Dual Regression-Based Semi-Supervised Infrared Image Instance Segmentation Method for Power Transmission Line Surveillance

Instance segmentation of infrared images is a key task for power transmission line surveillance. This paper proposes customized solutions to the following problems. First, for the semantic inconsistency in image translation, based on cycle-consistency concept, this paper presents a Dual-Regression (DR) framework, where forms a closed-loop learning between image translation and instance segmentation, thus improving the discriminative ability of instances. Second, for the data-label dependency in supervised training, this paper presents a semisupervised optimization scheme, which embeds the semantic prior extracted from a large number of unlabeled images into the model. The feature representations are enriched for highaccuracy instance segmentation. Experiments on real-world dataset demonstrate that the proposed method gains state-of-the-art performance compared with existing instance segmentation methods. Benefiting from the fine segmentation outlines and high category confidence, the proposed method is excepted to be invested in fully automatic instance annotation of power transmission line infrared images.