Infrared Image Object Detection of Substation Electrical Equipment Based on Enhanced RT-DETR

In response to the challenges posed by low contrast in infrared images of substation electrical equipment, the presence of small targets, and irregular target shapes in power systems, as well as the industry’s demands for model lightweighting and real-time detection, this paper proposes an improved real-time object detection model based on RT-DETR. The improved model enhances feature learning by integrating multi-path gradient information within the ResNet18 network, significantly reducing computational burden and redundant parameters. In the neck network, GConvC3, which utilizes depthwise separable convolutions and a Sigmoid activation function, extracts information from the fusion of different scale features during cross-scale feature interaction, further decreasing computational complexity and improving the efficiency of feature interaction. To address the characteristics of various targets and the limitations of the original loss function, Focaler-EIoU is employed as the bounding box regression loss in the classification-regression detection head, effectively solving the localization issues for small and irregularly shaped targets, and improving both precision and the model's convergence speed during training. Experimental results indicate that the improved model achieves a 1.9% increase in mAP and a 2.5% increase in AP50 while reducing the number of parameters and computational load by 46% and 51.8%, respectively, with an FPS of 148, thus meeting requirements for both accuracy and real-time detection.