Depth-Aware Feature Distillation Network for Surface Defect Detection

RGB-D surface defect detection methods leverage depth information to improve detection performance. However, their computational complexity often limits their applicability in resource-constrained scenarios. Besides, the limited availability of depth sensors makes acquiring depth data challenging during inference. To address these issues, this paper proposes a DepthAware Feature Distillation Network (DFDNet) for surface defect detection. DFDNet introduces a Dual-Attention Gated Interaction (DAGI) module to enable sufficient bidirectional cross-modal interaction, with a gating mechanism for adaptive feature fusion. A Multi-Scale Feature Transformation (MSFT) module enriches RGB feature representation by exploiting the multi-scale information, while the feature distillation strengthens spatial structure perception. During inference, this method relies solely on simulated RGB-D features for prediction to reduce model complexity and eliminate dependency on depth data.