Semantic Segmentation of Metal Surface Defects and Corresponding Strategies

Facing the demand of the industry for metal surface detection, this article addresses the segmentation of metal surface defects. Metal surface defect segmentation has many technical problems, such as serious imbalances in quantity and distribution, low contrast, and weak boundary information. To solve these problems, this article proposes a multi-scale attention feature fusion module, normalized mean square frequency category weight strategy, biased weight training sampling strategy, and category-boundary loss calculation strategy. The multi-scale attention feature fusion module fuses different levels of feature information to obtain defect features in images more effectively. The normalized mean square frequency category weight strategy suppresses the excessive weight of small defect categories and false positive classification. The biased weight training sampling strategy allocates different training frequencies to training samples and increases the attention to defect samples. The category-boundary loss calculation strategy imposes weights on the defect boundary and increases the attention to the boundary information. On four public datasets (Severstal, SD900, Magnetic-tile, and Crack), the proposed method achieves improvements of 7.8%, 4.0%, 18.4%, and 9.4% (absolute quantity), respectively, compared with the baseline method, while the calculation amount and the number of parameters remain almost unchanged. Hence, our method ranks as state of the art.