LSKAFF-YOLO: Large Separable Kernel Attentional Feature Fusion Network for Transmission Tower Detection in High-Resolution Satellite Remote Sensing Images

Abstract

High-resolution satellite remote sensing technology provides an effective solution for the efficient and stable inspection of high-voltage transmission lines. The accurate extraction of transmission towers is crucial for leveraging satellite imagery in transmission line monitoring. This article addresses the challenges in detecting transmission towers, which are often obscured by complex backgrounds, variable object sizes, and small-scale objects. We propose the large separable kernel attentional feature fusion (LSKAFF)-YOLO network to enhance the precision of transmission tower extraction. This model incorporates LSKA-AFF module into the backbone network to extend the receptive field. By effectively leveraging the contextual information of satellite remote sensing images, it provides richer feature details for transmission tower positioning. Moreover, a progressive path aggregation network replaces the original neck network, mitigating information loss or degradation during feature transfer and interaction, thereby realizing multiscale feature fusion of transmission towers. To comprehensively evaluate the model’s performance, this study constructs Gao Fen tower dataset, a multiscene high-resolution satellite remote sensing transmission tower dataset, using GaoFen-2 and GaoFen-7 images, with references to publicly available datasets satellite remote sensing power tower dataset. The experimental results show that LSKAFF-YOLO attains mAP0.5 values of 88.8$\%$ and 94.6$\%$, precision of 80.5$\%$ and 89.3$\%$, recall of 89.2$\%$ and 93.5$\%$, and F1-score of 0.85 and 0.91, respectively. LSKAFF-YOLO outperforms other existing methods in terms of precision and overall performance in transmission tower detection.