Liu Yue, Huang Xinbo
School of Electrical Engineering, Xi'an University of Technology, Xi'an 710054, China.
School of Electronics and Information, Xi'an Polytechnic University, Xi'an 710048, China.
Sensors (Basel). 2024 Jan 10;24(2):428. doi: 10.3390/s24020428.
Regular inspection of the insulator operating status is essential to ensure the safe and stable operation of the power system. Unmanned aerial vehicle (UAV) inspection has played an important role in transmission line inspection, replacing former manual inspection. With the development of deep learning technologies, deep learning-based insulator defect detection methods have drawn more and more attention and gained great improvement. However, former insulator defect detection methods mostly focus on designing complex refined network architecture, which will increase inference complexity in real applications. In this paper, we propose a novel efficient cross-modality insulator augmentation algorithm for multi-domain insulator defect detection to mimic real complex scenarios. It also alleviates the overfitting problem without adding the inference resources. The high-resolution insulator cross-modality translation (HICT) module is designed to generate multi-modality insulator images with rich texture information to eliminate the adverse effects of existing modality discrepancy. We propose the multi-domain insulator multi-scale spatial augmentation (MMA) module to simultaneously augment multi-domain insulator images with different spatial scales and leverage these fused images and location information to help the target model locate defects with various scales more accurately. Experimental results prove that the proposed cross-modality insulator augmentation algorithm can achieve superior performance in public UPID and SFID insulator defect datasets. Moreover, the proposed algorithm also gives a new perspective for improving insulator defect detection precision without adding inference resources, which is of great significance for advancing the detection of transmission lines.
定期检查绝缘子运行状态对于确保电力系统的安全稳定运行至关重要。无人机（UAV）巡检在输电线路巡检中发挥了重要作用，取代了以往的人工巡检。随着深度学习技术的发展，基于深度学习的绝缘子缺陷检测方法受到越来越多的关注并取得了很大的进步。然而，以往的绝缘子缺陷检测方法大多侧重于设计复杂精细的网络架构，这会增加实际应用中的推理复杂度。在本文中，我们提出了一种新颖的高效跨模态绝缘子增强算法，用于多域绝缘子缺陷检测，以模拟真实复杂场景。它还在不增加推理资源的情况下缓解了过拟合问题。高分辨率绝缘子跨模态转换（HICT）模块旨在生成具有丰富纹理信息的多模态绝缘子图像，以消除现有模态差异的不利影响。我们提出了多域绝缘子多尺度空间增强（MMA）模块，以同时增强不同空间尺度的多域绝缘子图像，并利用这些融合图像和位置信息来帮助目标模型更准确地定位各种尺度的缺陷。实验结果证明，所提出的跨模态绝缘子增强算法在公共UPID和SFID绝缘子缺陷数据集上能够实现卓越的性能。此外，所提出的算法还为在不增加推理资源的情况下提高绝缘子缺陷检测精度提供了新的视角，这对于推进输电线路检测具有重要意义。
