TWI842292B - Appearance defect inspection method and system for automated production line products - Google Patents

Abstract

An appearance defect inspection method and an appearance defect inspection system for automated production line products are disclosed. The system includes a test fixture, an image capture device, an automatic learning module and a light source. With learning image files obtained by the image capture device learned by the automatic learning module and these learning image files having been normalized, the system can more accurately predict the defects of the products, and then eliminate the defective products. Model training provided by the present invention is accurate, and also reduces the misjudgment rate when selecting defective products online.

Description

Automated production line product appearance defect inspection method and system

The present invention relates to a method and system for inspecting appearance defects, in particular, a method and system for inspecting appearance defects of automated production line products.

With the popularization of industrial automation equipment, the production process of many products has been standardized, and the manufacturing capacity can be greatly improved. It is conceivable that quality inspection of these large quantities of products is not an easy task, especially when these products are small and the defects are not easy to distinguish. Therefore, the speed of quality inspection work relying solely on human (eye) power can no longer keep up with the production capacity, and additional auxiliary equipment is needed to solve this problem.

Auxiliary inspection equipment has different characteristics according to the characteristics of different products. For example, if you want to inspect large products, such as the internal cracks of precast water pipes, you can use ultrasonic technology; if you want to inspect the ingredients of beverages, you can use chemical composition analysis; if you want to detect visible defects, you can do it through product appearance image analysis or comparison. For the appearance inspection of small products with inconspicuous defects on the automated production line, in addition to the aforementioned image analysis or comparison, automated image processing is also required.

In some prior art of appearance defect inspection, there are some references that can provide solutions. For example, the Republic of China Patent No. I479145 proposes an appearance defect detection system and method, which is used to obtain a current image of a sample to be tested taken by a camera unit of a detection device, the current image including a side image and a projection image of the sample to be tested; perform defect detection on the side image of the sample to be tested, and determine the rotation angle of the sample to be tested according to the projection image of the sample to be tested; when a defect is detected in the side image of the sample to be tested, obtain a corresponding side image template according to the rotation angle of the sample to be tested; compare the side image of the sample to be tested with the obtained side image template to determine the side block where the defect appears in the sample to be tested. This patent is about using image comparison to find out the appearance defects of products. However, the definition of appearance defects is limited to existing data. If the data storage is incomplete, the comparison results may miss some defects. In addition, the Republic of China's invention patent No. I667575 proposes a defect detection system connected to an automatic appearance inspection device, including the following devices: a re-inspection server station receives defect images and defect positions; a training terminal stores trained modules; a classification terminal receives defect images and defect positions, reads a target training module corresponding to the defect image in the trained module, and classifies the defect image according to the target training module to generate and transmit a marked defect image to the re-inspection server station; the re-inspection terminal receives the marked defect image, receives and transmits the verification operation of the corresponding marked defect image to the re-inspection server station; and the marked re-inspection terminal receives the verification operation and the marked defect image, and receives the marking result of the corresponding marked defect image. The re-inspection server station transmits the marking results and the marked defect image to the training terminal so that the training terminal can train the corresponding training module accordingly. The aforementioned patent uses artificial intelligence to learn the appearance defects of defective products, and then finds the defects and picks out the defective products. However, the image usage range of the patent is too large, and the image has not been normalized. In addition to causing inaccuracy in the learning stage of artificial intelligence, it will also lead to an increase in the error rate when selecting defective products online.

In order to solve the problems existing in the prior art, the present invention is proposed.

This paragraph extracts and compiles certain features of the invention. Other features will be revealed in subsequent paragraphs. Its purpose is to cover various modifications and similar arrangements within the spirit and scope of the attached patent application.

In order to solve the problems existing in the prior art, the present invention discloses an automated production line product appearance defect inspection method, comprising the steps of: in a data preparation stage, executing: a) photographing the appearance of a plurality of defective products of a product in a certain direction at a fixed distance to obtain a plurality of learning image files; b) generating a red primary color image file, a blue primary color image file, a green primary color image file, a Sobel image file and an edge effect (Edge) image file for each learning image file, wherein the total number of pixels and the pixel arrangement method of each image file are the same; and c) concatenating the image files in step b) into a training image file; in a model training stage, executing: d) using all the image files to obtain a training image file; The training image and the corresponding defect type are used to generate an estimation model through a convolution neural network to estimate the defect types of other image files; and in the first quality control stage, the following steps are performed: e) obtaining a confirmation image file of the appearance of a product under inspection of the product in the fixed direction photographed at a fixed distance in step a); f) generating the red primary color image file, the blue primary color image file, the green primary color image file, the Sobel image file and the edge effect image file of the confirmation image file, wherein the total number of pixels and the pixel arrangement method of each image file are the same; g) concatenating the image files in step f) into a confirmation image file; and h) inputting the confirmation image file into the estimation model to obtain the corresponding defect type.

