defect item that is determined to be inconsistent, then the inspection image Wi is extracted as an extraction image Wp.

[0056] (8) The user appropriately annotates the extracted inspection images Wi. In this way, only inspection images Wi with inconsistent defect items are extracted, so the extracted inspection images Wi can be easily and efficiently annotated for machine learning.

[0057] (8-1) Here, the local AI logic R2 or the cloud AI logic G can be appropriately allowed to learn from the annotated image Wa that has been annotated on the extracted image Wp. This can improve the accuracy of good / defective product discrimination and multi-label classification.

[0058] (9) In comparison step S4, if the comparison result shows that the defect items of the first classification result and the second classification result are inconsistent, the second classification result is output first (step S7 is output first). Thus, the over-detection of defect items by the first good / defective product discrimination is corrected by the second good / defective product discrimination, so that a reliable inspection result can be output.

[0059] (Appearance screening procedure)

[0060] The appearance screening program has the function of enabling the computer to perform each of the processes in each step of the appearance screening method described above. Specifically, according to each step of the appearance screening method, the appearance screening program enables the computer to perform: an imaging function to capture an image of the object to be inspected, W, and generate an inspection image, Wi; a first discrimination function to perform a first good / defective discrimination on the inspection image Wi locally based on a first inspection scheme 11R; a second discrimination function to perform a second good / defective discrimination on the inspection image Wi in the cloud based on a second inspection scheme 21R; and a parallel processing function to execute the first good / defective discrimination and the second good / defective discrimination in parallel. It should be noted that the appearance screening program can be installed on the server computer 2, the local computer 1, or other computers (not shown) on the network N. This allows time-consuming processing to be distributed to the two processors on the local computer 1 and the server computer 2 respectively, efficiently advancing the computation and improving the overall processing speed, while also combining the two discrimination results for screening. Furthermore, this eliminates the need to rely on a local computer 1 with limited performance (specifications). Instead, a server computer 2 with sufficient performance can be used for the discrimination process. This allows for supplementary discrimination processing by the server computer 2 while maintaining the equipment's cycle time. Consequently, the results from both discrimination processes can be combined for high-precision selection. Thus, even with limitations imposed by the image processing PC's specifications and the image processing time constraints imposed by the production equipment's cycle time, the appearance screening process can more accurately identify qualified products.

[0061] (Second Implementation)

[0062] Next, the appearance screening system 200 according to the second embodiment will be described. Figure 4 This is a diagram showing an outline of the appearance screening system 200 according to the second embodiment. Hereinafter, in the appearance screening system 200 according to the second embodiment, elements having the same function as those in the appearance screening system 100 according to the first embodiment will be given the same reference numerals or symbols.

[0063] In the appearance screening system 100 according to the first embodiment, the case where the local AI logic R2 in the first inspection scheme 11R is learned using the annotated image Wa has been described. However, in order to learn on the local computer 1, the inspection device related to the local computer 1 needs to be disabled, which may take time. In addition, learning on the local computer 1 may cause the content of the local AI logic R2 running on the local computer 1 to be automatically modified, which is unmanageable by the user. Therefore, in order to solve these problems, the appearance screening system 200 according to the second embodiment differs from the appearance screening system 100 according to the first embodiment, mainly as follows.

[0064] like Figure 4 As shown, the appearance screening system 200 according to the second embodiment has a third discrimination unit 25 on the server computer 2 with the same structure as the first discrimination unit 11.

[0065] The third discrimination unit 25 has a third inspection scheme 25R with the same structure as the first inspection scheme 11R.

[0066] In other words, the third inspection scheme 25R includes an alternative rule base discrimination logic R3 with the same structure as the rule base discrimination logic R1 and an alternative local AI logic R4 with the same structure as the local AI logic R2, and is constructed by synthesizing the alternative rule base discrimination logic R3 and the alternative local AI logic R4. For example, if an image that should be judged as a good product is judged as a defective product by the alternative rule base discrimination logic R3, and if the alternative local AI logic R4 judges the image as a good product / defective product and judges it as a good product, then the third inspection scheme 25R judges the image as a good product.

[0067] Furthermore, instead of passing the annotated image Wa to the local AI logic R2, the user passes it to the alternative local AI logic R4, which then learns from it.

[0068] When the user determines that the performance of the third inspection scheme 25R is significantly higher than that of the first inspection scheme 11R, the user may appropriately replace the first inspection scheme 11R in the local computer 1 with the third inspection scheme 25R at any time.

[0069] At this time, the comparison extraction unit 23 can replace the second discrimination unit 21 to compare the first discrimination unit 11 and the third discrimination unit 25. That is, in the appearance screening system 200, the comparison extraction unit 23 compares the first classification result of the first multi-label classification performed by the first discrimination unit 11 with the third classification result of the third multi-label classification performed by the third discrimination unit 25, and extracts the image Wp that is inconsistent between the two.

[0070] Thus, according to the appearance screening system 200 of the second embodiment, the server computer 2 is provided with a third discrimination unit 25 having the same structure as the first discrimination unit 11. Therefore, instead of transmitting the annotated image Wa to the local AI logic R2 in the local computer 1, it can be transmitted to the alternative local AI logic R4 included in the third discrimination unit 25 in the server computer 2, where it will be learned. Consequently, without disabling the inspection devices associated with the local computer 1, the user can update the first discrimination unit 11 of the local computer 1 at appropriate times.

[0071] It should be noted that the technical scope of the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention. Furthermore, without departing from the spirit of the present invention, the constituent elements in the above embodiments can be appropriately replaced with well-known constituent elements. In addition, the above variations can be appropriately combined without departing from the spirit of the present invention.

