Defect Detection Of Ceramic Substrate Based On Deep Learning

As the packaging substrate of electronic devices,the quality of ceramic substrate has a great impact on the long-term stable operation of electronic products.However,due to the complex production environment and special production technology,various defects are easy to appear in the production process of ceramic substrate,which damages the structure of electronic devices and makes the whole electronic devices unable to work normally.Therefore,the flaw detection of ceramic substrate is an important step to ensure the production quality of electronic devices.The traditional manual microscope detection method has low efficiency and inconsistent detection standards,so it is difficult to meet the high efficiency production requirements of automated factories.The generalization ability and flexibility of detection algorithms based on traditional image algorithms and machine vision are poor,which is difficult to meet the requirements of accuracy of ceramic substrate defect detection.Stain,foreign matter,gold edge bulge,ceramic gap and damage are five typical defects commonly found on ceramic substrate.This topic in-depth analyzes the shortcomings of detection algorithms based on traditional manual detection and machine learning in these five typical defects detection,combines the target detection method of deep learning and makes optimization to improve the defect detection ability,and develops the corresponding defect detection system.The practicability of the proposed method is verified.The main research contents include:1.A defect detection method for ceramic substrate based on improved YOLOv4 is studied.In view of the difficulty of defect detection caused by small size,changeable color and shape of defects on ceramic substrate and large size variation among defects of different types,The improved YOLOv4 network introduced the confidence loss function based on the Gradient Harmonizing Mechanism(GHM)and the Criss-Cross Attention Net(CCNet)module.And add small scale target detection branch to improve defect detection capability.The experimental results show that for the 608?608 resolution ceramic substrate image,the average accuracy of the defect detection method based on the improved YOLOv4 for the ceramic substrate stain,foreign mater,gold edge bulge,ceramic gap and damage is 97.70%.The average recall rate reached 97.79%,which can meet the requirements of the precision of the detection of the defects of ceramic substrate in the industrial field.2.Aiming at the problem of decreasing detection accuracy of five types of defects on ceramic substrate under low-resolution(224 224?)images,a knowledge distillation based ceramic substrate defect detection algorithm,YOLOv4-CSKD,was proposed.Based on YOLOv4 algorithm framework,knowledge distillation algorithm was introduced to construct teacher network and student network.More accurate characteristic information learned by teacher network was used to guide the training of student network,so as to improve the defect detection ability of student network for low-resolution ceramic substrate image.At the same time,the feature fusion module based on Coordinate Attention(CA)is introduced in the teacher network,which makes the features learned by the teacher network adapt to both high-resolution(448 448?)image information and low-resolution image information,so as to better guide the training of students network.The experimental results show that YOLOV4-CSKD achieves an average accuracy of 96.51% and an average recall rate of 90.30% for low-resolution images.Compared with the compared YOLOv4 algorithm,the average accuracy and average recall rate are increased by 10.65% and 14.41%,respectively.Basically meet the requirements of efficient and rapid detection of ceramic substrate defects in low-resolution scenarios.3.A set of ceramic substrate defect detection software system is designed and developed for the requirements of ceramic substrate defect detection.The ceramic substrate defect detection software system integrates three modules: data set labeling,model training and defect detection.The data set labeling module can quickly label the ceramic substrate defect image and generate the data set that meets the training requirements of the ceramic substrate defect detection model.The model training module can train and optimize the model based on the ceramic substrate data set generated by the data set labeling module to obtain the defect detection model and detect the training results in real time.The defect detection module realizes the automatic detection of ceramic substrate defects,displays the detection results in real time on the system interface,and automatically saves the image of ceramic substrate with defects and the corresponding defect information.Without affecting the system performance,various defect information is collected in the background and saved in My SQL for subsequent checking and verification.