HDRANet: Hybrid Dilated Residual Attention Network for SAR Image Despeckling

Abstract

In order to remove speckle noise from original synthetic aperture radar (SAR) images effectively and efficiently, this paper proposes a hybrid dilated residual attention network (HDRANet) with residual learning for SAR despeckling. Firstly, HDRANet employs the hybrid dilated convolution (HDC) in lightweight network architecture to enlarge the receptive field and aggregate global information. Then, a simple yet effective attention module, convolutional block attention module (CBAM), is integrated into the proposed model to constitute a residual HDC attention block through skip connection, which further enhances representation power and performance of the model. Extensive experimental results on the synthetic and real SAR images demonstrate the superior performance of HDRANet over the state-of-the-art methods in terms of quantitative metrics and visual quality.

SAR Image Despeckling Using Continuous Attention Module

“…To verify the performance of the proposed SAR-CAM model, experiments conducted on simulated and real SAR data are described. For comparison, three NL methods and six deep learning methods were chosen: the probabilistic patch-based (PPB) [13], the SAR block-matching 3D (SAR-BM3D) [14], the fast adaptive nonlocal SAR (FANS) [15], the dilated residual network-based SAR (SAR-DRN) [30], the hybrid dilated residual attention network (HDRANet) [32], the U-Net [35], the spatial and transform domain convolutional neural network (STD-CNN) [33], the multi-objective network (MONet) [29], and the multiscale residual dense dual attention network (MRDDANet) [34].…”

“…Attention modifies existing convolutional features by multiplying generated attention weights to highlight strong features while restraining non-important features. A couple of works have been designed for SAR image despeckling based on attention blocks [32]- [34], [40], and have achieved superior results compared with non-attention CNNs. Thus, the proposed method also utilizes an attention mechanism to enhance its despeckling capability.…”

“…Another end-to-end method, the SAR-DRN [30] employed a dilated residual network (DRN) to increase the receptive field size; this is well known for aiding image denoising tasks. In addition to the SAR-DRN, a hybrid dilated convolution (HDC) with skip connections and convolutional block attention modules (CBAMs) [31] are implemented in the HDRANet [32], avoiding gridding artifacts that are commonly found in dilated convolutional operation, and improving feature extraction capability. Speaking of attention modules, a spatial-channel attention module (SCAM) was proposed in the STD-CNN [33] to enhance the isolation of the noise-related features from the real image for developing despeckling performance.…”

“…The common feature of these models is that they all depend on data-driven deep learning methods that can be divided into two parts-synthetic and multitemporal approach [37]. Synthetic approach employs a fully developed simulated speckled noisy input from a grayscale converted version of a clean optical reference image and [28]- [30], [32]- [35] filters rely on this method. On the other hand, multitemporal approach produces noise-free reference by averaging a stack of multitemporal SAR acquisition, obtaining regions with no significant temporal differences for reliability.…”

Nonlocal CNN SAR Image Despeckling

“…Along the same line, Gui et al [37] used dilated convolution and residual learning with a densely connected network. In addition, Li et al [38] relied on dilated convolution and residual training, the main innovation being the use of a convolutional block attention module to enhance representation power and performance. All these methods use synthetic training on the UC-Merced [39] dataset.…”