Recognition And Classification Of Surface Defects Of Cylindrical Lithium Battery Steel Shell Based On Deep Learning

In the production of cylindrical lithium batteries,due to friction,bumps and other reasons,many defects such as deformation,pits,scratches,etc.will be generated,which seriously affects the quality of the product.The detection of surface defects in cylindrical lithium batteries becomes an indispensable part in its production process.Traditional visual inspection algorithms are based on manual designed features and different inspection algorithms are required to deal with different surface defects.It is difficult to form a unified inspection framework and the inspection results are difficult to meet the industrial production demand.In this study,the surface defect detection of cylindrical lithium battery steel shells is treated as object detection in computer vision and deep learning is used in the surface defect detection.The recognition and classification of the surface defects in image of cylindrical lithium battery steel shells is completed by object detection methods based on deep learning.In this research,1389 images of the surface of the cylindrical lithium battery steel shell were obtained through the lithium battery surface defect detection equipment,and the surface defects were divided into six types: pits,cracks,stains,hard printings,wire drawing and shell deformations.A deep learning data set with 1037 defect images was constructed,including 2061 pits,112 cracks,190 stains,241 hard printings,97 wire drawings and 50 shell deformation defects.We did this work with Label Image,a good annotation tool for deep learning object detection.Faster R-CNN,Cascade R-CNN and YOLO v3 were implemented on our dataset,and we compared and analyzed the performance of these three methods.Our research shows that YOLO v3 has higher accuracy in detecting defects,but at the cost of more missed detections;Faster R-CNN has a lower missed detection rate,but the accuracy of detected defects is also lower;Cascade R-CNN has a comprehensive performance in false detection and missed detection,and is more suitable for surface defect detection of cylindrical lithium battery steel shell.In terms of detection speed,Faster R-CNN can detect 12 images per second,Cascade R-CNN can detect 10 images per second,and YOLO v3 can detect 23.8 images per second,both of which can meet the requirements of lithium battery defect detection equipment.We made a detailed analysis of the results of Cascade R-CNN,and improved the performance of the algorithm by adjusting the confidence threshold,excluding minor defects,and eliminating the missing and mislabeling of labeled data sets.The average missed detection rate of 6 defects is 6.21%,and the average false detection rate is 3.91%.