Improved Surface Defect Detection Algorithm For Hot-dip Galvanized Sheet Based On Deep Learning

With the rapid development of China’s economy,people’s demand for high-quality hot-dip galvanized sheet is increasing.However,surface defects will seriously affect the appearance and performance of the product.Therefore,it is very important to perform high-speed and accurate online surface defect detection on hot-dip galvanized sheet..At present,the detection of hot-dip galvanized sheet surface defects is still mainly manual,which not only has low efficiency,but also has a high rate of missed detection.Research and development of efficient hot-dip galvanized sheet surface defect detection equipment and systems has become a focus of common concern of iron and steel enterprises.Among many detection methods,the machine vision method is the current mainstream solution,in which the hardware part adopts a charge-coupled device camera array has basically reached consensus,and mainly improves the surface defect detection algorithm of the software part.After sorting out many traditional target detection algorithms based on candidate regions and classic target detection algorithms based on regression methods,this paper proposes two deep learning-based hot-dip galvanized sheet surface defect detection algorithms,considering the realization of online detection under high accuracy under different conditions.The main contributions are as follows:Introduced the related concepts of deep learning theory,clarified the concept of layers,introduced the calculation methods and design concepts of convolutional layer,pooling layer,and fully connected layer in detail,and explained the link relationship between the layers.The source of the model’s nonlinear expression ability is explained,and the optimization methods and strategies of network design are explained.Considering the higher computing power equipment,the classic target detection algorithm based on regression method is improved from three aspects and the characteristics of the hot-dip galvanized sheet surface defect data set are analyzed,thereby designing the precision and speed of hot-dip galvanized sheet surface detection The G-SSD algorithm with equal emphasis mainly makes the following improvements: First,a method for enhancing over-segmented data with overlapping regions on the image is designed to increase the relative proportion of small-sized targets,thereby improving the detection ability of small targets;second,in The base network part uses deep separable convolution to achieve acceleration;finally,improve its training strategy,increase the training complexity while further sparse the network model,and enhance its generalization ability.Another model is designed on a device with lower computing power.By adding the input information of the fourth channel,that is,the spatial information channel,to the output layer of the network,the visual activation and assisted manual re-examination are realized by means of the class activation graph algorithm.In general,a classification network can be used as the base network to design a G-Mobile Net model to achieve target detection and location capabilities.This changes the conventional thinking of traditional design target detection algorithms,speeds up the detection speed and greatly reduces the overall network structure.Number of parameters.By comparing the experimental results with different parameter values,the network structure and parameters suitable for actual production conditions based on the hot-dip galvanized sheet surface defect data set were determined.