Based On The Deep Learning Strip Surface Defect Detection And Identification Methods Of Research

Iron and steel industry,as the pillar industry of national economy,plays a pivotal role in the process of maintaining rapid economic development.As one of the important products in the iron and steel industry,strip steel is widely used in automobile,national defense,military industry and other fields.The quality of strip steel directly affects the quality of products.In order to meet the increasing demand,it is of great significance to develop an efficient and accurate strip steel surface defect detection system.Considering the complexity of strip production environment,deep learning and surface defect detection algorithm are studied in this paper.The existing target detection network and target classification network are improved to realize the detection and recognition of strip surface defects based on deep learning.The main research results of this paper are as follows:1、Design of strip surface defect classification network.Aiming at the limited data sets of strip surface defects,the methods of data expansion,such as image rotation and brightness change,are used to simulate the complex production environment to get enough data sets;The Dense Net network was selected as the classification network,and the transfer learning method was adopted to shorten the training time.In the training process,label smoothing,learning rate attenuation,Dropout layer and other training techniques were used to prevent over-fitting.Finally,the classification accuracy of 98.72% was achieved on the test set.2、Design of strip surface defect detection network.YOLOV3 network is a classic target detection network.In order to ensure that YOLOV3 can extract more accurate high-dimensional features in line with the surface defects of strip steel,the network structures of Inception Net and Dense Net were integrated into the feature extraction network;In order to make the bounding box output of target detection more consistent with the size information of strip defects,the clustering algorithm was improved and the anchor box size of YOLOV3 was optimized by means of Kmeans++;Considering the redundant features in YOLOV3,attention mechanism was introduced to suppress the redundant features.After three improvements,the New-YOLO network was obtained,and the m AP value of each defect reached84.17%,an increase of 5.43% compared with that before the improvement.3、Design and construction of GUI interface of defect detection and identification system.In order to improve user experience and enable users to quickly master using methods without professional knowledge,Py QT5 is used to build GUI interface of defect detection and identification system.Through this GUI interface,non-professionals can also carry out network training and defect detection.