GSAC-YOLO:Global-Selective Attention and Convolutional Networks for Transmission and Distribution Tower Detection in High-Resolution Satellite Remote Sensing Images

With the rapid advancement of remote sensing technology, the detection of transmission and distribution towers based on high-resolution satellite remote sensing images is gradually becoming a key approach for large-scale automated power grid inspection. Although existing detection methods have achieved certain progress in this task, they still face notable challenges when dealing with power tower that are frequently obscured, especially in accurately identifying distribution towers that are easily confused with backgrounds such as streetlights and trees. To address these issues, this paper proposes the global-selective attention and convolution-you only look once (GSAC-YOLO) network model. GSAC-YOLO incorporates two plug-and-play innovative modules, one structural enhancement operation, and an innovative bounding box generation system: the pooling-grouped dual-driven feature extraction module (PGD FEM), the information dynamic integration module (IDIM), multiple up-sampling operations (MUO), and the optimal bounding box generation system (OBGS). Among them, PGD-FEM and IDIM respectively enhance the model's global perception capability and selective feature fusion mechanism, enabling the model to better capture the distinct structural characteristics of transmission and distribution towers; MUO reduces information loss during feature transmission; and OBGS improves the final localization accuracy. These improvements are designed to remain lightweight as much as possible, in order to avoid compromising detection efficiency. Experimental results on the SRSPTD dataset demonstrate that the GSAC-YOLO network model achieves <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</i>1 score and mAP50 of 0.787 and 78.3%, respectively. Compared with other detection methods, GSAC YOLO achieves superior detection accuracy. This paper's code is available at: https://github.com/CNOFT0/GSAC-YOLO.