A multi-attribute decoupled learning network for precise segmentation of complex defects on steel surface

Abstract

In the steel manufacturing industry, precise segmentation of surface defects is critical for achieving high-quality product control. Steel surface defects encompass various types such as cracks, scratches, and weld slag, which exhibit significant disparities in size, shape, and textural features. Furthermore, steel surfaces are frequently characterized by complex background interference, which further increases the difficulty of segmentation. These complex scenarios render it difficult for existing segmentation methods to adapt stably to different defect types, leading to fluctuations in both accuracy and robustness. To address these challenges, this paper proposes an Attribute-Aware Routing and Dynamic Loss Weighting-based Adaptive Segmentation Model (ARD-SAM). By constructing an attribute-aware routing mechanism to guide features toward specialized segmentation heads and incorporating multiple dedicated losses with a dynamic weighting strategy, this approach achieves decoupled learning and differentiated optimization for defects. Experiment validation on a private casting billet surface defects dataset, collected from the production line of a major steel enterprise, demonstrates that ARD-SAM achieved a comprehensive score of 81.46%, representing an improvement of 13.39 percentage points over the leading baseline. The model exhibits superior accuracy and stability across multi-attribute coupled defects, providing a robust solution for fine-grained industrial quality control.