摘要
Abstract
Steel is a crucial material in industrial manufacturing,and any defects on its surface can significantly impact its functionality.Hence,it is crucial to conduct prompt and precise detection of faults on the surface of steel.This work presents an enhanced YOLOv7-IMN al-gorithm designed to accurately identify steel surface defects.Firstly,the parameter of network is decreased by developing the Inception-ELAN module,which utilizes large-kernel convolution to extend the receptive field and enhance the model's detection capability.Secondly,the multi-scale feature extraction(MFE)strategy is used in the backbone to acquire additional feature and detail information.The efficient multi-scale attention(EMA)mechanism is introduced to enhance the extraction ability of defect position information.Finally,the normalized wasser-stein distance(NWD)is introduced to improve the ability to detect small defects.The enhanced algorithm on the NEU-DET dataset achieves a detection accuracy of 81.3%,surpassing the YOLOv7 by 7.6%.Additionally,the number of parameters is reduced by 24.3%,amount of com-putation is reduced by 31.8%,and inference speed is improved by 19.3%.These improvements make the model highly suitable for industrial applications,as it offers superior detection accuracy and faster speed.关键词
缺陷检测/YOLOv7/特征提取/EMA/NWDKey words
defect detection/YOLOv7/feature extraction/EMA/NWD分类
计算机与自动化
