A Practical Approach for SAR Image Despeckling Using Deep Learning

Khilan Ravani, Shivam Saboo, Jignesh S. Bhatt

Abstract: In this paper, we propose a deep learning based practical approach to despeckle synthetic aperture radar (SAR) images without using ground truth (despeckled images). Conventionally, the learning based approaches need the despeckled images for the training. However, in real scenarios, it is often difficult to acquire the ground truth. To this end, we demonstrate a technique to restore the speckled images using the available SAR data only. We modify the UNet architecture in order to train it with the given speckled images. We 1) introduce an extra residual connection in each Conv-Block of UNet, and 2) replace the transpose convolution with parameter free bi-linear interpolation. We show the training of this modified network in the mean-squared error (MSE) sense. Experiments have been conducted on ImageNet dataset as well as on the real SAR images. The results are validated both qualitatively as well as quantitatively and compared with state-of-the-art-approaches.

External IDs:dblp:conf/igarss/RavaniSB19