Combining Unsupervised Domain Adaptation and Semi-Supervised Learning for Power Line and Transmission Tower Segmentation

Abstract

Fully supervised image segmentation can effectively extract power line (PL) and transmission tower (TT) from aerial images. However, its performance is constrained by the lack of sufficiently detailed and high-confidence annotations. Furthermore, PL is the hard sample due to its slender shape and low proportion of feature information. To address the aforementioned challenges, this work innovatively introduces unsupervised domain adaptation (UDA) and semi-supervised learning (SSL) into the PL and TT segmentation task, and designs a new framework named UDASSL-Seg. Specifically, UDA is employed for pretraining, enabling the segmentation network to learn generic features and knowledge of hard sample. Subsequently, SSL is employed for fine-tuning, enabling the segmentation network to acquire generalization capabilities on the target dataset. Additionally, in order to further augment the segmentation network’s performance, the new designed dynamic co-perturbation consistency (DCPC) was proposed to extend the perturbation space by combining multiple image-level and dynamic feature-level perturbations. Extensive experiments were conducted on both self-built and public datasets. The results demonstrate the superiority of the proposed UDASSL-Seg over several state-of-the-art semi-supervised segmentation methods.