Substrate defect detection system and method based on deep learning

Abstract

The invention discloses a substrate defect detection system and method based on deep learning, and belongs to the technical field of machine vision, and the system comprises a camera detection control unit which is used for carrying out the position compensation motion of a robot or feeding back a detection result to an upper computer for presentation; the camera visual guidance unit is used for receiving the image data and the sensor data in real time, processing the image data and the sensor data and sending a processing result instruction to the collaborative robot for motion compensation; and the detection algorithm software unit is used for performing result comparison and presentation with a detection picture shot by the camera and obtained by the upper computer software platform. According to the invention, high-precision and high-efficiency automatic defect detection can be realized.

Description

Technical Field

[0001] The present invention relates to the field of machine vision technology, and more specifically, to a substrate defect detection system and method based on deep learning. Background Art

[0002] The rapid development of substrate process technology, with finer pitch and more complex coating processes, has made it difficult for traditional substrate defect detection methods (such as manual visual inspection, optical inspection and electronic testing) to meet quality requirements.

[0003] Chinese patent application number CN116993718A discloses a machine vision-based method for defect detection of TFT array substrates. This method processes the basic grayscale image of the TFT array by analyzing grayscale values, constructs a target grayscale run-length matrix, compares the relative position distribution of the maximum data value and the data value to obtain the comprehensive noise intensity, and performs filtering to improve the accuracy of the detection results. Chinese patent application number CN116894991A discloses a deep learning model optimization method for weld target detection, a weld target detection method and system based on the optimized deep learning model, and uses the ADMM algorithm to perform joint weight pruning, quantization, and dequantization on an SSD deep learning model initialized with pre-trained FP32 weights. The optimized deep learning model is then quantized again to obtain a further optimized deep learning model. A weld target detection method based on the optimized deep learning model is also disclosed. Chinese patent publication number CN217748156U discloses a package substrate defect detection device, which is equipped with a loading device, a rear-side suction platform module, a front-side suction platform module, a detection device, a flipping device, and an unloading device on a frame. Visual cameras are installed on the front and rear suction platform modules for identification and detection. Chinese patent application publication number CN116087231A discloses a Faster-RCNN-based LED substrate defect detection system. This system utilizes image acquisition, processing, model building, training, and defect detection modules. Each module is responsible for acquiring images of the LED substrate from different angles, processing the images, building a detection model, training the model, and detecting defects.

[0004] In his paper "Research on Defect Detection Technology for Flexible Packaging Substrates Based on Machine Learning," Zheng Shuchao of Guangzhou University proposed an abnormal region detection algorithm based on a combination of accelerated robust features and a random forest algorithm, as well as a surface defect detection algorithm based on transfer learning. These algorithms effectively reduced the false positive rate, built a motion detection platform, and completed defect detection tests on high-density flexible packaging substrates. In his paper "Research on Defect Detection of Copper-clad Surfaces of High-Density Ultra-Thin Flexible Integrated Circuit Packaging Substrates," Li Lu of South China University of Technology proposed a copper foil surface segmentation method based on local histogram constraints, built a DAG-SVMS classifier based on color features to classify the degree of oxidation defects, and built a defect detection system based on an ultra-precision microscope. In his paper "Research on Surface Defect Detection for High-Density Flexible Packaging Substrates," Jiang Baochao of South China University of Technology used the residual network RES NET-101 as a feature extraction network and introduced a feature pyramid to improve the loss function, which to a certain extent solved the problem of difficult defect feature extraction. In his paper "High-Density Flexible Packaging Substrate Defect Detection Method Based on CRS-YOLO Algorithm", Wang Kaixin from South China University of Technology replaced the original SPP red pyramid with the more performant Basic RFB pooling pyramid, reducing the missed detection rate, and replaced the original CloU loss function with SloU to accelerate model convergence. In his paper "Research on Surface Defect Detection Algorithms and Systems for High-Density Flexible IC Substrates", Wang Siyuan from South China University of Technology proposed an improved wavelet transform image algorithm based on information to effectively reduce image noise, and proposed an image segmentation algorithm based on improved K-means clustering, which improved both segmentation accuracy and convergence speed. In his paper "Trace Defect Detection for Flexible Integrated Circuit Packaging Substrates", Huang Jiayang from South China University of Technology introduced a gradient matrix to optimize the existing relative total variation filtering algorithms, and proposed a random sampling consensus algorithm for trace defect detection, which is feasible in trace defect detection.

