Research on Geospatial Object Recognition Algorithm in Power Transmission and Transformation Engineering Based on FIseg-YOLO

The preliminary work of power transmission and transformation engineering design tasks relies on the precise survey of terrain to ensure the high-quality completion of design tasks. Therefore, developing a technology that can accurately and efficiently survey the construction area has become a hot topic in the field of electrical design research. However, identifying the location and precise boundaries of geospatial targets of different scales and densities from high-resolution natural environment images captured by drones remains a challenge. To address this issue, we improved upon the YOLOv8 algorithm by incorporating the concept of the YOLACT algorithm, proposing a novel instance segmentation algorithm—FIseg-YOLO, specifically designed for recognizing geospatial targets in power transmission and transformation line projects. This algorithm integrates various attention mechanism modules in the backbone network, effectively improving the accuracy of target recognition against complex backgrounds. Simultaneously, we optimized the neck network structure by enhancing the multi-path fusion of the network, achieving multi-level feature fusion between different layers. Additionally, the algorithm adds a multi-scale detection head to the basic network structure to extract shallow information from the input image and combine it with the feature fusion network to detect objects across a broader scale. Experimental validation on the geospatial orthophoto dataset shows that the mAP@.5 of the FIseg-YOLO model reached 67.3%, a 6.6% increase compared to the YOLOv8 base model, indicating a significant improvement in model performance. This improvement is of great significance for enhancing the efficiency of preliminary survey tasks in power transmission and transformation engineering.