In the automated production line product appearance defect inspection method, in step b), a binarized (Threshold) image file can be further produced for each learning image file, and in step f), the binarized image file related to the confirmation image file can be further produced.

In the automated production line product appearance defect inspection method, in step b), a high-pass filter (High Pass Filter) processed image file can be further produced for each learning image file, and in step f), the high-pass filter processed image file for the confirmation image file can be further produced.

Preferably, the differences in background brightness and color temperature of the learning image files should not exceed 1%.

Preferably, the difference in the number of pixels between any two learning image files in any direction on the plane is less than 0.5%.

The present invention also discloses an automated production line product appearance defect inspection system, comprising: an inspection jig, which systematically arranges a plurality of defective products or inspected products of a product; an image capture device, which maintains a substantially similar distance from each defective product or inspected product on the inspection jig, to sequentially photograph the appearance of the defective products in a certain direction to obtain a plurality of learning image files, and to sequentially photograph the appearance of the inspected products in the certain direction to obtain a plurality of confirmation image files; and an automatic learning module, which is installed in a working host and comprises the following submodules: a data pre-processing submodule, which receives the learning image files and the confirmation image files from the image capture device and removes background images irrelevant to the product; and an image generation submodule, which performs the following operations: generating a plurality of learning image files for the learning images; The file and each of the confirmation image files are used to generate a red primary color image file, a blue primary color image file, a green primary color image file, a Sobel image file and an edge effect image file, and the total number of pixels and pixel arrangement of each image file are the same; the image files generated from the same learning image file are concatenated into a training image file; and the image files generated from the same confirmation image file are concatenated into a confirmation image file; a model learning submodule, using all the training images and the corresponding defect types, generates an estimation model through a convolutional neural network to estimate the defect types of other image files; and a quality judgment submodule, inputs the confirmation images into the estimation model to obtain the corresponding defect types.

According to the present invention, the image generation submodule can further generate a binary image file associated with each of the learning image files and the confirmation image files.

According to the present invention, the image generation submodule can further generate a high-pass filter processed image file associated with each of the learning image files and the confirmation image files.

The automated production line product appearance defect inspection system may further include a light source that maintains the difference in background light brightness and color temperature in the closed space where the image capture device is located within 1% when emitting light.

Preferably, the difference in the number of pixels between any two learning image files in any direction on the plane is less than 0.5%.

By using the automatic learning module to learn the learning image files obtained by the image capture device, and these learning image files have been normalized, the system can more accurately predict product defects and then eliminate them. The model training of the present invention is accurate, and the error rate will also be reduced when selecting defective products online.

10: Inspection fixture

10’: Inspection fixture

20: Image capture device

20’: First image capture device

20”: Second image capture device

30: Automatic learning module

31: Data pre-processing submodule

32: Image generation submodule

33: Model learning submodule

34: Quality judgment submodule

40: Light source

A: Closed space

B: Blue primary color image file

E: Edge effect image file

G: Green primary color image file

H: High-pass filter processing image file

L: Learning video file

P:Products

R: Red primary color image file

S: Sauber image file

S01~S08: Steps

Figure 1 is a flow chart of an automated production line product appearance defect inspection method according to an embodiment of the present invention.

FIG2 shows a red primary color image file, a blue primary color image file, a green primary color image file, a Sauber image file, and an edge effect image file made from a learning image file.

FIG3 shows the red primary color image file, the blue primary color image file, the green primary color image file, the Sobel image file, the edge effect image file, a binarized image file, and a high-pass filter processed image file made from the learning image file.

Figure 4 is a schematic diagram of the components of an automated production line product appearance defect inspection system according to an embodiment of the present invention.

Figure 5 shows another view of the image capture device and inspection fixture of the automated production line product appearance defect inspection system.

Figure 6 shows the architecture of an automatic learning module.

The present invention will be described in more detail with reference to the following embodiments.

The present invention discloses an automated production line product appearance defect inspection method (hereinafter referred to as the present invention). The present invention is divided into three implementation stages: a data preparation stage, a model training stage, and a quality control stage. The following describes the specific steps of different stages in order.

