Power-DETR: end-to-end power line defect components detection based on contrastive denoising and hybrid label assignment

Abstract

Maintenance of power transmission lines is essential for the safe and reliable operation of the power grid. The use of deep learning-based networks to improve the performance of power line defect detection faces significant challenges, such as small target sizes, shape similarities, and occlusion issues. In response to these challenges, a transformer-based end-to-end power line detection network called Power-DETR is introduced. Initially, building upon Deformable DETR, a large pre-trained model (Swin-large) is utilized to increase the number of multi-scale features, and activation checkpoint technology is applied to ensure effective training within limited memory capacity. Subsequently, a contrastive denoising training strategy is integrated to combat ambiguity and instability of the Hungarian matching algorithm during training, aiming to expedite model convergence. Additionally, a hybrid label assignment strategy combining OHEM and cost-based ATSS is proposed to provide the model with high-quality queries, ensuring adequate training for the decoder and enhancing encoder supervision. Experimental results substantiate the efficacy of the proposed Power-DETR model as a novel end-to-end detection paradigm, surpassing both one-stage and two-stage detection models. Furthermore, the model demonstrates a significant 15.7% enhancement in mAP0.5 compared to the baseline.