AGCA: An Adaptive Graph Channel Attention Module for Steel Surface Defect Detection
Xin Xiang, Zenghui Wang, Jun Zhang, Yi Xia, Peng Chen, Bing Wang
Abstract
Surface defect detection is an important part of the steel production process. Recently, attention mechanisms have been widely used in steel surface defect detection to ensure product quality. The existing attention modules cannot distinguish the difference between steel surface images and natural images. Therefore, we propose an adaptive graph channel attention (AGCA) module, which introduces graph convolutional theory into channel attention. The AGCA module takes each channel as a feature vertex, and their relationship is represented by an adjacency matrix. We perform non-local (NL) operations on features by analyzing graphs constructed in AGCA. The operation significantly improves the feature representation capability. Similar to other attention modules, the AGCA module has lightweight and plug-and-play characteristics. It enables the module easily embedded into defect detection networks. The experimental results on various backbone networks and datasets show that the AGCA outperforms state-of-the-art methods. Code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/C1nDeRainBo0M/AGCA</uri> .