Research And Application Of Steel Surface Defect Detection Based On Improved YOLOv4 And Mask R-CN

Steel is an indispensable material for national construction and relates to the safety of infrastructure.During the production,processing,and use of steel,various types of defects may occur on its surface due to modern production processes and external factors.These defects not only affect the integrity and functionality of the steel but can also cause accidents and endanger personal safety.Intelligent,automated defect detection on steel surfaces is therefore of particular significance.To solve the problems of many types of steel surface defects,small targets,dense distribution,complex background,and high noise,this thesis optimizes and improves the YOLOv4 and Mask R-CNN algorithms intending to provide a more accurate and efficient solution for daily steel defect detection.The details of the research are as follows:In the data pre-processing stage,to improve the image quality,increase the amount of training data and solve the problem of less defective data,this thesis adopts the Ray distributed computing framework for data enhancement of existing image data.Experiments show that when the amount of tasks is the same,the time of data preprocessing is significantly shortened as the number of distributed nodes increases,and the method can effectively speed up the data pre-processing speed.To address the problems of time-consuming,high error rate,and low efficiency of manual and conventional defect detection,this thesis proposes a strip steel surface defect detection algorithm based on the improved YOLOv4 algorithm.A convolutional attention mechanism is embedded in the YOLOv4 backbone network and the path aggregation network is replaced with the RFB＿aug structure to improve the ability of the network model to extract relevant characteristic information;experiments demonstrate that the mean average precision of the YOLOv4 algorithm with the improved network structure is 2.66% higher than that of the original algorithm.Meanwhile,to reduce the number of parameters and the size of the model,the improved YOLOv4 algorithm was subjected to a structured pruning,and the pruned YOLOv4 model reduced the number of parameters by 19.7% compared with the improved YOLOv4 model,and improved the mean average precision by 2.02% compared with the original YOLOv4 algorithm.The pruned model is lighter,reducing network complexity and memory consumption,and is more suitable for deployment in real industrial environments.To address the challenges of the diversity of rail defect forms,randomness,and poor accuracy of currently used methods,a Mask R-CNN-based rail surface defect segmentation network is proposed.A new feature pyramid structure,DFPN,is used to design the detection network to achieve multi-scale fusion;a new evaluation indicators,CIOU,is adopted in the region proposal network to overcome the limitations of IOU in some special cases.Experimental results show that the new model,compared with Mask R-CNN,improves the mean average precision from 97.53% to 98.70% of the original algorithm with little increase in the number of parameters,and can extract defect features more finely and locate defect locations accurately.This thesis uses the flask framework to build a prototype system of steel surface defect detection,including single picture prediction,batch picture prediction,and video prediction functions.The prototype system applies a deep learning algorithm,which can quickly detect and identify steel surface defects,and provide high-definition image positioning information so that inspectors can more quickly and efficiently complete the detection of steel surface defects.