融合注意力机制的金属缺陷图像分割方法  

Metal defect image segmentation algorithm combined with attention mechanism

作　　者：赵鹤 杨晓洪[1] 杨奇 尹丽琼 ZHAO He;YANG Xiao-hong;YANG Qi;YIN Li-qiong(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Yunnan Kunming,650500,China;WISCOGroup Kunming Iron&Steel Co,Ltd,Anning Company,Yunnan Kunming,650302,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500 [2]武钢集团昆明钢铁股份有限公司安宁公司,云南昆明650302

出　　处：《光电子．激光》2021年第4期403-408,共6页Journal of Optoelectronics·Laser

基　　金：国家重点研发计划(2017YFB0306405);云南省重点研发计划(2018BA070);国家自然科学基金(61364008)资助项目。

摘　　要：由于金属表面缺陷图像的特性,有效精确分割是图像处理任务中的一大挑战。为了获得缺陷的类型、大小及位置信息,本文提出一种融合注意力机制的金属缺陷图像分割网络。该网络分为两条路径,语义信息路径主要由残差块构成的卷积网络获得特征图,采样过程中分步融合注意力机制以增强特征与背景对比度。旁路路径设计注意力机制模块获得位置信息的权重图,后将同尺寸的特征图与权重图融合,通过空间金字塔结合多尺度特征。实验结果表明,运用该算法可以提高金属表面缺陷图像的分割精度。Due to the characteristics of metal surface defect images,effective and accurate segmentation is a major challenge in image processing tasks.In order to obtain defect type,size and location information,this paper proposes a metal defect image segmentation network that incorporates attention mechanism.The network is divided into two paths.The semantic information path is mainly composed of a convolutional network composed of residual blocks to obtain a feature map.During the sampling process,the attention mechanism is integrated step by step to enhance the contrast between the feature and the background.The side road path design attention mechanism module obtains the weight map of the location information,and then combines the feature map of the same size with the weight map,and combines multi-scale features through the spatial pyramid.Experimental results show that the algorithm can improve the segmentation accuracy of metal surface defect images.

关 键 词：语义分割 膨胀卷积 注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]