基于通道和空间注意力的带钢表面缺陷显著性目标检测  

Channel and Spatial Attention Based Salient Object Detection of Surface Defects in Strip Steel

作　　者：郭华平 李锡瑞 张莉 孙艳歌 付志鹏 GUO Huaping;LI Xirui;ZHANG Li;SUN Yange;FU Zhipeng(College of Computer and Information Technology,Xinyang Normal University,Xinyang 464000,China;Henan Provincial Key Laboratory of Education Big Data Analysis and Application,Xinyang Normal University,Xinyang 464000,China)

机构地区：[1]信阳师范大学计算机与信息技术学院,河南信阳464000 [2]信阳师范大学河南省教育大数据分析与应用重点实验室,河南信阳464000

出　　处：《信阳师范学院学报（自然科学版）》2024年第4期470-476,共7页Journal of Xinyang Normal University(Natural Science Edition)

基　　金：国家自然科学基金项目(62002307);河南省自然科学基金项目(222300420275);河南省科技计划项目(242102210092);河南省研究生教育优质课程项目(YJS2022KC34);信阳师范大学研究生科研创新基金项目(2024KYJJ010)。

摘　　要：钢材表面缺陷检测在工业产品质量控制中越来越重要。由于钢材表面缺陷具有复杂背景、缺陷种类多样、尺度不一等特点,精确、高效地检测带钢表面缺陷仍然是一项极具挑战性的任务。针对这些问题,提出了一种基于通道和空间注意力的带钢表面缺陷显著性目标检测模型。首先,基于Transformer提取带钢图像的多尺度特征,以捕获目标图像的长距离依赖关系;接着,将获取的多尺度特征图送入所设计的两种不同的注意力模块(通道注意力模块和空间注意力模块),以强调带钢表面缺陷特征而抑制不相关的背景特征,从而加强对缺陷目标和背景的区分能力;最后,采用多尺度渐进融合模块融合多尺度特征图,以便不同尺度的特征信息能够进行互补,获取具有丰富语义信息的特征图,使模型能够更高效且精确地检测出带钢表面缺陷。大量实验结果表明,该模型在显著目标检测任务中具有显著的优势,表现出更高的准确性和更强的鲁棒性。Steel surface defect detection is becoming increasingly important in industrial product quality control.Due to the complex background,variety of defects and different scales of steel surface defects,accurate and efficient detection of strip surface defects is still a challenging task.In order to solve these problems,a strip surface defect salient object detection model based on channel and spatial attention was proposed.Firstly,based on Transformer,the multiscale features of strip steel images were extracted to capture the long-distance dependence of the target image;Then,the acquired feature map was fed into two different attention modules:The channel attention module and the spatial attention module,in order to emphasize the characteristics of the strip steel surface defect and suppress irrelevant background features,so as to strengthen the ability to distinguish between the defect target and the background;Finally,the multiscale progressive fusion module was used to integrate the multiscale feature map,so that the feature information of different scales could be complementary,obtaining feature maps with rich semantic information which enables the model to detect strip surface defects highly efficient and accurately.A large number of experimental results showed that the proposed model had great advantages in the salient object detection task with higher accuracy and stronger robustness.

关 键 词：带钢 缺陷检测 空间注意力 通道注意力 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]