Research On Defect Detection Of Steel Surface Methods Based On Deep Learning

The presence of defects on the steel surface will directly determine the steel quality and product grade.Detecting defects on the steel surface,due to the problems of slow speed and low accuracy of photoelectric detection and image processing detection,this paper applies Deep Learning theory and technology to realize efficient and accurate detection of steel surface defects.The main content includes the following aspects:First,prepare the steel surface defect image data set.In the case of no data sets,the data sets are collected by network and marked by the tool Label Img.In view of the large number of defects and miscellaneous data in the data set,the statistical data of multiple defects are used to analyze the distribution and scale characteristics of various defects.Aiming at the shortage of steel surface defect data,a method of on-line dynamic data enhancement was proposed to enlarge the data set.A transfer learning method is proposed to make up for the lack of data,speed up model training and improve the ability and accuracy of model generalization.Secondly,improve the network model architecture.According to the different structure and performance of the backbone network,Res Net-50 is selected as the backbone network model by comparing the visual feature diagram and network parameter table.For the weak detection capability of ROI Pooling small targets,the ROI Align module is replaced by the ROI Pooling module to improve the accuracy of the small target regression box.Add DCN network module to strengthen the ability of extracting irregular defect points;The Feature Pyramid Network is used to modify the backbone network structure and enhance the feature extraction ability of multi-scale defects.In addition,the Faster-RCNN steel surface defect detection algorithm is improved.For multi-scale defects in the data set,K-means algorithm was used to regenerate 5 groups of priori frames to replace the original priori frames,so as to ensure the regression ability of the Regressor to multi-scale defects.The model step-by-step training method is proposed to accelerate the training speed.Histogram equalization was used to enhance the color brightness of the test data set.A method of simultaneous Center Loss based on Softmax was proposed to optimize the Loss function and improve the model generalization performance.Finally,the test data set shows that the m AP value and the FPS of the improved Faster RCNN algorithm are improved,the m AP value is increased from 73.6% of the origin Faster RCNN model to 75.1%,and the FPS is increased from 18 frames/s to 19 frames/s.