MSAM: a multi-scale attention mechanism for improving industrial defect segmentation

We propose a novel multi-scale attention mechanism (MSAM) that specifically addresses the challenges of environmental noise in industrial defect segmentation. The MSAM is composed of two submodules, namely the multi-scale feature filtering (MSFF) submodule and the scale selection (SS) submodule. The MSFF generates spatial attention weights that provide precise spatial location guidance for multi-scale features. This enhances the regions of interest while suppressing irrelevant regions. Additionally, the SS adaptively models the interdependencies between the channels of multi-scale features and assigns different weights to different channels. Experiments on the three industrial datasets demonstrate that the application of the MSAM module can effectively tackle the influence of complex noise in industrial segmentation and significantly improve the accuracy of industrial defect segmentation.