In the data preparation stage, the first step of the method is to shoot the appearance of a plurality of defective products of a product in a certain direction at a fixed distance to obtain a plurality of learning image files (S01). Regardless of how the defective products are arranged, such as in a straight line or in a ring, the same side (in a fixed direction) of each defective product must face the image capture device for shooting images, and the distance between the side of each defective product and the image capture device must be as fixed as possible to ensure that the size of the defective product and the pixel range occupied in each learning image file captured by the image capture device have small differences. According to the requirements of the present invention, the difference ratio of the number of pixels in any direction (vertical and horizontal directions) of any two learning image files on the plane is less than 0.5%. For example, in the learning image file of defective product A, the longest horizontal part occupies 4,000 pixels, and in the learning image file of defective product B, the corresponding location can only be between 3,980 and 4,020 pixels. In practice, the number of pixels occupied by all defective products in that location can be averaged, and the defective learning image files that exceed the average value by more than 0.5% can be removed. The same process is done in the vertical direction. In theory, the products of the automated production line have a consistent appearance, and under the aforementioned image capture conditions, there will be consistent image positions and pixel numbers. However, this theoretical value will be different for each learning image file due to operator errors, differences in installation on the fixture, temperature and humidity affecting light, and even problems with the image capture device itself. In order to make the subsequent defect learning model training process more effective and reduce learning errors, the learning image files that do not meet the above restrictions are eliminated for model training. Since the number of defective products (here refers to the same defect, such as cracks) can be picked out from the bulk, the total number of learning image files will not be a problem. In addition, the background light source will also affect the quality of the learning image files. Therefore, according to the present invention, the difference in the brightness and color temperature of the background light obtained for the learning image files does not exceed 1%, and the learning image files that exceed the standard should also be eliminated. Image The capture device may be, but is not limited to, a charge-coupled device (CCD) image capture device or a complementary metal-oxide-semiconductor (CMOS) image capture device. The type may be a camera (not limited to a wired or wireless type). The file format of the learning image file is not limited by the present invention, and any commonly used image file format may be used, but the image file specified in the next step must be generated according to the learning image file.

The second step performed in the data preparation stage is to generate a red primary color image file, a blue primary color image file, a green primary color image file, a Sobel image file, and an edge effect (Edge) image file for each learning image file, wherein the total number of pixels and the pixel arrangement method of each image file are the same (S02). For the description of the related image files, please see FIG. 2, which shows a red primary color image file R, a blue primary color image file B, a green primary color image file G, a Sobel image file S, and an edge effect image file E generated from a learning image file L. The red primary color image file R, the blue primary color image file B, and the green primary color image file G are images based on red, blue, and green by using the three primary color light mode to separate the colors of the learning image file L. The red primary image file R, the blue primary image file B and the green primary image file G can be synthesized into the original learning image file L. In Figure 2, since the original appearance color of the defective product is mainly gray, the red primary image file R, the blue primary image file B and the green primary image file G are all biased towards a combination of gray colors. For the convenience of explanation, the defects in the learning image file L in Figure 2 are represented by the three colored lines in the upper right corner. The red primary image file R, the blue primary image file B and the green primary image file G also have three corresponding defect lines at the same position. The Sobel image file S uses the Sobel operator and applies the gradient of the image as the judgment basis. Separate masks are used to detect boundaries of different directions. When the gradient change exceeds a threshold value, it is judged as a boundary. The edge effect image file E uses the edge detection algorithm to find the characteristic edges from the learning image file L. It should be noted that the total number of pixels and the arrangement of pixels in each produced image file are the same. For example, if each image file is 4,000 x 3,000, it does not have to be the same as the total number of pixels and the arrangement of pixels in the learning image file L, but it is best to be consistent. This is for the next step.

The third step performed in the data preparation phase is to concatenate the image files in step S02 into a training image file (S03). As shown in Figure 2, "+" indicates the order of concatenation: red primary image file R, blue primary image file B, green primary image file G, Sobel image file S and edge effect image file E. Therefore, 5 different patterns can be seen in the training image file, regardless of whether they are arranged horizontally or vertically, but the arrangement order must be consistent. The number of training images will be generated for the number of learning image files L. The format of the training image file is not limited, and can be a raster image file such as a JPEG file, a vector image file such as a WMF file, or a lossless compressed bitmap image file such as a PNG file.