[0072] As described above, the appearance screening system 100, 200 according to the embodiment includes: an imaging device 10, which captures an inspection object W and generates an inspection image Wi; a local computer 1, having a first discrimination unit 11, which performs a first good / defective product discrimination on the inspection image Wi locally based on a first inspection scheme 11R; a server computer 2, having a second discrimination unit 21, which performs a second good / defective product discrimination on the inspection image Wi in the cloud based on a second inspection scheme 21R; and a parallel processing unit 24, which performs the first good / defective product discrimination and the second good / defective product discrimination in parallel. Thus, time-consuming processing can be distributed to the two processors respectively owned by the local computer 1 and the server computer 2, efficiently advancing the processing, thereby improving the overall processing speed, and combining the two discrimination results for screening. Furthermore, this eliminates the need to rely on a local computer 1 with limited performance (specifications). Instead, a server computer 2 with sufficient performance can be used for the discrimination process. This allows for supplementary discrimination processing by the server computer 2 while maintaining the equipment's cycle time. Consequently, the results from both discrimination processes can be combined for high-precision screening. Thus, even with limitations imposed by the image processing PC's specifications and the image processing time constraints imposed by the production equipment's cycle time, the appearance screening system 100 can more accurately select qualified products.

[0073] The appearance screening method according to the embodiment includes: an imaging step S1, capturing an image of the object to be inspected W and generating an inspection image Wi; a first discrimination step S2, performing a first good / defective product discrimination on the inspection image Wi locally via a local computer 1 based on a first inspection scheme 11R; and a second discrimination step S3, performing a second good / defective product discrimination on the inspection image Wi via a server computer 2 in the cloud based on a second inspection scheme 21R. Furthermore, the first discrimination step S2 and the second discrimination step S3 are executed in parallel in the appearance screening method. Therefore, even with limitations imposed by the specifications of the image processing PC and the image processing time constraints caused by the machine cycle time of the product manufacturing equipment, good products can be screened more accurately.

[0074] The appearance screening procedure according to the embodiment enables the computer to perform: an imaging function to capture an image of the object to be inspected, W, and generate an inspection image, Wi; a first discrimination function to perform a first good / defective product discrimination on the inspection image Wi locally based on a first inspection scheme 11R; a second discrimination function to perform a second good / defective product discrimination on the inspection image Wi in the cloud based on a second inspection scheme 21R; and a parallel processing function to execute the first good / defective product discrimination and the second good / defective product discrimination in parallel. Therefore, even with limitations imposed by the specifications of the image processing PC and the image processing time constraints caused by the machine cycle time of the product manufacturing equipment, good products can be screened more accurately.

[0075] Explanation of reference numerals in the attached figures

[0076] 100, 200: Appearance screening system; 1: Local computer; 10: Imaging device; 11: First discrimination unit; 11R: First inspection scheme; 12: Communication unit; 2: Server computer; 21: Second discrimination unit; 21R: Second inspection scheme; 22: Communication unit; 23: Comparison and extraction unit; 24: Parallel processing unit; 25: Third discrimination unit; 25R: Third inspection scheme; N: Network; R1: Rule base discrimination logic; R2: Local AI logic; R3: Alternate rule base discrimination logic; R4: Alternate local AI logic; G: Cloud AI logic; W: Inspection object; Wi: Inspection image; Wp: Image extraction; Wa: Image annotation.

Claims

1. An appearance screening system, comprising: Imaging device, used to capture images of the object being inspected and generate inspection images; The local computer has a first discrimination unit, which is used to perform a first good / defective product discrimination on the inspection image locally based on a first inspection scheme; The server computer has a second discrimination unit, which is used to perform a second good / defective product discrimination on the inspection image in the cloud based on a second inspection scheme; The parallel processing unit is used to perform the first good / defective product determination and the second good / defective product determination in parallel.

2. The appearance screening system according to claim 1, When the first good / defective product discrimination result is a defective product, the first discrimination unit performs a first multi-label classification on the inspection image, classifies it as a defective product mode and labels it; When the second good / defective product discrimination result is a defective product, the second discrimination unit performs a second multi-label classification on the inspection image, classifies it as the defective product pattern and labels it; It also includes a comparison and extraction unit, which compares the first classification result of the first multi-label classification and the second classification result of the second multi-label classification, and extracts inspection images that are inconsistent between the two.

3. The appearance screening system according to claim 1 or 2, The first inspection scheme is composed of a rule-based discrimination logic based on predefined rules and local AI logic implemented by a trained discrimination AI.

4. The appearance screening system according to claim 1 or 2, When the first good / defective product identification result is different from the second good / defective product identification result, the second good / defective product identification result is output.

5. An appearance screening method, comprising: The imaging step involves capturing images of the object being inspected and generating inspection images. The first discrimination step involves, based on the first inspection scheme, performing a first good / defective product discrimination on the inspection image locally using a local computer. The second discrimination step involves, based on the second inspection scheme, performing a second good / defective product discrimination on the inspection image via a server computer in the cloud. The first good / defective product determination and the second good / defective product determination are performed in parallel.

6. An appearance screening program for causing a computer to perform: Imaging function: Captures the object under inspection and generates inspection images; The first discrimination function, based on the first inspection scheme, performs a first good / defective product discrimination on the inspection image locally; The second discrimination function, based on the second inspection scheme, performs a second good / defective product discrimination on the inspection image in the cloud; The parallel processing function executes the first good / defective product determination and the second good / defective product determination in parallel.

Images