PAM-DETR: Parallel Attention-MLP for Insulator Defect Detection

Insulators are indispensable components for reliable electrical power transmission, and the significance of insulator defect detection lies in safeguarding power system operation, mitigating potential hazards, and maintaining continuous and secure electricity supply. This paper proposes Parallel Attention-MLP for insulator defect detection (PAM-DETR), a novel approach with a highly adaptable encoder compared to other models. The encoder in PAM-DETR comprises parallel branches responsible for capturing global and local insulator defect features, along with channel attention for processing channel-dimension information in these features. Within the parallel branches, one branch employs a modified self-attention mechanism to capture long-range dependencies, while the other branch extracts local token relationships using MLP networks. Furthermore, we leverage channel attention instead of a feedforward network (FFN), enabling the model to prioritize channel-based information over spatial information. In the insulator defect datasets, our model achieves a higher Average Precision (AP) value compared to state-of-the-art.