According to the present invention, the following steps are performed in the model training stage: all training images and corresponding defect types are used to generate an estimation model through a convolutional neural network to estimate the defect types of other image files (S04). All training images refer to all training images used for input learning regardless of the type of defect. Each defective product corresponding to each training image may have more than one defect, such as cracks, laser engraving position deviation, paint color deviation, etc., which are all the aforementioned defect types. The convolutional neural network consists of one or more convolutional layers and a fully connected layer at the top, as well as associated weights and pooling layers. The convolutional neural network can predict the characteristics of other homogeneous two-dimensional structures based on the two-dimensional structure and characteristics of the existing input data. The convolutional neural network has good results in image recognition. The present invention does not limit the design of the convolutional neural network, but only optimizes the input training image file with special technology, so that the same convolutional neural network structure can have more accurate prediction results for predicting specific defect types in the process of learning training images.

In the quality control stage, the present invention performs the following steps in sequence. First, obtain a confirmation image (S05) file of the fixed direction appearance of an inspected product of the product photographed at a fixed distance by the synchronization step S01. Here, the inspected product and the defective product are the same product, but the inspected product is the target for defect type estimation using the estimation model, and can be a series of products produced on the production line. The inspected product is photographed using the same shooting method as the defective product (fixed direction, fixed distance, same color temperature, same brightness) to obtain a confirmation image file. The content type and file format of the confirmation image file are the same as those of the learning image file. Second, prepare the red primary color image file, the blue primary color image file, the green primary color image file, the Sauber image file and the edge effect image file for the confirmation image file, wherein the total number of pixels and the pixel arrangement method of each image file are the same (S06). Similarly, both the confirmation image file and the learning image file must prepare the aforementioned five types of image files to have a consistent basis for comparison, and the difference ratio of the number of pixels occupied by the product is also within 0.5%. Third, concatenate the image files in step S06 into a confirmation image file (S07). The confirmation image file is the same as the training image file, which is concatenated by the red primary color image file, the blue primary color image file, the green primary color image file, the Sauber image file and the edge effect image file. The format and size of the confirmation image file and the training image file are the same, and the order of concatenating the image files is also the same. Fourth, the confirmed image file is input into the estimation model to obtain the corresponding defect type (S08), which is the result of deep learning.

In the aforementioned method, if you want to further increase the learning effect of the convolutional neural network and produce a more accurate estimation model, you can further generate a binarized (Threshold) image file for each learning image file in step S02. Simultaneously, step S06 also generates the binarized image file for the confirmation image file. The binarized image is to convert the grayscale image into a black and white image according to a specific clip-value, so as to find the image edge of the product and mark its defects. See Figure 3, the binarized image file T of the learning image file L is shown in the figure. The estimation model has an additional object of learning the binarized image file T of the learning image file L, and the estimation will also be based on the part of the binarized image file in the confirmation image file for judgment.

According to the present invention, the aforementioned method can further increase the learning effect of the convolutional neural network and produce the most accurate estimation model. In this case, in step S02, a high-pass filter (High Pass Filter) processing image file is further produced for each learning image file, and in step S06, the high-pass filter processing image file for the confirmation image file is further produced. "High pass" refers to the action of sharpening or edge enhancement graphics. The image file obtained by executing the filter (software) is the high-pass filter processing image file. Please refer to Figure 3 again. The high-pass filter processed image file H of the learning image file L is shown in the figure. The red primary color image file R, the blue primary color image file B, the green primary color image file G, the Sauber image file S, the edge effect image file E, the binarized image file T and the high-pass filter processed image file H are sequentially connected in series to form the most complete training image file or confirmation image file for learning or estimation.

Please see Figure 4, which is a schematic diagram of the components of an automated production line product appearance defect inspection system (hereinafter referred to as the system) of an embodiment of the present invention. The system is to be installed in a closed space A that is not disturbed by external light sources, so that images of defective products or inspected products in the closed space A can be accurately taken as a basis for intelligent learning and judgment. The system includes an inspection fixture 10, an image capture device 20, an automatic learning module 30 and a light source 40. The types, functions and overall operations of these technical components will be explained in detail with the relevant diagrams.

The inspection jig 10 systematically arranges several defective products or inspected products of a product P. "System" means arranging them in a regular manner to operate systematically. When the product P is a defective product, the inspection jig 10 is used to perform the steps of the data preparation stage of the above method; when the product P is an inspected product, the inspection jig 10 is used to perform the steps of the quality control stage of the above method.

