Research On Surface Defect Detection Technology Of Battery Shell Based On Deep Learning

During the production and transportation of cylindrical steel battery shells,due to materials,stamping processes,friction,collisions,etc.,defects such as strains,scratches,pits,etc.will occur on individual battery shells,which seriously affect product quality and even Will cause safety hazards.Traditional digital image processing and machine learning technologies have been widely used in product defect detection,but for the defect detection of parts such as battery cases,due to the complex defect categories,it is difficult to extract complete features,and the detection effect is relatively poor.Inspired by computer vision technology,this topic is based on neural network technology to detect the surface defects of the battery case,which has reference significance for the defect detection research of the reflective rotating body similar to the battery case.The main contents of this paper are as follows:The characteristics and difficulties of battery case defect detection are analyzed,and the overall plan of defect detection is determined.Use machine vision technology to detect battery shell defects.The entire defect detection system is divided into hardware acquisition module,preprocessing module and neural network detection module according to functions.The three modules follow in time and advance in function,and finally The three modules are unified into one platform according to the logical relationship of the business to realize the whole process of defect detection.Aiming at the hardware acquisition module,this article selects the camera and lens according to the needs,and designs a special lighting scheme according to the characteristics of the reflection of the battery shell,which can achieve a relatively uniform distribution of light.For the picture preprocessing,this article selects a threshold-based segmentation method based on the characteristics of the battery case picture,and designs a segmentation algorithm,which can accurately segment the ROI that only contains the battery case area.Aiming at the problem of image enhancement,this paper compares multiple enhancement algorithms,verifies that the image contrast enhancement algorithm based on the exposure fusion framework is more suitable for reflective objects,and finally determines the use of semantic segmentation as the algorithm for defect detection in this article.The neural network detection algorithm is the core content of this subject.Aiming at the problem of battery shell defect detection,this subject collects pictures of various battery shell defects and makes them into a data set for training the segmentation network.According to the characteristics of detection,determine the appropriate loss function,optimizer and evaluation index.Use U-Net,HRNet and Fast SCNN three classic semantic segmentation networks to conduct experiments.According to the results,summarize the small sample and small target problems in this topic,and give two solutions for the improvement of the training program and the improvement of the network.By analyzing the results,it is determined to improve U-Net.Inspired by Inception network widening network channels and Res Net network deepening the number of network layers,this article combines the structural characteristics of Inception-Res Net V1 to transform the modules in Inception-Res Net V1 to make the encoder and decoder modular.The improved network is trained,and the optimizer and loss function are determined according to the situation.The final result is greatly improved compared with U-Net,and the average intersection ratio and defect category intersection ratio are both above 0.8.Finally,export the inference model,implement the Python deployment of the model,use the Py Qt5 development interface,encapsulate all functions,and realize the visual interface.The detection effect and the real-time requirements meets the task requirements.