An insulator defect detection algorithm based on an improved RT-DETR

Abstract

Insulator defect detection is critical to ensuring the safe and stable operation of power transmission lines. To address the challenges of high missed detection rates for small targets, weak multi-scale feature fusion capability, and poor real-time performance in existing detection models, an improved RT-detection transformer-based insulator detection model is proposed. Specifically, a lightweight backbone network, EfficientViT-SCSA, is introduced to enhance feature extraction capability while reducing model parameters. Meanwhile, a dynamic upsampling module, DySample, is adopted to improve fine-detail reconstruction ability while maintaining high computational efficiency. Furthermore, a neck network with multi-scale fusion and feature enhancement capability, termed GV-neck, is constructed to effectively bridge the backbone and the detection head. Finally, an Inner-MPDIoU loss function is introduced to impose additional geometric constraints on bounding-box regression, thereby improving localization accuracy, particularly for small defect targets. Experimental results on a self-constructed insulator defect detection dataset demonstrate that the improved model achieves superior performance over the baseline in terms of precision, recall, mean average precision, and inference speed. This method provides an effective theoretical reference for real-time insulator defect detection tasks.