The image capture device 20 maintains a substantially similar distance from each defective product or inspected product on the inspection jig 10, so as to sequentially photograph the appearance of the defective products in a certain direction to obtain a plurality of learning image files (used in each step of the data preparation stage), and sequentially photograph the appearance of the inspected products in the certain direction to obtain a plurality of confirmation image files (used in each step of the quality control stage). The shape of the image capture device 20 has been disclosed in the previous text and will not be repeated here. The image capture device 20 interacts with the inspection jig 10 to complete the shooting operation. In this embodiment, the inspection jig 10 is a fixed device, on which products P (all defective products or all inspected products) are fixedly installed in sequence and at equal intervals. The image capture device 20 sequentially photographs the product P from the left (indicated by the solid line image) to the right (indicated by the dotted line image). This is a combination of the image capture device 20 and the inspection jig 10. In other embodiments, as shown in FIG5, the combination of the image capture device and the inspection jig of the present system has another form. In this embodiment, the inspection jig 10' is a ring-shaped device, on which the products P are placed sequentially and at equal intervals, and the inspection jig 10' can be rotated (clockwise) to drive the product P. Product P has two surfaces that need to be inspected for appearance (product P is drawn as a trapezoid, and the two surfaces are the surfaces corresponding to the upper and lower bases respectively). The first image capture device 20' is used to photograph the front surface of product P, and the first image capture device 20" is used to photograph the back surface of product P. This combination is that the inspection fixture 10' moves but the image capture device does not move, which facilitates the operation of multiple image capture devices.

The automatic learning module 30 is installed in a working host 1, and a specific operation is completed by operating the working host 1. The hardware architecture of the working host 1 of this system is not much different from the general server architecture, and may include a central processing unit, memory, storage device (such as a hard disk), input and output units, etc. Based on different needs, the characteristics and functions of various components of the hardware will be different. These hardware are architectures that technical personnel in the server field should understand. In addition, the various modules and sub-modules of the automatic learning module 30 of this system introduced below are technical requirements for operating by utilizing or cooperating with the above-mentioned existing hardware equipment. Therefore, they can be software, including specific program codes and data, and run in at least part of the hardware architecture under the operating system (for example, the program codes and related data files are stored in the storage device, temporarily stored in the memory under the operation of the operating system, and dynamically called and executed by the central processor). On the other hand, these modules can also be special hardware, such as application-specific integrated circuits (ASIC) or external cards, to perform the functions assigned by these modules or sub-modules. What's more, these technical requirements can be partly software and partly hardware, and can be effectively integrated according to the needs of product designers, all within the technical scope advocated by this patent.

Please see Figure 6, which shows the architecture of the automatic learning module 30. The automatic learning module 30 includes the following sub-modules: a data pre-processing sub-module 31, an image generation sub-module 32, a model learning sub-module 33 and a quality judgment sub-module 34.

The data pre-processing submodule 31 receives the learning image files and the confirmation image files from the image capture device 10, and removes the background image that is not related to the product. In this way, the learning image files and the confirmation image files will be "cleaner", but the file size (number of pixels) will differ from each other within 0.5%. The image generation submodule 32 can perform the following operations: First, for each of the learning image files and the confirmation image files, a red primary color image file, a blue primary color image file, a green primary color image file, a Sobel image file and an edge effect image file are produced. The total number of pixels and the pixel arrangement of each image file are the same. If the difference ratio of the number of pixels in any direction of the plane between any two learning image files is less than 0.5%, the image generation submodule 32 can exclude the learning image files that do not meet the requirements. Second, the image files produced from the same learning image file are concatenated into a training image file for the data preparation stage. Third, the image files produced from the same confirmation image file are concatenated into a confirmation image file for the quality control stage. The model learning submodule 33 uses all the training images and the corresponding defect types to generate an estimation model through a convolutional neural network to estimate the defect types of other image files. The operation method is the same as that described in the aforementioned method step S04, and will not be repeated here. The quality judgment submodule 34 inputs the confirmation images into the estimation model to obtain the corresponding defect types.

The image generation submodule 32 further generates a binary image file related to each of the learning image files and the confirmation image files, and adds the binary image file to the red primary color image file, the blue primary color image file, the green primary color image file, the Sobel image file, and the edge effect image file for learning or for estimating the defect type. Adding the binary image file content to the training image file can increase the learning effect of the convolutional neural network and produce a more accurate estimation model. Similarly, the image generation submodule 32 can also further generate high-pass filter processed image files related to each of the learning image files and the confirmation image files, and sequentially connect the red primary color image file, the blue primary color image file, the green primary color image file, the Sauber image file, the edge effect image file, the binarization image file and the high-pass filter processed image file to form a training image file or a confirmation image file for learning or estimating the defect type.

