LFAF-DETR: Lightweight Dynamic Feature-sensing Fusion Network for Industrial Metal Surface Defect Detection

Defect identification in industrial sheet metal is critical to ensure quality control in the manufacturing process. However surface defect detection faces severe challenges in balancing detection accuracy, efficiency and model complexity in industrial environments. In this study, we propose a lightweight feature-aware fusion network to optimize feature sensing for defect recognition. First, we propose a gated feature aggregation module This module combines multi-scale depth-separable convolution with dynamic minus-selection gating to achieve adaptive fusion of local details and global semantics. Second, we propose a bidirectional cross-attention that improves cross-scale geometric feature differentiation by simplifying two-path channel alignment and multiscale interaction. Finally, a dynamic adaptive focused attention module that collaboratively optimizes spatial channel feature localization and contextual aggregation through a heterogeneous attention path approach. Experiments on NEU-DET and GC-10 datasets show that the proposed method has excellent performance in terms of accuracy and parameters.