DSA-YOLO: Coupling Spectral and Spatial Cues for Lightweight Strip Steel Surface Defect Detection

Abstract

Strip steel surface defect detection is a critical task in industrial quality control, yet it remains persistently challenging owing to the low contrast between defects and the rolling texture, the sparse and uneven defect distribution, and the highly variable defect morphology. Existing detectors tend to operate in the spatial domain with uniformly allocated computation, making them ill-suited to these characteristics. This article proposes DSA-YOLO, a lightweight one-stage detector built upon YOLO11-N that addresses these issues through three complementary modules. A dual-domain adaptive spectral enhancement (DASE) block applies magnitude-based spectral filtering and spatial-residual-guided modulation in the shallow backbone stages to strengthen weak defect responses against the rolling-texture background. A bi-level routing attention (BRA) unit routes self-attention through a sparse subset of defect-relevant regions, modeling long-range dependencies among scattered defects while suppressing background computation. An adaptive synergistic pyramid network (ASPN) couples linear deformable convolution with large-kernel and orthogonal strip convolutions, and adopts concatenation-based fusion to preserve weak, shape-variant responses. With only 3.27 M parameters and an inference speed of 154.4 FPS, DSA-YOLO achieves an $\mathrm {mAP}_{50}$ of 81.60% on NEU-DET and 73.8% on GC10-DET, demonstrating consistent effectiveness across two public strip-steel defect benchmarks and its potential for real-time industrial inspection.