The light source 40 has the function of controlling the image quality of the image capture device 20. When emitting light, the light source 40 maintains the difference in the background light brightness and color temperature in the closed space A where the image capture device 40 is located within 1%, so as to assist in learning the image file and confirm the normalization of the image file.

Although the present invention has been disclosed in the form of implementation as above, it is not intended to limit the present invention. Anyone with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of the patent application attached hereto.

S01~S08: Steps

Claims (10)

An automated production line product appearance defect inspection method comprises the steps of: in a data preparation phase, executing: a) photographing the appearance of a plurality of defective products of a product in a certain direction at a fixed distance to obtain a plurality of learning image files; b) generating a red primary color image file, a blue primary color image file, a green primary color image file, a Sobel image file and an edge effect (Edge) image file for each learning image file, wherein the total number of pixels and the pixel arrangement method of each image file are the same; and c) concatenating the image files in step b) into a training image file; in a model training phase, executing: d) using all training images and corresponding defect images to obtain a training image file; Type, through a convolution neural network to generate an estimation model for estimating the defect type of other image files; and in the first quality control stage, perform: e) obtain a confirmation image file of the fixed direction appearance of a product under inspection of the product photographed at a fixed distance in synchronization with step a); f) generate the red primary color image file, the blue primary color image file, the green primary color image file, the Sobel image file and the edge effect image file related to the confirmation image file, wherein the total number of pixels and the pixel arrangement method of each image file are the same; g) concatenate the image files in step f) into a confirmation image file; and h) input the confirmation image file into the estimation model to obtain the corresponding defect type. As described in the first item of the patent application scope, the automated production line product appearance defect inspection method, wherein in step b), a binarization (Threshold) image file is further produced for each learning image file, and in step f), the binarization image file related to the confirmation image file is further produced. As described in the second item of the patent application scope, the automated production line product appearance defect inspection method, wherein in step b), a high pass filter (High Pass Filter) processing image file is further produced for each learning image file, and in step f), the high pass filter processing image file of the confirmation image file is further produced. As described in Item 1 of the patent application scope, the method for inspecting appearance defects of products on an automated production line, wherein the difference in the background light brightness and color temperature of the learning image files does not exceed 1%. As described in Item 1 of the patent application scope, the method for inspecting appearance defects of products on an automated production line, wherein the difference in the number of pixels of defective products in any two learning image files in any direction on the plane is less than 0.5%. An automated production line product appearance defect inspection system comprises: an inspection jig, which systematically arranges a plurality of defective products or inspected products of a product; an image capture device, which maintains a substantially similar distance from each defective product or inspected product on the inspection jig, to sequentially photograph the appearance of the defective products in a certain direction to obtain a plurality of learning image files, and to sequentially photograph the appearance of the inspected products in the certain direction to obtain a plurality of confirmation image files; and an automatic learning module, which is installed in a working host and comprises the following submodules: a data pre-processing submodule, which receives the learning image files and the confirmation image files from the image capture device and removes background images irrelevant to the product; an image generation submodule, which performs the following operations: generating a plurality of learning image files and the confirmation image files; Each of the confirmed image files generates a red primary color image file, a blue primary color image file, a green primary color image file, a Sobel image file and an edge effect image file, and the total number of pixels and pixel arrangement of each image file are the same; the image files generated from the same learning image file are concatenated into a training image file; and the image files generated from the same confirmed image file are concatenated into a confirmed image file; a model learning submodule, using all the training images and the corresponding defect types, generates an estimation model through a convolutional neural network to estimate the defect types of other image files; and a quality judgment submodule, inputs the confirmed images into the estimation model to obtain the corresponding defect types. As described in Item 6 of the patent application scope, the automated production line product appearance defect inspection system, wherein the image generation submodule further produces a binary image file associated with each of the learning image files and the confirmation image files. As described in Item 7 of the patent application scope, the automated production line product appearance defect inspection system, wherein the image generation submodule further produces a high-pass filter processed image file associated with each of the learning image files and the confirmation image files. The automated production line product appearance defect inspection system as described in Item 6 of the patent application further includes a light source that maintains the difference in background light brightness and color temperature in the closed space where the image capture device is located within 1% when emitting light. As described in Item 6 of the patent application scope, the automated production line product appearance defect inspection system, wherein the difference ratio of the number of pixels of defective products in any two learning image files in any direction of the plane is less than 0.5%.