多尺度增强特征融合的钢表面缺陷目标检测  

Object detection of steel surface defect based on multi-scale enhanced feature fusion

作　　者：林珊玲 彭雪玲 王栋 林志贤[1,2,3] 林坚普 郭太良 LIN Shanling;PENG Xueling;WANG Dong;LIN Zhixian;LIN Jianpu;GUO Tailiang(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362252,China;China Fujian Photoelectric Information Science and Technology Innovation Laboratory,Fuzhou 350116,China;School of Physics and Information Engineering,Fuzhou University,Fuzhou 350116,China)

机构地区：[1]福州大学先进制造学院,福建泉州362252 [2]中国福建光电信息科学与技术实验室,福建福州350116 [3]福州大学物理与信息工程学院,福建福州350116

出　　处：《光学精密工程》2024年第7期1075-1086,共12页Optics and Precision Engineering

基　　金：国家重点研发计划资助项目(No.2021YFB3600603,No.2022YFB3603705);福建省自然科学基金资助项目(No.2020J01468);国家自然科学基金青年基金资助项目(No.62101132);国家重点研发计划资助(2023YFB3609400)。

摘　　要：针对轻量级目标检测算法在钢表面缺陷检测任务中识别精度低的问题,提出一种多尺度增强特征融合的钢表面缺陷目标检测算法。该算法采用提出的自适应加权融合模块为不同层级特征自适应计算融合权重,将深层语义与浅层细节进行加权融合,使得浅层特征在不丢失细节信息的同时获得丰富的深层语义。利用提出的空间特征增强模块从3个独立方向强化融合特征,通过引出残差旁路增强网络结构的稳定性,使卷积过程能够挖掘到更多的关键信息。根据先验框与真实框的整体交并程度为模型选择更为合适的训练样本。实验结果表明,该算法的检测精度达到80.47%,相比原始算法提升6.81%。该算法的参数量为2.36 M,计算量为952.67 MFLOPs,能快速且高精度检测钢材表面的缺陷信息,具有较高的应用价值。To address the issue of low recognition accuracy in lightweight algorithms for steel surface defect detection,this paper introduces a Multi-scale Enhanced Feature Fusion(EFF)technique.Initially,an Adaptive Weighted Fusion(AWF)module calculates fusion weights adaptively for different feature levels.This allows shallow features to enrich with deep semantics without compromising detail.Subsequently,the Spatial Feature Enhancement(SFE)module boosts the fused features from three distinct directions and improves network stability by integrating residual pathways,enabling the convolution process to extract more critical information.The model then selects better training samples based on the overlap between the prior box and the ground truth.Experimental outcomes show that the proposed method achieves a detection accuracy of 80.47%,marking a 6.81% increase over the baseline algorithm.Moreover,with 2.36 M parameters and 952.67 MFLOPs,this algorithm efficiently and accurately identifies steel surface defects,demonstrating significant practical utility.

关 键 词：缺陷检测 单发多框检测器 增强特征融合 自适应加权融合 空间特征增强 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]