Robust steel surface defect detection with edge-aware-semantic collaboration and efficient distillation network

Abstract Steel is a critical material for modern infrastructure, vehicles, and industrial equipment, driving the development of sectors including construction, manufacturing, and transportation. Surface defects on steel degrade both its mechanical properties and corrosion resistance, potentially causing structural failure or reduced service life. Current methods often struggle to delineate the indistinct boundaries of blurred defects, resulting in insufficient precision and high false-negative rates. To address these issues, we propose an edge-aware edge-semantic collaboration and efficient distillation network (3ESNet) for steel surface defect detection. To overcome the limitations in handling blurred boundaries and leverage complementary information, we develop an edge-aware feature fusion backbone within 3ESNet, which adaptively integrates edge and semantic streams for enhanced edge-sensitive feature extraction. Concurrently, to address scale variations and fragmentation, we design a dynamic edge-semantic FPN to employ adaptive weighting and multi-scale semantic fusion for robust feature integration. Moreover, a spatial attention module is constructed in 3ESNet to enhance the representation of high-level semantics. Finally, to improve accuracy without increasing model complexity or computation, we devise multi-decoder logical distillation to enable 3ESNet to achieve superior accuracy while preserving its lightweight design. Experiments on NEU-DET, GC10-DET, and APSDD datasets demonstrate that 3ESNet achieves 79.7% AP50 on NEU-DET with 16.1M parameters and 44.2 GFLOPs. Compared to advanced methods like D-FINE and YOLOv11, 3ESNet demonstrates higher accuracy, efficiency, and generalizability across diverse industrial scenarios.