Power Tower Detection Method for SAR Images Based on Deep Learning

In high-resolution Synthetic Aperture Radar (SAR) images, power tower detection is challenging due to clutter interference and variations in local incident angles. To address this, we propose GA-YOLO, a power tower detection method based on the YOLO network, using GF-3 SAR images as experimental data. GA-YOLO offers two key advantages: (1) Improved feature extraction and loss function. We introduce the Global Convolution Attention Layer (GCAL) to enhance feature extraction for power tower detection. Additionally, a focal mechanism is incorporated to refine the traditional IoU loss function, improving the accuracy of detecting hard samples. (2) Knowledge distillation is applied to reduce computational complexity while maintaining high detection accuracy. Experimental results show that GA-YOLO achieves over 90% F1-score and AP@0.5, with minimal missed or false detections. Moreover, distillation learning effectively reduces model complexity while preserving high detection performance.