A Large Foundation Model Based Finetuning Network for Transmission Towers and Powerlines Segmentation

Nowadays, in the autonomous flight missions of UAVs (unmanned aerial vehicles) , precise observation and prompt decision-making are of paramount importance. Accurate segmentation and detection of transmission towers and powerlines from a UAV’s perspective is practically for ensuring safe and efficient path guidance for these UAVs. However, the complexity of the scene often makes it difficult to capture powerlines, as they are extremely thin and can easily blend into the background, posing a significant challenge in their detection. To solve this problem, we utilize a large foundation model known as the Segment Anything Model (SAM), which is designed to enhance the capacity for localizing and accurately segmenting powerlines within confusing scenes. Based on SAM, we propose a finetuning decoder to transfer the generalized class-agnostic knowledge of SAM to class-aware downstream task. Extensive experiments on multiple benchmark datasets (TTPLA、PLID、PLDU、PLDM) demostrate the efficiency of our proposed method. The comparison results with previous methods show that our method achieves the current state-of-the-art performance on transmission towers and powerlines segmentation task.