Osroad-cycleGAN: An Efficient Approach for SAR Road Image Augmentation

Deep learning algorithms for image interpretation tasks like object detection, segmentation, and recognition require large, high-quality labeled data. However, SAR (Synthetic Aperture Radar) images pose distinct challenges in the labeling process due to their unique scattering mechanisms. Transforming optical images into SAR images via generative adversarial networks (GANs) offers a new approach for SAR data augmentation. Focusing on road extraction in SAR images, we have refined the cycleGAN architecture in this paper. Dilated convolution was introduced to the generator to expand its receptive field, while a spatial attention mechanism was added to the discriminator. The image structural similarity was also integrated into the loss function, boosting the quality of generated SAR images. Experiments on the Massachusetts and Umbra SAR image datasets show that the proposed method effectively augments data for SAR road images.