Explicitly guided context adaptive optimisation networks for strip surface defect detection

The detection of surface defects on steel strips, a critical task in industrial quality control, is hampered by vast scale variations, complex morphologies, and elusive small targets. To overcome these limitations, we propose an Explicitly Guided Context-Adaptive Network (EGCA-Net). The network introduces three core innovations. First, an explicit scale-aware module (ESGFA) predicts defect scales to guide feature adaptation across layers. Second, an optimized attention mechanism (PGA+DLA) combines defect-specific priors with dynamic local attention, significantly enhancing sensitivity to tiny, low-contrast defects. Third, a context-aware feature fusion network (CA-BiFPN) intelligently aggregates multi-scale features using adaptive spatial weights and long-range dependency modeling, improving discrimination in complex backgrounds. Extensive experiments on generic (PASCAL VOC, MS COCO) and specialized industrial defect datasets (KolektorSDD, GC10-Det, NEU-DET) validate our approach. EGCA-Net demonstrates state-of-the-art performance, particularly in detecting challenging small targets, and exhibits strong robustness. Our work provides an accurate and adaptable solution for real-world industrial inspection.