基于全局感知和高斯分布的钢材缺陷检测  

Steel defect detection based on global perception and Gaussian distribution

作　　者：苏佳[1] 贾泽 秦一畅 侯卫民[1] SU Jia;JIA Ze;QIN Yi-chang;HOU Wei-min(School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang 050018,China)

机构地区：[1]河北科技大学信息科学与工程学院,河北石家庄050018

出　　处：《计算机工程与设计》2024年第12期3688-3694,共7页Computer Engineering and Design

基　　金：国家自然科学基金青年科学基金项目(62105093)。

摘　　要：针对钢材表面缺陷检测易受背景干扰且尺度信息差距大等问题,提出一种改进的钢材表面缺陷检测算法MN-Yolo。引入全局感知GC模块,能够有效利用上下文信息,增强特征提取能力;设计一种多尺度融合模块MSF,以便更深层次整合多种尺度的特征信息,提高网络检测性能;引入WDloss,通过使用NWD度量BBox的相似性,改善对小目标缺陷的误检和漏检问题。在NEU-DET上验证可知,改进算法与原Yolov7-tiny算法比较,mAP、Precision、F1分别提升4.7%、5%、5.5%,FPS可达141,结果验证了改进算法的有效性。Aiming at the problems of steel surface defect detection being susceptible to background interference and large scale information gap,an improved steel surface defect detection algorithm MN-Yolo was proposed.A global perception GC module was introduced,effectively utilizing contextual information to enhance the feature extraction capability.A multi-scale fusion module MSF was designed to integrate multi-scale feature information at a deeper level and the network detection performance was improved.The WDloss was introduced to improve the false detection and missed detection of small target defects using NWD to measure the similarity of BBoxes.Validation on NEU-DET shows that compared with the original Yolov7-tiny algorithm,the mAP,Precision and F1 of the improved algorithm are improved by 4.7%,5%,and 5.5%,respectively,and the FPS reaches 141,verifying the effectiveness of the improved algorithm.

关 键 词：缺陷检测 高斯分布 深度学习 背景干扰 上下文信息 多尺度融合 全局感知 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]