A Weakly Supervised Contrastive Learning Pretraining Method for Visual Defect Detection of Transmission Lines

To address the issue of low detection accuracy caused by the similarity between inspection targets and backgrounds in power transmission line inspection images, this paper designs a visual defect inspection framework for electrical components based on weakly supervised contrastive learning pretraining. First, two power transmission line inspection datasets are constructed: one containing coarse labels and another with fine-grained labels. The coarse label dataset is used for weakly supervised contrastive learning pretraining, where the coarse label information distinguishing the target from the background is utilized to perform feature metric learning on one-dimensional (global) and two-dimensional (local) features. This significantly improves the ability to distinguish between background and target features in local regions. The fine-grained label dataset is used for supervised target detection model transfer learning, which provides precise target location and category information, further enhancing the model’s accuracy and robustness in detecting targets against complex backgrounds. Additionally, this paper introduces a deformable convolution operation at the global image level to learn the boundaries between the background and the target. Furthermore, a novel multi-path coordinated attention mechanism is designed to dynamically optimize the boundary offset, further enhancing the global feature distinction and target detection performance. Experimental results show that the proposed method improves the detection accuracy from 76.0% to 79.4% compared to the original algorithm. When transferred to similar target detection algorithms, the method achieves an improvement ranging from 0.7% to 2.8%, effectively enhancing the performance of existing algorithms in detecting targets in complex background environments. The dataset and code are available at https://github.com/19971483525/IEEE-TIM.