Cross layer feature aggregation network for multi-form power line instance detection

Abstract

The use of UAV for power line inspection has become a common method for real-time monitoring of power lines. However, automatic detection of multi-form power lines in complex backgrounds remains a highly challenging task. Most existing power line detection methods are based on segmentation techniques and are primarily focused on detecting linear power lines, lacking the capability to adequately represent multi-form power lines in complex environments. To address this issue, inspired by lane instance detection methods based on row detection, we propose a novel Multi-Form Power Line Instance Detection Network to detect various forms of power lines in real-world inspection scenarios. First, we leverage high-level semantic features to coarsely locate the power lines. Then, we refine and aggregate the image features of the power lines to achieve precise localization. In addition, to address the challenge posed by occlusions and other situations lacking visual clues, we introduce the OD_AttentionROI module, which allows the network to focus more on global contextual information, enhancing the representation of multi-form power lines. Finally, to tackle the problem of similarity computation for multi-form power lines, we propose the PLineIoU Loss, treating the pixel set of multi-form power lines as a whole to improve regression accuracy. Experimental results demonstrate that our detection framework can accurately detect multi-form power lines in real-world inspection scenarios. Moreover, the proposed framework can be easily extended to intelligent inspection tasks involving other multi-form linear or strip-shaped objects. The code is available at: https://github.com/DearPerpetual/MFPLNet.