Power-Synth: Synthetic Dataset for Power Line Inspection
收藏资源简介：
Accurately detecting power line defects under diverse weather conditions is crucial for ensuring power grid reliability and safety. Existing power line inspection datasets, while valuable, often lack the diversity needed for training robust machine learning models, particularly for adverse weather scenarios like fog, rain, and nighttime conditions. This paper addresses this limitation by introducing a novel framework for generating synthetic power line images under diverse weather conditions, thereby enhancing the diversity and robustness of power line inspection systems. The proposed approach employs a combination of novel heuristic image processing techniques, and a multi-domain Generative Adversarial Network (GAN) called StarGAN-v2. Initial transformations using heuristic methods simulate rain, fog, and night conditions, providing a foundation for the GAN to learn accurate mappings between weather domains. The StarGAN-v2 model, achieving its best performance with a latent dimension of 16, yielded a Frechet Inception Distance (FID) score of 24.72 and a Learned Perceptual Image Patch Similarity (LPIPS) score of 0.37 for fog, indicating high fidelity and perceptual similarity to real images. Furthermore, the impact of incorporating these synthetic images into the training process of various object detection models is thoroughly examined. The results show that models trained on a combination of synthetic and real data outperform those trained solely on either real data only or synthetic data only.
在多种天气条件下精准检测电力线路缺陷，对保障电网运行的可靠性与安全性至关重要。现有电力线路巡检数据集虽具备一定应用价值，但往往缺乏训练鲁棒性机器学习模型所需的数据多样性，尤其在雾、雨及夜间等恶劣天气场景下存在明显短板。针对这一局限，本文提出一种可在多样化天气条件下生成合成电力线路图像的新型框架，以此提升电力线路巡检系统的多样性与鲁棒性。所提方法结合了创新启发式图像处理技术与名为StarGAN-v2的多域生成对抗网络（Generative Adversarial Network, GAN）。首先通过启发式方法完成初始变换，模拟雨、雾及夜间场景，为生成对抗网络学习不同天气域间的精准映射奠定基础。StarGAN-v2模型在潜在维度为16时达到最优性能，针对雾天场景的弗雷歇特初始距离（Frechet Inception Distance, FID）得分为24.72，学习感知图像块相似度（Learned Perceptual Image Patch Similarity, LPIPS）得分为0.37，表明生成图像与真实图像具备极高的保真度与感知相似性。此外，本文还全面评估了将此类合成图像融入多种目标检测模型训练流程的影响。实验结果显示，结合合成数据与真实数据训练的模型，其性能优于仅使用真实数据或仅使用合成数据训练的模型。