[0005] The above patents and papers show that machine learning is feasible in substrate defect identification, but the above solutions still have certain limitations due to the small number of high-precision substrate defect samples and unclear defect characteristics, and there are still deficiencies in automated identification and high-efficiency detection. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a substrate defect detection system and method based on deep learning to achieve high-precision and high-efficiency automated defect detection.

[0007] The object of the present invention is achieved through the following solutions: A substrate defect detection system based on deep learning, comprising: The camera detection control unit is connected to the industrial computer through the control module, so that after the camera completes shooting, the processed image is transmitted to the visual guidance unit and the detection algorithm software unit. After the image is recognized and detected, it is fed back to the control module to transmit the corresponding signal, perform position compensation movement of the robot, or feed back the detection result to the host computer for presentation; The camera vision guidance unit is used to provide images for the computing platform. The AI ​​computing platform is integrated into the industrial computer to receive and process image data and sensor data in real time, and send the processing result instructions to the collaborative robot for motion compensation. The detection algorithm software unit is used to build a first-order target detection algorithm based on a deep learning model, add an attention mechanism to the baseline model, design a loss function algorithm to detect substrate defects, train a deep learning-based substrate defect detection algorithm model, store and convert the generated model in a unified format, and finally compare the results with the detection images taken by the camera obtained by the host computer software platform.

[0008] Furthermore, the camera detection control unit uses the camera to transmit the image parsed by the preprocessing module with image enhancement function to the visual guidance unit and the detection algorithm software unit, and uses Gaussian blur to remove noise in the image acquired by the camera in real time during the image preprocessing stage. The calculation formula is as follows:

[0009] Where G is the filter function, I is the input image, and σ is the blur level.

[0010] Furthermore, the camera vision guidance unit includes a PID control module for performing precise motion control using a PID control algorithm. The control formula is as follows:

[0011] Where u(t) is the control output, e(t) is the difference between the target position and the current position, Kp, Ki, and Kd are the proportional, integral, and differential coefficients respectively, and w p 、w i 、w d are the weight parameters of proportional, differential and integral terms respectively.

[0012] Furthermore, in the detection algorithm software unit, a first-order target detection algorithm is built with a deep learning model, and an attention mechanism is added to the baseline model. A loss function algorithm is designed to detect substrate defects, and a substrate defect detection algorithm model based on deep learning is trained. The generated model is stored and converted in a unified format by the ONNX model framework, and finally the results are compared with the detection pictures taken by the camera obtained by the host computer software platform. Specifically, the following are presented: First, the feature map (C×H×W) is globally averaged pooled in width and height using pooling kernels of size (H, 1) and (1, W) to obtain a wide-dimensional feature map (C×1×W) and a high-dimensional feature map (C×H×1). Secondly, the wide-dimensional feature map and the high-dimensional feature map are merged and then 1*1 convolution is performed; Then, a nonlinear activation function is used to obtain a feature map of size C / r×1×(H+W); Finally, the generated feature map is encoded into features in the width (C×1×W) and height (C×H×1) directions, and combined with the original feature map to obtain the final attention feature map.

[0013] Furthermore, in the detection algorithm software unit, the following loss function is used to solve the problem that the direction mismatch between the predicted box and the true box causes the predicted box to oscillate around the true box during training, resulting in poor model convergence and slow convergence speed. The loss function expression is:

[0014] in:

[0015]

[0016] ; in, represents the intersection-and-union ratio, represents the shape cost, Represents the distance cost; B is the area value of the identified target box; is the area value of the identified object; w x , w y are weight parameters in the x and y directions; α and β are used to represent the nonlinearity of the exponential function; c x , c y Used to set the initial position of the distance cost; 、 are the union of the predicted and actual values ​​of the image height / width in the corresponding directions; 、 express; is the attention degree coefficient of shape cost; 、 The width and height of the recognition box.

[0017] Furthermore, the camera detection control unit includes an industrial CCD camera.

[0018] Furthermore, the camera vision guidance unit includes a laser sensor.

[0019] Furthermore, the control module includes a PLC module.

[0020] Furthermore, the detection algorithm software unit includes an ONNX model framework, which stores and converts the generated model in a unified format.

