Deep Convolutional Network for Spatially Correlated RAYLEIGH Noise Suppression on TerraSAR-X Images

The task of design of deep convolutional network for denoising of single-look amplitude images of TerraSAR-X spaceborne remote sensing system is considered. The images are distorted by spatially correlated Rayleigh noise which is difficult to remove. Noise suppression on TerraSAR-X images is complicated also by necessity of improved preservation of fine details and textures on the images. A deep convolutional network structure is proposed as well as custom loss function for its training providing for this network a detail preserving optimization. A comparative analysis of the proposed network training with different loss functions and different training test sets is carried out. It is shown that the proposed network provides PSNR for denoised images in average on 1.3 dB better than well known DCT based filter adapted for the noise. It is shown also that usage of the proposed custom loss function allows to provide weighted PSNR value on 0.8 times better than usage of conventional loss function, and on 2 dB better than DCT based filter. Results of processing of real TerraSAR-X images are presented.