EMRA-Net: A Pixel-Wise Network Fusing Local and Global Features for Tiny and Low-Contrast Surface Defect Detection

Abstract

The problem of tiny and low-contrast surface defect detection is a nontrivial one. To solve the problems, this article proposes an edge and multi-scale reverse attention network (EMRA-Net), of which includes feature extraction and feature fusion. In the process feature extraction, the global dynamic convolution features, global dynamic multi-scale fusion (MSF) feature, and local pyramid edge feature are obtained through the pre-training backbone Resnet 34, the fresh MSF module, and the innovative pyramid edge module, respectively. In the process of feature fusion, these features are blended by a new self-learning scale module and a novel spatial channel domain reverse attention (SCRA) module step by step. The experimental results of five widely used datasets show that the EMRA-Net outperforms the existing methods. In addition, the mean intersection of union (mIoU)<xref ref-type="fn" rid="fn1">¹</xref><xref ref-type="fn" rid="fn2">²</xref> on the printed circuit board (PCB) industry dataset reaches 95.31&#x0025;. Moreover, the results of EMRA-Net indicated that the local edge feature can improve the performance of the defect detection network. The EMRA-Net has great potential in the application in the detection of tiny and low-contrast defects.