Knowledge-Guided Rotated Network for Power Line Detection in Infrared Images

Detecting power lines accurately remains a challenging task for low-altitude aircraft due to their low radiation and complex background, making them one of the most perilous impediments for such aircraft. To overcome this issue, in this letter, a knowledge-guided rotated network (KRNet) is proposed for power line detection in complex infrared scenes. The proposed method consists of the following main three steps. First, we develop a linear-distribution perception module (LDPM), which enhances feature representation by extracting and exploiting important linear and spatial distributional prior knowledge. Second, to more accurately locate the power line, we design a novel loss function rotation-aware complete intersection-over-union (RACIoU), which utilizes the prior knowledge of power line non-directionality to guide model learning. Finally, using this prior knowledge of the power line, we obtain good detection results and verify the effectiveness and advantages of the proposed method using the power line rotation detection dataset (PRDD) that we built. Our method outperforms other state-of-the-art benchmarks for power line detection.