Research On Lightweight Surface Defect Detection Algorithm Based On Deep Convolutional Neural Network

In industrial production,surface defect detection is one of the key methods to control product quality.In the early days,it mainly relied on manual detection,which had the problems of high rate of missed detection,low efficiency,and slow speed.Although the defect detection algorithm based on traditional machine learning can greatly improve the detection efficiency,it requires manual feature extraction,the design process is complicated,and it relies too much on the experience of experts.In recent years,deep learning has developed rapidly,especially the development of convolutional neural networks,which has made a major breakthrough in the field of computer vision.Therefore,the surface defect detection algorithm based on deep learning plays a very important role in improving industrial production efficiency.However,at present,a large amount of research on convolutional neural networks is aimed at building deeper and more complex network structures to improve accuracy.These networks have high computational complexity and it is difficult to meet the requirements for real-time detection speed in industrial production,so they cannot be directly applied to defect detection.In response to this problem,this thesis combines the characteristics of surface defect detection,with the goal of establishing lightweight surface defect detection algorithm,mainly doing the following work:(1)This thesis proposes an end-to-end and lightweight surface defect detection network(LSDDN).Based on the idea of YOLO directly predicting the bounding box and class probability,a lightweight defect detection network was designed.The network is composed of many basic modules.The basic modules are composed of channel split,1×1convolution,depthwise separable convolution and channel shuffle.This structure can significantly reduce network parameters.LSDDN verified its effectiveness on the DAGM2007 dataset,with an average classification accuracy of 97.23% and the prediction time for a picture was only 20 ms.(2)This thesis proposes a lightweight surface defect detection network based on improved SSD(LW-SSD).This thesis realizes the lightweight of the original SSD network,and at the same time improves its lack of location loss.Experimental results show that the improvement of the loss function can increase the m AP value by 1.51%.The average classification accuracy of LW-SSD reaches 97.57%,and the time to predict a picture is 37 ms.(3)This thesis proposes a new pruning method based on joint zero-value percentage to further compress the LW-SSD model.This pruning method can greatly reduce the amount of network parameters and significantly improve the detection speed without sacrificing the minimum classification accuracy and defect localization accuracy.The experimental results show that the average classification accuracy of LW-SSD after pruning is 97.03%,which is only 0.54% lower than that before pruning,but the time to predict a picture is indeed shortened by 16 ms to only 21 ms.