[0021] A substrate defect detection method based on deep learning, based on any of the substrate defect detection systems based on deep learning as described above, performs the following steps: S1-1, the workpiece arrives at the monitoring point; S1-2, the robot moves to the default position; S2-1, camera takes pictures and processes them, and the robot performs compensation movements; S2-2, recognition using camera photo algorithm; S3, upload the test results to the host computer.

[0022] The beneficial effects of the present invention include: (1) The present invention uses deep learning technology to detect substrate defects and uses an industrial CCD camera for image acquisition, which has the characteristics of higher recognition rate; the algorithm part adds an attention mechanism and adopts a new loss function to replace the original CIoU_Loss loss function, which can significantly improve the convergence speed of the model during the training stage and has better reasoning ability.

[0023] (2) The present invention uses machine vision to process substrate inspection, is compatible with the image characteristics brought by the substrate coating, can effectively identify the defect types unique to the substrate, can expose low-visibility processing defects through attention mechanisms, etc., and has objectivity and standardization of inspection.

[0024] (3) The present invention has built a fully automatic detection system, which eliminates visual fatigue compared to traditional visual inspection, increases detection accuracy, effectively reduces inspection costs, reduces the introduction of human subjective factors, improves detection accuracy, and increases substrate detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 Flowchart of a substrate defect detection method based on deep learning according to an embodiment of the present invention; Figure 2This is a workflow diagram of a visual guidance system according to an embodiment of the present invention; Figure 3 This is a network model diagram of a camera detection algorithm system according to an embodiment of the present invention; Figure 4 This is a module diagram of the attention mechanism of an embodiment of the present invention. DETAILED DESCRIPTION

[0027] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0028] In a preferred embodiment, a deep learning-based substrate defect detection system and method are proposed, aiming to achieve high-precision, high-efficiency automated defect detection. Specifically addressing the imbalance of substrate defect samples, an attention mechanism is added to the baseline model to effectively address the problem of substrate defects occupying a small number of pixels in the image to be detected, and the lack of obvious features of some defects (such as scratches and indentations). To address the problem of feature information loss during the convolution process, a new loss function is proposed to address the directional mismatch between the predicted and true frames. This significantly improves the model's convergence speed during training and provides better inference performance.

[0029] More specifically, a substrate defect detection system based on deep learning includes: The camera detection control unit can be specifically composed of a hardware combination of an industrial CCD camera, a camera fixture, a collaborative robot, and a camera light source. The camera detection control unit is connected to an industrial computer via a PLC control module. After the camera completes shooting, it transmits the processed image to the vision guidance unit and detection algorithm software. After recognition and detection, the image is fed back to the PLC control module, transmitting the corresponding signal, performing position compensation movement for the robot, or feeding the detection results back to the host computer for presentation.

[0030] The camera vision guidance unit is implemented using a combination of industrial CCD cameras, a computing platform, and laser sensors. The industrial CCD camera provides high-quality, real-time images to the computing platform. The high-performance AI computing platform, integrated into the industrial computer, receives and processes image and sensor data in real time, then transmits the resulting instructions to the collaborative robot for motion compensation.

[0031] The detection algorithm software unit is specifically composed of a deep learning model algorithm, an ONNX model framework, and a host computer software platform. A first-order object detection algorithm is built using a deep learning model, and an attention mechanism is added to the baseline model. A new loss function algorithm is proposed to replace the original CIoU_Loss loss function for substrate defect detection. A deep learning-based substrate defect detection algorithm model is trained. The ONNX model framework stores and converts the generated model in a unified format, and the results are ultimately compared with inspection images captured by an industrial CCD camera and obtained by the host computer software platform.

[0032] In a further invention concept, an innovative method is adopted to use an industrial CCD camera to transmit the image analyzed by an intelligent pre-processing module with image enhancement function to a visual guidance unit and detection algorithm software, so that the image clarity and detail capture ability can be significantly improved in complex lighting environments, thereby improving training efficiency and reducing training costs.

[0033] The images acquired by the industrial CCD camera in real time are de-noised using Gaussian blurring in the image preprocessing stage. The calculation formula is as follows:

[0034] Where G is the filter function, I is the input image, and σ is the blur level.

