基于特征融合SSD的微电连接器缺陷检测  

Defect detection of micro-electrical connector based on multi-scale feature fusion SSD

作　　者：刘群坡[1,2] 方源 张建军 苏波[1] LIU Qunpo;FANG Yuan;ZHANG Jianjun;SU Bo(School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454000,Henan China;Henan International Joint Laboratory of Direct Drive and Control of Intelligent Equipment,Henan Polytechnic University,Jiaozuo 454000,Henan China)

机构地区：[1]河南理工大学电气工程与自动化学院,河南焦作454000 [2]河南理工大学河南省智能装备直驱技术与控制国际联合实验室,河南焦作454000

出　　处：《华中科技大学学报（自然科学版）》2022年第3期49-54,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家重点研发计划资助项目(2016YFC0600906);河南省高校科技创新团队项目(20IRTSTHN019);河南省创新型科技人才队伍建设工程资助项目(CXTD2016054);河南省科技攻关资助项目(212102210508)。

摘　　要：提出基于多尺度特征融合单次多框检测器(SSD)算法,对微精密玻璃封装电连接器的缺陷进行检测.针对SSD算法在背景复杂、噪声干扰多、目标特征微弱环境下,特征提取能力弱、检测精度低的问题,在主干网络加入深度残差结构,丰富细节信息;针对在卷积网络中关键信息逐步丢失问题,提出了自顶向下的多尺度特征融合方法将含有上下文信息的高语义层与含有位置信息的低层特征进行融合,使得融合后的特征层包含丰富的边界信息和背景信息;在此基础上,构建了一种轻量级的通道注意力模块,增强SSD算法对特征层重要特征的提取并弱化无关特征,从而提高SSD网络的特征提取能力.实验结果表明:改进算法相对于原始的SSD算法,精度由86.42%提高到了91.28%.a multi-scale feature fusion based SSD(sing shot multibox detector) algorithm was proposed to detect the defects in micro-precision glass-encapsulated electrical connectors. In order to solve the problem of weak feature extraction ability and low detection accuracy of SSD algorithm in the environment with complex background,multiple noise interference and weak target features,a depth residual structure was added to the backbone network to enrich the detailed information.Aiming at the problem of key information loss in convolution network,a top-down multi-scale feature fusion method was proposed to fuse the high semantic layer containing context information with the low-level feature containing location information,so that the fused feature layer contains rich boundary information and background information;based on which,a lightweight channel attention module was constructed to enhance the extraction of important features of feature layer by SSD algorithm and to weaken irrelevant features,thereby the feature extraction ability of SSD network was improved.Experimental results show that compared with the original SSD algorithm,the accuracy of the proposed algorithm is increased from 86.42% to 91.28%.

关 键 词：缺陷检测 深度学习 深度残差结构 多尺度特征融合 轻量级通道注意力机制 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]