Source-linked AI summary
Learning a Dilated Residual Network for SAR Image Despeckling
Qiang Zhang, Qiangqiang Yuan, Jie Li, Zhen Yang, Xiaoshuang Ma
TL;DR
SAR image despeckling must address multiplicative speckle noise that degrades image quality and limits downstream use. The paper proposes SAR-DRN, a nonlinear end-to-end network using dilated convolutions, skip connections, and residual learning, and reports state-of-the-art quantitative and visual performance, especially for strong speckle noise.
Problem
Multiplicative speckle noise degrades SAR image quality, while traditional linear despeckling models are limited in preserving details and handling nonlinear noise.
Method
SAR-DRN learns an end-to-end mapping from noisy to clean SAR images using dilated convolutions, skip connections, and residual learning without relying on manually adjusted critical parameters.
Results
SAR-DRN shows state-of-the-art quantitative and visual performance in simulated and real SAR despeckling experiments, especially for strong speckle noise.
Takeaways & Limitations
Dilated convolutions enlarge the receptive field in a lightweight structure, while skip connections and residual learning maintain image details and reduce vanishing-gradient problems.
Takeaways & Limitations
The current method requires separate training for each number of looks, and future work targets complex real scenes, polarimetric SAR, and multi-temporal data.
Abstract
from arXiv · showhide
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 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 superior performance over the state-of-the-art methods on both quantitative and visual assessments, especially for strong speckle noise.
1. Introduction
SAR images support important remote-sensing applications but are inherently degraded by multiplicative speckle noise. The paper reviews limitations of existing despeckling methods and introduces SAR-DRN, a lightweight nonlinear end-to-end approach designed to preserve details.
- Motivation: Multiplicative speckle noise severely reduces SAR image quality and the efficiency of interpretation, retrieval, and other applications.SAR is useful in mapping, surveying, and reconnaissance because it operates across broad surface areas and in adverse conditions.
- Existing methods: Spatial linear filters trade homogeneous-region smoothing against edge preservation and can fail when the equivalent number of looks is small.Reported deficiencies include loss of average value, blurring of strong targets, and incomplete removal of noise in dark scenes.
- Existing methods: Wavelet-based, nonlocal-means, and variational methods address speckle reduction through transforms, self-similarity, or optimization with prior assumptions.These approaches extend beyond simple spatial filtering but remain part of the traditional despeckling landscape reviewed by the paper.
- Motivation: Existing methods can fail to preserve sharp features in complicated textures or create block artifacts in speckled images.These limitations motivate a nonlinear mapping between speckled and clean SAR images.
- Proposed approach: SAR-DRN learns a nonlinear end-to-end mapping using dilated convolutions, skip connections, and residual learning.Dilated convolutions enlarge the receptive field while maintaining filter size and layer depth in a lightweight structure; skip connections help maintain details and reduce vanishing gradients.
- Evaluation: The proposed approach is evaluated on simulated and real SAR experiments using quantitative and visual assessments, with especially strong performance under strong speckle noise.The paper presents the experiments in Section 4 and conclusions in Section 5.
2. Related Work
The paper frames SAR despeckling as recovery under a multiplicative noise model and reviews CNN-based processing as a nonlinear alternative. It introduces the convolutional operations, nonlinear activation, and training procedures underlying this approach.
- SAR noise model: SAR speckle degradation is modeled multiplicatively as y = x · n, where y is the speckled image, x the clean image, and n the speckle noise.For SAR amplitude images, the speckle is described as following a Gamma distribution.
- SAR noise model: The equivalent number of looks L is used as a quantitative evaluation index for real SAR despeckling experiments in homogeneous areas.The supplied passages define L as the equivalent number of looks and specify L ≥ 1.
- CNN motivation: Because speckle is nonlinear multiplicative noise, the paper identifies nonlinear expressions as an important strategy for speckle reduction.CNNs are introduced as a way to process low-level and top-level image features without relying solely on traditional filtering assumptions.
- CNN formulation: A CNN layer combines local patches from the previous layer with learned weights and biases, then applies a nonlinear activation to produce output feature maps.The output is indexed by feature-map and pixel locations, with filter size and the number of output maps as layer parameters.
- CNN formulation: The ReLU activation used to introduce nonlinearity is defined as F(O) = max(0, O).The activation is presented as a way to ensure a nonlinear relationship between CNN inputs and outputs.
- CNN training: Back-propagation updates convolution kernels and biases through gradients of the loss function and gradient descent, with α serving as the learning rate.The update proceeds from higher layers toward lower layers using the chain rule.
- CNN-based despeckling: Prior despeckling CNNs include DnCNN and SAR-CNN, with SAR-CNN combining convolutional layers, batch normalization, ReLU, and homomorphic transforms for multiplicative noise.These methods provide context for the paper’s proposed SAR-DRN architecture.
3. Proposed Method
SAR-DRN learns an end-to-end nonlinear mapping for SAR despeckling using dilated convolutions, skip connections, and residual learning. These components enlarge contextual coverage while supporting lightweight modeling, detail preservation, and trainability.
- 3. Proposed Method: SAR-DRN is trained end-to-end with dilated convolutions, skip connections, and residual learning instead of log-transform or manually specified noise models.The network learns complex nonlinear relations directly from training data and corresponding labels.
