Learning a Dilated Residual Network for SAR Image Despeckling

Abstract

Abstract:In this paper, to break the limit of the traditional linear models for synthetic aperture radar (SAR) image despeckling, we propose a novel deep learning approach by learning a non-linear end-to-end mapping between the noisy and clean SAR images with a dilated residual network (SAR-DRN). SAR-DRN is based on dilated convolutions, which can both enlarge the receptive field and maintain the filter size and layer depth with a lightweight structure. In addition, skip connections and a residual learning strategy are added to the despeckling model to maintain the image details and reduce the vanishing gradient problem. Compared with the traditional despeckling methods, the proposed method shows a superior performance over the state-of-the-art methods in both quantitative and visual assessments, especially for strong speckle noise.

DoPAMINE: Double-Sided Masked CNN for Pixel Adaptive Multiplicative Noise Despeckling

“…x + µ 2 x + c 1 )(σ 2 x + σ 2 x + c 2 ) and is designed to cover the weakness of PSNR, which is known to be sensitive to shift or brightness of the images Results Table 1 is the PSNR and SSIM results on the UCML test dataset. SAR-DRN (Zhang et al 2018) is a stateof-the-art baseline model in despeckling that predicts residual Z − x before doing despeckling. It is constructed by 7 dilated convolution layers and skip connections.…”

“…In addition to the experiments on the benchmark dataset, we also despeckled real-world SAR-obtained images with 4 number of looks: Death Valley, Flevoland, and San Francisco Bay. These are also widely used benchmark images for SAR despeckling (Zhang et al 2018). We only report the result for Flevoland image with 512 × 512 resolution due to space limit.…”

“…Namely, by casting the despeckling as a supervised learning problem, the CNN models are trained to learn a mapping from the noisy patch to the clean patch. For example, (Wang, Zhang, and Patel 2017) proposed ID-CNN that uses exactly the same architecture as DnCNN on the Gammadistributed multiplicative noise and introduced a total variation (TV) regularization; (Chierchia et al 2017) transformed both clean and noisy images to the log domain then carried out the same process as in the additive noise case; (Zhang et al 2018) proposed SAR-DRN, which added dilated convolution and skip connections to the ID-DCN architecture. As in the additive noise denoising case, the CNN-based models have achieved superior despeckling performances than the classical approaches.…”

Deep Multi-Scale Recurrent Network for Synthetic Aperture Radar Images Despeckling

“…Bai et al [26] added fractional total variational loss to the loss function to remove the obvious noise while maintaining the texture details. The authors of [27] proposed a CNN framework based on dilated convolutions called SAR-DRN. This network amplified the receptive field by dilated convolutions and further improved the network by exploiting the skip connections and a residual learning strategy.…”

“…In this study, we design an end-to-end multi-scale recurrent network for SAR image despeckling. Unlike [9,[25][26][27], which only utilized CNN to acquire speckle distribution characteristics and additional division operation or subtraction operation to remove speckle, we use the network to learn the distribution characteristics of speckle noise, meanwhile automatically implementing speckle suppression to output clean images. The proposed network is based on the encoder-decoder architecture.…”