Research On Surface Defect Detection Algorithm For Electronic Components For Edge Computing

With the development of intelligent manufacturing technology,more and more factories are undergoing digital and intelligent transformation,at the meantime various AI applications are appearing in factory workshops.However,due to the complexity of the industrial production environment,edge intelligence-related technologies have not been widely deployed in the field of intelligent manufacturing.Thesis focuses on the practical needs of automation,rapidity,and intelligence in the surface defect detection process of varister.Based on deep learning and image processing technologies,a deep neural network model was trained to detect whether there are defects on the surface of varister.Image processing algorithms were designed for edge devices,and an experimental system for varister detection and screening at the edge was built.The specific implementation contents are as follows:In the pre-processing of varister data,first,the varister image is collected,and then,based on the problem of impurities in the image background,image noise,a large proportion of useless information in the image,and non-uniform placement angles of varister in the collected data,a varister image pre-processing algorithm based on Open CV is designed and implemented to solve the problems in the collected images.In addition,targeted data augmentation was performed to solve the problem of a small dataset.In the varister detection algorithm,the Res Net model and Vision Transformer model were introduced for comparative testing.Two different strategies,self-training and transfer learning,were adopted for training on the cloud server.After training,the Res Net50＿trans model was selected as the varister detection model with the best performance,with an ACC of 0.984 and an AUC of 0.99,based on the comparison of the Loss,ACC,confusion matrix,ROC,and corresponding AUC graphs with the Epoch variation.In the optimization and deployment of the varister surface defect detection model at the edge,model quantization and deployment testing were carried out based on Open VINO for edge devices,and the inference speed and accuracy of different inference frameworks were compared.To meet the practical needs of detecting multiple varister in a single frame image in the actual factory detection environment,a varister multi-target extraction algorithm based on Open CV was designed and implemented.In the design and implementation of the varister detection and screening experimental system at the edge,the development of the camera control system and the execution mechanism control program was first completed.Then,a camera calibration algorithm and coordinate transformation algorithm were designed and implemented based on the requirements of camera distortion and the need to convert image pixel coordinates to execution mechanism world coordinates during execution mechanism grasping.Finally,all algorithms and programs were integrated,and the varister detection and screening system was developed and validated to achieve automatic detection and screening of varister.