基于自注意特征融合的钢材表面小目标缺陷检测  

Defect detection of small targets on steel surface based on self-attention feature fusion

作　　者：冯夫健 罗太维[1,2] 谭棉 汪小梅 王岳继[1,2] Feng Fujian;Luo Taiwei;Tan Mian;Wang Xiaomei;Wang Yueji(School of Mathematical Sciences and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;Key Laboratory of Pattern Recognition and Intelligent Systems of Guizhou Province,Guizhou Minzu University,Guiyang 550025,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵阳550025 [2]贵州民族大学贵州省模式识别与智能系统重点实验室,贵阳550025

出　　处：《电子测量技术》2024年第19期172-180,共9页Electronic Measurement Technology

基　　金：贵州省科技计划项目(QKHJCZK2022YB195,QKHJCZK2023YB143,QKHPTRCZCKJ2021007,QKHJCZK2022YB197);贵州省教育厅自然科学研究项目(QJJ2022015,QJJ2023061,QJJ2023012,QJJ2022047,QJJ2024063);贵州民族大学博士研究启动项目(GZMUZK[2024]QD04);贵州民族大学基金科研项目(GZMUZK[2023]YB14);贵州省模式识别与智能系统重点实验室开放课题(GZMUKL[2022]KF01)资助。

摘　　要：针对钢材表面小目标缺陷占比小,对比度低,导致钢材表面小目标缺陷检测模型提取丰富缺陷特征失效的问题。基于联系上下文信息和增强特征融合之间的关系,对钢材表面小目标缺陷检测问题提出以下解决方案:首先,结合滑动窗口机制Swin Transformer,利用分层结构和局部窗口整合不同特征块的特征信息,以在降低卷积操作密集性的基础上增强小目标缺陷特征信息的对比度;其次,采用坐标注意力机制使模型获得更多的位置信息,以增强小目标缺陷特征信息的多样性;最后结合具有丰富梯度流信息的特征融合模块CSP-FCN,提出了基于自注意特征融合的钢材表面小目标缺陷检测模型SFNet,该模型将不同尺度特征融合以产生丰富的语义信息,增强钢材表面小目标缺陷的特征表达能力。实验结果表明,SFNet在NEU-DET和GC10-DET公开数据集上的检测性能优于目前经典的目标检测模型。此外,所提模型在参数量减少为原来1/2的基础上平均精度值分别提升了3%和3.7%。In addressing the issue of ineffective extraction of rich defect features for small-scale surface defects on steel due to their low contrast and small proportion,this paper proposes a solution for small-target defect detection.Leveraging the relationship between contextual information integration and enhanced feature fusion,we introduce the following approaches:incorporating the sliding window mechanism Swin Transformer,which integrates feature information from different blocks hierarchically and through local windows to enhance the contrast of defect features while reducing convolutional operation density;the model employs Coordinate Attention to obtain more positional information,enhancing the diversity of features related to small-target defects.Additionally,we propose the steel surface small-target defect detection model SFNet based on self-attention feature fusion,integrating features with richer semantic information across different scales using the CSP-FCN feature fusion module.Experimental results demonstrate that SFNet achieves superior detection performance on the NEU-DET and GC10-DET public datasets compared to current classic object detection models.Furthermore,the proposed model achieves an average precision improvement of 3%and 3.7%,respectively,while reducing the parameter count to half of its original size.

关 键 词：钢材表面缺陷 小目标缺陷检测 Swin Transformer 位置信息 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TN919.8[自动化与计算机技术—计算机科学与技术]