Self-calibrated dilated convolutional neural networks for SAR image despeckling

Abstract

ABSTRACT Synthetic aperture radar (SAR) imaging process is essentially disturbed by speckle noise due to its unique mechanism. Speckle noise causes severe degradation of SAR image quality, which significantly limits its practical application. Recently, convolutional neural networks (CNNs) have indicated good potential for various image processing tasks. In this article, we propose a self-calibrated dilated convolutional neural network for SAR image despeckling called SAR-SCDCNN. The main body of SAR-SCDCNN is formed by seven self-calibrated blocks (SeCaBlock). Firstly, in each SeCaBlock, the input features are split into two branches: one represents the contextual features in the original space, and another represents those in the long-range space. Then, the down-up sampling operation and the convolutions with hybrid dilated rates are employed to enlarge the receptive field. The weights for calculating features themselves are adaptively extracted in the second branch. And then, two branch features are concatenated and fused through a convolution. Finally, a skip connection is used between the input and output of each SeCaBlock to give full play to the expressive power of the deep neural network and enhance training stability. Experiments on synthetic speckled and real SAR images are conducted to perform objective quantitative and subjective visual evaluations of image quality. Results show that our proposed method can effectively suppress speckle noise and adequately preserve detailed features.