[0035] In a further inventive concept, in step S2, a high-performance AI computing platform processes image and sensor data in real time, improving system response speed. The laser sensor enhances the system's intelligence, the image processing module ensures recognition accuracy, and the PID control algorithm enhances the robot's motion precision. The combination of hardware and software creates a synergistic effect across the entire vision guidance unit, enabling efficient data collection and processing, ensuring the robot's intelligence, and ultimately improving work efficiency.

[0036] The PID control algorithm is used for precise motion control, and the control formula is as follows:

[0037] Where u(t) is the control output, e(t) is the difference between the target position and the current position, Kp, Ki, and Kd are the proportional, integral, and differential coefficients respectively, and w p 、w i 、w d are the weight parameters of proportional, differential and integral terms respectively.

[0038] In a further inventive concept, in step S3, a new attention mechanism is added to address the problem of feature information loss during the convolution process caused by the fact that substrate defects occupy fewer pixels in the image to be detected and the characteristics of some defects (such as scratches and indentations) are not obvious. First, the feature map (C×H×W) is globally averaged pooled in the width and height directions using pooling kernels of size (H, 1) and (1, W), respectively, to obtain a wide-dimensional feature map (C×1×W) and a high-dimensional feature map (C×H×1). Secondly, the wide-dimensional feature map and the high-dimensional feature map are combined and 1*1 convolution is performed. Then, a nonlinear activation function is used to obtain a feature map of size C / r×1×(H+W). Finally, the generated feature map is encoded into features in the width (C×1×W) and height (C×H×1) directions, and combined with the original feature map to obtain the final attention feature map.

[0039] In a further inventive concept, the detection algorithm in step S3 uses a new loss function to solve the problem that the direction mismatch between the predicted box and the true box causes the predicted box to oscillate around the true box during training, resulting in poor model convergence and slow convergence speed. The loss function is expressed as follows (see the implementation example for details of the public derivation):

[0040] in:

[0041]

[0042] .

[0043] In summary, the present invention applies deep learning technology to the substrate defect detection method, performs automatic feature extraction and analysis through substrate images, quickly and accurately identifies various types of defects, and greatly improves the accuracy and efficiency of detection; through the training and optimization of deep learning models, adaptive detection of different types of substrates and various complex defects is achieved, enhancing the versatility and flexibility of the detection system; in addition, the substrate defect detection system based on deep learning can also provide real-time feedback on detection results, reduce production costs, and enhance the overall competitiveness of the electronics manufacturing industry.

[0044] In another embodiment, a method for detecting substrate defects based on deep learning is specifically provided. Based on the above system, it includes an industrial CCD camera module, a collaborative robot, an industrial computer, a PLC control module, and related tooling brackets, etc., for performing defect detection on substrates. The system uses deep learning algorithms, a high-performance AI computing platform (such as NVIDIA Jetson), and an ONNX model framework to build a detection algorithm software platform, compare and identify the scanned images of the substrate, determine the defect location, identify and mark it, and feed the results back to the industrial computer. The process is as follows: Figure 1 As shown, the specific steps include: 1) Install the industrial CCD camera and collaborative robot. The collaborative robot base is mounted on the boom beam, allowing the entire robot to droop. The industrial CCD camera, illuminated by a ring light source, is mounted at the end of the robot. This facilitates subsequent motion compensation and establishes a TCP coordinate system for the robot, with the camera's main viewpoint as the positive direction.

[0045] 2) Install the PLC control module. The PLC control module and the high-performance AI computing platform control the collaborative robot through the PID algorithm to perform target position compensation movement. It has the function of receiving the workpiece request detection signal and receiving the detection results.

[0046] 3) The camera vision guidance system of the present invention is completed by combining hardware such as industrial CCD cameras, computing platforms, and laser sensors. The industrial CCD camera is mainly used to provide high-quality real-time images to the computing platform. The computing platform is deployed on an industrial computer and is used to receive image data and sensor data in real time. The processing result instructions are sent to the collaborative robot through the PLC, allowing the robot to perform position compensation movements. The workflow is as follows: Figure 2 shown.

[0047] Among them, the real-time acquired image uses Gaussian blur to remove noise in the image preprocessing stage, and the calculation formula is as follows.

[0048]

[0049] Where G is the filter function, I is the input image, and σ is the blur level.

