基于多尺度局部与全局上下文信息的钢材缺陷检测方法  

Steel Surface Defect Detection Model Based on Multiscale Local and Global Context Information

作　　者：张莉 付志鹏[1] 郭华平 孙艳歌 李锡瑞 ZHANG Li;FU Zhipeng;GUO Huaping;SUN Yange;LI Xirui(College of Computer and Information Technology,Xinyang Normal University,Xinyang 464000,China;College of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou 450001,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]郑州大学计算机与人工智能学院,河南郑州450001

出　　处：《信阳师范学院学报（自然科学版）》2024年第4期477-484,共8页Journal of Xinyang Normal University(Natural Science Edition)

基　　金：国家自然科学基金项目(62403405);河南省自然科学基金项目(222300420275);河南省科技计划项目(242102210092);河南省重点研发计划(241111212200);河南省研究生教育优质课程项目(YJS2022KC34);信阳师范学院“南湖学者奖励计划”青年项目。

摘　　要：钢材表面缺陷对许多工业产品的质量和性能有重大影响,会给生产带来巨大的经济损失。因此,对钢材表面进行实时监控,及时发现缺陷是非常有必要的。为提升对尺度差异较大、背景复杂的钢材表面缺陷的检测性能,提出一种基于多尺度局部与全局上下文信息的钢材缺陷检测模型。该模型使用卷积神经网络模型中带下采样机制的卷积操作,获取粗糙多尺度局部特征图,再使用自注意力机制分别在每一尺度作用于经过卷积提取的局部特征图,以获取像素间的长相关信息(如图像划痕、斑块、夹杂物等),增强缺陷类间判别能力;然后,采用特征金字塔结构,融合多尺度的特征图,以此提升对多尺度目标的检测能力;最后,引入通道与空间注意力模块与WIoU损失函数。实验结果表明,相比于Faster RCNN和EDDN等模型,该方法对于提升钢材表面缺陷检测性能行之有效。Steel surface defects have a significant impact on the quality and performance of many industrial products,which will bring huge economic losses to production.Therefore,it is very necessary to detect the steel surface in real time and find defects in time.In order to improve the detection performance of steel surface defects with large scale differences and complex backgrounds,a steel surface defect detection model based on multiscale local and global context information was proposed.Convolution operation with down-sampling mechanism in the convolutional neural network model was used to obtain rough multi-scale local feature maps.Then,self-attention mechanism was used to act on the local feature map extracted by convolution at each scale,to obtain long-distance dependencies between pixels(such as scratches,patches,inclusions,etc.),thus to enhance the inter-class discrimination ability of defects.Afterwards,the feature pyramid structure was used to fuse multi-scale feature maps,to improve the detection ability of multi-scale objects.Finally,channel and spatial attention module and WIoU loss function were introduced.The experimental results showed that,compared with algorithms such as Faster RCNN and EDDN,the proposed method was effective in improving the detection performance of steel surface defects.

关 键 词：自注意力 表面缺陷检测 卷积神经网络 多尺度特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]