SAR Superpixel Segmentation Based on Spatial Attention Network

Superpixel segmentation is a key preprocessing step in image analysis, and deep learning-based methods have made significant progress in optical images. However, their application in Synthetic Aperture Radar (SAR) images is challenged by speckle noise and low contrast. Superpixel Sampling Network (SSN), an end-to-end trainable deep learning method originally designed for optical images, is applied to the SAR image dataset SSDD in this study. In order to cope with the noise characteristics of SAR images, we add a spatial attention model to the network architecture of SSN to enhance its attention to critical regions and improve its robustness to noise. In addition, we use <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\operatorname{IoU}$</tex> (Intersection over Union) loss instead of cross-entropy loss to directly optimize the overlap accuracy of the superpixel region, so as to improve the boundary regression accuracy. Experimental results show that the proposed method achieves 86% boundary recall on SSDD datasets, which significantly improves the superpixel segmentation ability in the SAR domain.