[0050] Use PID control algorithm for precise motion control. The control formula is as follows:

[0051] Where u(t) is the control output, e(t) is the difference between the target position and the current position, K p , K i , K d are proportional, integral, and differential coefficients respectively, w p 、w i 、wd They are the weight parameters of the proportional, differential and integral terms respectively. The specific examples are as follows: An industrial machine needs to move from its current position A (2, 3) to its target position B (5, 5). The current position error e is calculated as follows.

[0052]

[0053] is the horizontal coordinate of point b, is the horizontal coordinate of point a, is the ordinate of point b, is the vertical coordinate of point a. The values ​​of each parameter in the PID control algorithm are K p =1,K i =0, K d =0.05, w p =1,w i =0,w d =1, =0.5, u(t)=5.135, ∫e(t)dt=15, which means the robot needs to exert a control force of 5.135 to move to the target position B.

[0054] 4) Build a detection algorithm software system. Figure 3 As shown in the figure, this system consists of a deep learning model algorithm, an ONNX model framework, and a host computer software platform. It uses a deep learning model to build a first-order target detection algorithm, adds an attention mechanism to the baseline model, and proposes a new loss function algorithm to replace the original CIoU_Loss loss function to detect substrate defects.

[0055] After model training is completed, the ONNX model framework will store and convert the generated model in a unified format, and finally compare the results with the detection images taken by the industrial CCD camera obtained by the host computer software platform, and send the corresponding results to the PLC module.

[0056] Among them, the attention mechanism can perform global average pooling on the input feature map (C×H×W) in both width and height directions using two pooling kernels of size (H, 1) and (1, W) to extract a wide-dimensional feature map (C×1×W) and a high-dimensional feature map (C×H×1). Each attention feature map can obtain long-range dependencies along the spatial direction. The two attention maps are merged together for 1*1 convolution, and then a nonlinear activation function is used to obtain a feature map of size C / r×1×(H+W). The feature map is encoded into features in both width (C×1×W) and height (C×H×1) directions, and combined with the original feature map to obtain the final attention feature map. The specific mechanism is as follows: Figure 4 shown.

[0057] In the detection algorithm of the present invention, a new loss function is used to replace the original CIoU_Loss loss function. The loss function includes four penalty terms: angle cost, distance cost, shape cost, and IoU cost.

[0058] (1) Angle cost: By introducing the angle-aware LF component, we make predictions on the X or Y axis and then continue to move closer along the closest axis to reduce the number of distance-related variables, improve the speed of model training and the accuracy of inference. The calculation formula is:

[0059] in:

[0060] 、 are the true width and predicted width of the recognition box respectively; 、 are the true height and predicted height of the recognition box respectively; For control The actual parameter of the flexible amplitude.

[0061] (2) The definition of distance cost takes into account the above angle cost, and its calculation formula is:

[0062] in:

[0063] w x , w y is the weight parameter in the x and y directions; α and β are used to indicate the degree of nonlinearity of the exponential function; c x , c y Used to set the initial position of the distance cost; 、 are the union of the predicted and actual values ​​of the image height / width in the corresponding directions.

[0064] (3) The calculation formula of shape cost is:

[0065] in,

[0066] 、 is the actual width and height of the image; 、 The width and height of the recognition box; is the attention degree coefficient of shape cost.

[0067] (4) The calculation formula of IoU cost is:

[0068] Among them, B is the area value of the recognition target box, is the area value of the identified object.

[0069] The camera detection motion system, visual guidance system, and detection algorithm software mentioned above are systematically combined to complete the construction of the entire defect detection system. The final regression loss function is calculated as follows: .

[0070] The units involved in the embodiments of the present invention may be implemented in software or hardware, and the units described may also be provided in a processor. In some cases, the names of these units do not limit the units themselves.

[0071] According to one aspect of an embodiment of the present invention, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0072] As another aspect, embodiments of the present invention further provide a computer-readable medium, which may be included in the electronic device described in the above embodiments, or may exist independently and not incorporated into the electronic device. The computer-readable medium carries one or more programs, and when executed by the electronic device, the electronic device implements the methods described in the above embodiments.

