The Research Of Surface Defect Detection Based On Deep Learning

Surface defect detection is an active research direction in target detection,and it is also a technology with great practical significance,which can play a huge role in improving factory automation and intelligent manufacturing.With the development of target detection based on deep learning in recent years,the development of surface defect detection has also made a qualitative leap.However,the accuracy of surface defect detection is still disturbed by factors such as a large number of small targets,unbalanced samples,and a small number of samples.Therefore,how to overcome these factors and improve the detection accuracy is a difficult point in research.At present,solving these problems mainly includes improving network structure and innovative data enhancement methods.In order to improve the accuracy of surface defect detection,this paper proposes a surface defect detection network based on the fusion of deformable convolution Faster RCNN and FPN.This network mainly consists of Faster RCNN,FPN,Deformable Convolution,RoI Align.The FPN feature map pyramid can extract strong semantic information because of the fusion of low-level features and high-level features;Deformable Convolution makes the network more adaptable to the change of target shape;RoI Align can solve the problem of RoI Pooling mismatch(misAlignment)and reduce The characteristic information of the small target is lost.In this paper,a comparative experiment is performed on the fusion network of Faster RCNN,Faster RCNN and FPN,and the surface defect detection network based on the fusion of deformable convolution and Faster RCNN and FPN.The experimental results on the aluminum profile data set show that Faster RCNN and FPN The FPN fusion network is 2.023%higher than the Faster RCNN and FPN fusion network mAP,and 4.279%higher than the Faster RCNN mAP.The experimental results show that the surface defect detection network based on the fusion of Faster RCNN and FPN based on deformable convolution.Aiming at the difficulty of detecting small targets in the field of surface defect detection and the impact of sample imbalance on the detection effect,this paper proposes a surface defect detection method based on enhanced data of small targets.This method is used to enhance the data of aluminum profiles:there are a large number of small targets in the data set,such as scratches,dirty spots,and paint bubbles,which are difficult to detect.Copy and paste the small targets of scratches,dirty spots,and paint bubbles many times.Increase the number of small targets(that is,expand the small targets);increase the number of samples by copying the targets of the samples with a small number of brushes,sprays,and paint bubbles in the original data set to the non-defective samples with similar backgrounds,so that the data set The sample is more balanced.Comparative experiments were conducted on three networks:Faster RCNN,Faster RCNN and FPN fusion network,and Deformable Convolution-based Faster RCNN and FPN fusion surface defect detection network.The enhanced data set increased by 2.614%compared to the original data set mAP,2.469%,2.48%,the experimental results show that the use of surface defect detection method based on small target expansion data enhancement for aluminum profile data set can significantly improve the detection accuracy.