Enhancing SAR Semi-Supervised Ship Detection with Hard-Sigmoid Weighted Pseudo-Labels and Space-Frequency-Selection Convolution

Synthetic Aperture Radar (SAR) images suffer from speckle noise, low contrast, and complex backgrounds, which degrade object detection. In maritime ship detection, semi-supervised object detection (SSOD) is essential when labels are scarce. We enhance the Soft Teacher framework by integrating hard-sigmoid Weighted Pseudo-Labels (WPL) and Space-Frequency-Selection Convolution (SFS-Conv) to dynamically weight pseudo-labels and extract both spatial and frequency cues. On the HRSID dataset, our method outperforms baseline approaches, showing particularly strong improvements in largescale vessel detection while maintaining competitive performance across other metrics.