Claims

1. A substrate defect detection system based on deep learning, characterized in that: include: The camera detection control unit is connected to the industrial computer through the control module, so that after the camera completes shooting, the processed image is transmitted to the visual guidance unit and the detection algorithm software unit. After the image is recognized and detected, it is fed back to the control module to transmit the corresponding signal, perform position compensation movement of the robot, or feed back the detection result to the host computer for presentation; The camera vision guidance unit is used to provide images for the computing platform. The AI ​​computing platform is integrated into the industrial computer to receive and process image data and sensor data in real time, and send the processing result instructions to the collaborative robot for motion compensation. The detection algorithm software unit is used to build a first-order target detection algorithm based on a deep learning model, add an attention mechanism to the baseline model, design a loss function algorithm to detect substrate defects, train a deep learning-based substrate defect detection algorithm model, store and convert the generated model in a unified format, and finally compare the results with the detection images taken by the camera obtained by the host computer software platform.

2. The substrate defect detection system based on deep learning according to claim 1, characterized in that: The camera detection control unit uses the camera to transmit the image analyzed by the preprocessing module with image enhancement function to the visual guidance unit and the detection algorithm software unit, and uses Gaussian blur to remove noise in the image acquired by the camera in real time during the image preprocessing stage. The calculation formula is as follows: Where G is the filter function, I is the input image, and σ is the blur level.

3. The substrate defect detection system based on deep learning according to claim 1, characterized in that: The camera vision guidance unit includes a PID control module for precise motion control using a PID control algorithm. The control formula is as follows: Where u(t) is the control output, e(t) is the difference between the target position and the current position, Kp, Ki, and Kd are the proportional, integral, and differential coefficients respectively, and w p 、w i 、w d are the weight parameters of proportional, differential and integral terms respectively.

4. The substrate defect detection system based on deep learning according to claim 1, characterized in that: In the detection algorithm software unit, a first-order target detection algorithm is built using a deep learning model, and an attention mechanism is added to the baseline model. A loss function algorithm is designed to detect substrate defects, and a substrate defect detection algorithm model based on deep learning is trained. The generated model is stored and converted in a unified format by the ONNX model framework, and finally compared with the detection images taken by the camera obtained by the host computer software platform. Specifically, the following steps are performed: First, the feature map (C×H×W) is globally averaged pooled in width and height using pooling kernels of size (H, 1) and (1, W) to obtain a wide-dimensional feature map (C×1×W) and a high-dimensional feature map (C×H×1). Secondly, the wide-dimensional feature map and the high-dimensional feature map are merged and then 1*1 convolution is performed; Then, a nonlinear activation function is used to obtain a feature map of size C / r×1×(H+W); Finally, the generated feature map is encoded into features in the width (C×1×W) and height (C×H×1) directions, and combined with the original feature map to obtain the final attention feature map.

5. The substrate defect detection system based on deep learning according to claim 4, characterized in that: In the detection algorithm software unit, the following loss function is used to solve the problem that the direction mismatch between the predicted box and the true box causes the predicted box to oscillate around the true box during training, resulting in poor model convergence and slow convergence speed. The loss function expression is: in: ； in, represents the intersection-and-union ratio, represents the shape cost, Represents the distance cost; B is the area value of the identified target box; is the area value of the identified object; w x , w y are weight parameters in the x and y directions; α and β are used to represent the nonlinearity of the exponential function; c x , c y Used to set the initial position of the distance cost; 、 are the union of the predicted and actual values ​​of the image height / width in the corresponding directions; 、 express; is the attention degree coefficient of shape cost; 、 The width and height of the recognition box.

6. The substrate defect detection system based on deep learning according to claim 1, characterized in that: The camera detection control unit includes an industrial CCD camera.

7. The substrate defect detection system based on deep learning according to claim 1, characterized in that: The camera vision guide unit includes a laser sensor.

8. The substrate defect detection system based on deep learning according to claim 1, characterized in that: The control module includes a PLC module.

9. The substrate defect detection system based on deep learning according to claim 1, characterized in that: The detection algorithm software unit includes an ONNX model framework, which stores and converts the generated model in a unified format.

10. A substrate defect detection method based on deep learning, characterized in that: Based on the substrate defect detection system based on any one of claims 1 to 9, perform the following steps: S1-1, the workpiece arrives at the monitoring point; S1-2, the robot moves to the default position; S2-1, camera takes pictures and processes them, and the robot performs compensation movements; S2-2, recognition using camera photo algorithm; S3, upload the test results to the host computer.

Images