Detection And Recognition Of Strip Surface Defects Based On Deep Learning

As one of the main products of steel,strip steel is widely used in building structure,aviation rocket,automobile manufacturing,household appliances,machinery manufacturing,electronic instruments and other fields.Due to the complexity of strip processing environment,there are many defects on strip surface.For defects in multi-objective and complex scenes,it is difficult to extract defect features based on manual and machine vision detection methods.Therefore,in this paper,deep learning target detection algorithm is adopted,and the excellent feature extraction ability of convolutional neural network is utilized to achieve the classification and location of six kinds of strip surface defects.The main research contents and achievements of this paper are as follows:By comparing and analyzing the mainstream target detection algorithms,YOLOv4 network is selected as the benchmark model,and its network structure and algorithm principle are deeply studied.By analyzing the experimental results of six kinds of strip surface defects detected by YOLOv4 network model,the following three improvements are made to the network model:(1)In view of the low accuracy of YOLOv4 network model in identifying crazing,rolled-in＿scale and other small target defects on strip steel surface,this paper designed a multi-scale feature prediction network and added a detection layer of 152×152 shallow network,in order to improve the network model of crazing,rolled-in＿scale and other small targets recognition accuracy.(2)In order to make the prediction box output by the network model more consistent with the dimension characteristics of the six types of strip surface defects,the K-means++clustering algorithm is introduced to re-cluster the data sets of the six types of strip surface defects,and the prior boxes suitable for the data sets of the six types of strip surface defects are selected to improve the accurate positioning of the six types of strip surface defects by the network model.(3)In order to extract the features of six kinds of defects on strip surface more efficiently,channel attention mechanism is embedded in the heads of four detectors in the network model.This mechanism can enhance the attention between channels and improve the extraction of key feature information by network model,so as to improve the identification accuracy of six kinds of strip surface defects by network model.In this paper,the YOLOv4-III network model improved by the above three points was adopted to test the six types of strip surface defects on the test set.The m AP value of the six types of strip surface defects reached 94.6%,which was 9.69 percentage points higher than the original YOLOv4 network model.The AP values of crazing and rolled-in＿scale are significantly improved,30.47 percentage points and 11.03 percentage points higher than the original YOLOv4 network model respectively.The experimental results show that the optimized YOLOv4-III network model can accurately detect the categories and locations of six kinds of defects on the strip surface and has a good practical application significance.