- 3.1 Dilated Convolutions: Dilated convolutions enlarge the receptive field while maintaining the filter size, enabling broader context without simply increasing kernel dimensions.With four layers, the cited receptive field is 31×31 for dilated convolution versus 9×9 for common convolution at the same depth.
- 3.1 Dilated Convolutions: The seven 3×3 dilated convolution layers use dilation factors 1, 2, 3, 4, 3, 2, and 1 to balance feature extraction ability and training time.The authors describe SAR-DRN as a lightweight model with only seven dilated convolution layers.
- 3.2 Skip Connection: Two skip connections link layers 1–3 and 4–7, passing earlier feature information forward to maintain image details and reduce vanishing-gradient problems.Their effects are compared with a model without skip connections in the discussion.
- 3.3 Residual Learning: Residual learning predicts the residual image, whose pixel values are described as mostly near zero and spatially sparse, making optimization smoother.Residual mapping is also reported to reduce training loss rapidly in multilayer networks.
- 3.3 Residual Learning: The training flow minimizes converged loss between the residual image and the network output to learn the nonlinear relation between speckled and original SAR images.The learned model is then prepared for real speckle SAR image processing.
1) Training and Test Datasets
The study trains SAR-DRN on optically derived, simulated SAR data because clean reference SAR images are difficult to obtain, then evaluates it on simulated and real scenes.
- 1) Training and Test Datasets: Clean reference SAR training images are difficult to obtain, so the UC Merced land-use dataset is used to simulate SAR despeckling data at different look numbers.The dataset contains 21 scene classes with 100 images per class, and optical amplitudes are processed before training for single-polarization SAR data.
- 1) Training and Test Datasets: 400 images sized 256×256 are cropped into 40×40 patches with stride 10, producing 193,664 training patches.Training uses batch size 128, with look numbers L = 1, 2, 4, and 8 for noise simulation.
- 1) Training and Test Datasets: Testing uses simulated Airplanes, Buildings, and Rivers examples plus real Flevoland, Deathvalley, and San Francisco SAR images.The real images are cropped to 500×600, 600×600, and 400×400, respectively.
3) Compared Algorithms and Quantitative Evaluations
SAR-DRN is evaluated against four mainstream despeckling methods on simulated and real SAR images. It achieves the strongest reported quantitative and visual performance, while dilated convolutions and skip connections improve training and runtime remains low.
- Compared Algorithms and Quantitative Evaluations: SAR-DRN is compared with PPB, SAR-BM3D, SAR-POTDF, and SAR-CNN using PSNR and SSIM across four simulated speckle levels.The simulated experiments use L=1, 2, 4, and 8 on Airplane, Building, and Highway images.
- Compared Algorithms and Quantitative Evaluations: SAR-DRN obtains all best PSNR results and nine of twelve best SSIM results across the four simulated noise levels.At L=1, it exceeds SAR-BM3D by about 1.1 dB, 0.6 dB, and 0.6 dB for Airplane, Building, and Highway, respectively.
- Compared Algorithms and Quantitative Evaluations: SAR-DRN shows superior quantitative and visual performance over state-of-the-art methods, especially under strong speckle noise.The comparison includes both simulated and real SAR images.
- Compared Algorithms and Quantitative Evaluations: On real AIRSAR images, SAR-DRN combines speckle reduction with preservation of edge and texture information better than the compared methods.Other methods show residual speckle, texture distortions, weaker edge preservation, or over-smoothing in different cases.
- Compared Algorithms and Quantitative Evaluations: ENL evaluation indicates that SAR-DRN has better speckle-reduction ability than the other methods on selected homogeneous regions.ENL was used for speckle reduction, while EPD-ROA was used for edge-preserving ability.
- Compared Algorithms and Quantitative Evaluations: Dilated convolutions reduce training loss and improve PSNR, while skip connections accelerate convergence and enhance model stability.Their combination improves PSNR by about 1.1 dB over the model without either component.
- Compared Algorithms and Quantitative Evaluations: SAR-DRN uses no batch-normalization layer and has the lowest reported runtime complexity among the compared algorithms.The lightweight network has seven layers, compared with 17 layers in SAR-CNN.
5. Conclusion
The paper concludes that SAR-DRN learns an end-to-end mapping for SAR despeckling and achieves state-of-the-art performance on simulated and real data, especially with strong speckle noise. Future work targets complex real scenes, integrated handling of different looks, polarimetric SAR, and richer training data.
- 5. Conclusion: SAR-DRN learns an end-to-end mapping between noisy and clean SAR images using dilated convolutions and skip connections.The design enlarges the receptive field while maintaining a lightweight structure and supports image-detail preservation.
- 5. Conclusion: SAR-DRN achieves state-of-the-art performance in simulated and real SAR despeckling experiments, especially for strong speckle noise.This conclusion is reported relative to traditional despeckling methods.
- 5. Conclusion: Future work will investigate more powerful models for complex real SAR scenes and an integrated model across different numbers of looks.The current method is trained separately for each number of looks.
- 5. Conclusion: The approach is planned for extension to polarimetric SAR, whose noise model is more complicated than single-polarization SAR.The authors also identify prior constraints and multi-temporal data as future directions for complex real SAR despeckling.