Instance Segmentation of Ice-Covered Transmission Line in Unmanned Aerial Vehicle Images

Unmanned aerial vehicle (UAV) detection of ice-covered transmission lines (ITL) is a crucial task for winter power grid maintenance. Instance segmentation offers robustness in handling complex scene variations in UAV images while providing instance-level information. However, the quality of instance segmentation heavily relies on precise pixel-level annotations, while predictions from low-cost annotations tend to be coarse. Therefore, this paper proposes a soft label mask optimization (SLMO) method to achieve high-quality segmentation results of ITL with only coarse annotations. Furthermore, to alleviate loss of feature details during multi-scale feature fusion, a pixel shuffle module (PSM) is introduced to replace the conventional interpolation sampling method. Combining these methods, this paper proposes a SLMO Network (SLMONet) based on the efficient SparseInst framework. Extensive experiments on our ITL datasets validate the effectiveness of SLMONet in segmenting ITL.