Infrared Insulator Object Detection Algorithm Based on Improved Deformable DETR

Thermal infrared image detection technology is extensively employed in the monitoring and diagnosis of electrical equipment due to its non-contact and rapid advantages. Insulators, as vital components of substations, hold significant influence over the operational state of the power grid, and infrared imagery proves highly effective for their evaluation. However, the acquisition of infrared images presents challenges, particularly as most insulators are located in outdoor environments, resulting in predominantly small datasets for insulator infrared imagery. The existing Deformable DETR method exhibits limited detection performance on small datasets, and further, it does not adequately account for the substantial aspect ratio of insulators, contributing to suboptimal results. In this paper, we have made improvements to the Deformable DETR framework and introduced a hybrid matching approach to address the issue of slow convergence on small datasets. Additionally, we have enhanced the loss function to rectify the challenges arising from the substantial aspect ratio of insulators. Extensive experimentation on our dataset substantiates the superior performance of our method.