Research On On-line Detection Algorithm Of Steel Plate Surface Defects Based On Deep Convolutional Neural Network

The surface quality of steel plates is of great significance to industrial productions.Currently,the domestic demand for high-quality steel plates is constantly increasing,and the existing on-line detection technologies for surface defects of steel plates are gradually becoming difficult to meet the production demand.However,the visual algorithms based on deep convolutional neural network are significantly superior to the traditional visual algorithms,which provide a new idea for the improvement of visual tasks.Therefore,it is necessary to use the new technology of deep convolutional neural network to study the on-line detection algorithm of steel plate surface defects.The core of the on-line detection algorithm for steel plate surface defects includes image screening and defect target detection.This paper focuses on the research of these two parts by using new technology,and proposes an on-line detection algorithm for steel plate surface defects based on deep convolutional neural network.The main research contents and innovations are as follows:(1)Aiming at the image filtering task,an improved Res Net50 image filtering algorithm based on deep convolutional neural network is proposed.This paper first builds the Res Net50 basic model based on the model principle and task requirements,and then uses the hidden layer and the Dropout layer to change the fully connected layers’ structure of the basic model to design and obtain a new model,and through experiments and explorations to find the improvement methods to effectively improve the filtering accuracy of the Res Net50 new model,including Adam optimization algorithm,and three data enhancement methods of horizontal flip,vertical flip and brightness jitter.Through experimental verification,the accuracy of image filtering of the final improved model reached 0.9925,and the harmonic average F1-score of precision and recall reached 0.9685.(2)Aiming at the task of defect detection,an improved Faster R-CNN defect detection algorithm based on deep convolutional neural network is proposed.This article first builds the Faster R-CNN basic model according to the model principle and task requirements,and then uses the better backbone Res Net50,Feature Pyramid Networks(FPN)and Ro I Align structure to change the Faster R-CNN model implementation,and at the same time,through the experiments and explorations to establish effective methods for optimizing the new model,including the default anchor boxes adjustment,Online Hard Example Mining(OHEM),and four data enhancement methods: horizontal flip,vertical flip,brightness jitter,and "random walk".Among them,the "random walk" off-line data enhancement algorithm was designed according to the similar background characteristics of the steel plate images,which realized data expansion by randomly extracting defects and filling them reasonably with gradual boundary transition.Experimental results show that the mean Average Precision(m AP)of defect detection of the final improved model reaches 0.708.(3)According to the principle of traditional detection system,the improved Res Net50 model and improved Faster R-CNN model are combined to design and implement an on-line detection algorithm for steel plate surface defects based on deep convolutional neural network.Experimental results show that the on-line detection algorithm of steel plate surface defects proposed in this paper can achieve a high detection accuracy of 0.696 m AP under the condition of 82 FPS real-time detection speed,which verifies the effectiveness of the algorithm studied in this paper.