Deformable Product Net: An Image Perception-Based Method for Power Line Defect Detection and Condition Assessment

The performance of power line structures is crucial for the reliable operation of transmission and distribution systems, with defects being the leading cause of failures. Current research lacks practical modeling and analysis of common defect types. Existing intelligent detection algorithms typically prioritize either prediction accuracy or computational efficiency but seldom address both simultaneously. In this study, we construct the 3-D spatial coordinates of power lines and define methods to calculate defect depth and area using mathematical relations, facilitating quantitative analysis and early-warning classification. We propose a lightweight network, deformable product net, which reduces model complexity by employing a simplified backbone and head structure. To balance prediction accuracy and computational efficiency, context anchor attention (CAA) and task-aligned learning (TAL) strategies are implemented. The network can dynamically adjust parameters to optimize performance for different tasks. Additionally, we have designed an upper computer software interface for real-time power line defect detection (PLDD), integrated with a hardware edge computing module, and explored the potential application of this method in semantic segmentation tasks. Experimental results demonstrate that our algorithm achieves state-of-the-art (SOTA) performance in terms of both parameter scale and prediction accuracy. Furthermore, to evaluate its portability and robustness, we perform additional comparisons using a power line defect verification dataset, two separate test sets, and the public PASCAL VOC 2007 datasets (a computer vision dataset). Our method outperforms existing approaches, and further comparative and ablation experiments confirm its efficacy. Our work encourages further exploration, and tasks related to the datasets can be accessed at: https://github.com/songzhiweiknight/Power-line-data-set.