An Improved RT-DETR Object Detection Method for Substation Equipment

This paper proposes an improved RT-DETR object detection algorithm tailored for substation equipment. Addressing challenges such as scale variations and occlusions in complex substation environments, we introduce two key innovations: (1) a multi-branch convolution module (conv3XC) that combines channel compression, feature extraction, and expansion to enhance multi-scale feature representation, and (2) a local window attention (LWA) mechanism that restricts attention ranges to improve local feature extraction and reduce computational overhead. Experiments on a publicly available substation equipment dataset demonstrate the model's superiority. With 97.6% precision and 89.7% [email protected], it outperforms RT-DETR, YOLOv8s, and DINO while maintaining moderate parameters (46.5M). Ablation studies confirm the individual contributions of conv3XC (+0.4% [email protected]) and LWA (+0.9% precision, reduced parameters). The model achieves an optimal balance between accuracy and efficiency, making it suitable for real-time substation monitoring applications.