Abstract
The existence of speckles in synthetic aperture radar (SAR) images affects its subsequent application in computer vision tasks, so the research of speckle suppression plays a very important role. Convolutional neural networks based speckle suppression algorithms cannot reach a good balance between despeckling effect and structure detail preservation. Considering these issues, a multiscale feature adaptive enhance network for suppressing speckle is proposed. Specifically, an encoder–decoder architecture embedded with multiscale operations is constructed to capture rich contextual information and remove speckles from coarse to fine. Then, deformable convolution is introduced to flexibly adapt changes in ground objects’ complex and diverse image features. Also, the constructed feature adaptive mixup module mitigates shallow feature degradation in deep networks by establishing connections between shallow image texture features and deep image semantic features with learnable weights. Experiments on synthetic and real SAR images show that the proposed method produces advanced results regarding visual quality and objective metrics.
Introduction
Synthetic Aperture Radar (SAR) is a powerful tool for remote sensing due to its ability to penetrate clouds and darkness. However, SAR images are inherently degraded by speckle, a multiplicative noise that appears as dense graininess. This noise significantly hinders downstream computer vision tasks such as ship detection or land-use mapping.
The primary challenge in despeckling is balancing noise suppression with the preservation of sharp structural details. Traditional filters often over-smooth edges, while many modern Convolutional Neural Networks (CNNs) struggle to adapt to the complex, irregular shapes of ground objects. MFAENet addresses these challenges by employing a multiscale architecture with adaptive feature enhancement to balance noise suppression and structural detail preservation.
Methodology
MFAENet utilizes an encoder-decoder backbone integrated with three specialized components designed to capture contextual information and adapt to complex image features.
- Multiscale Residual Attention Block (MRAB): This block splits feature channels and applies varying kernel sizes to capture context at multiple scales simultaneously. This ensures that high-resolution details are not lost during the downsampling process.
- Adaptive Enhancement Block (AEB): By incorporating deformable convolutions, the network can "warp" its sampling grid to match the irregular shapes of real-world objects like buildings and roads, rather than relying on rigid square grids.
- Feature Adaptive Mixup (FAM): This module uses learnable weights to adaptively blend shallow texture features from the encoder with deep semantic features in the decoder. This weighted fusion prevents the "ghosting" or blurring effects common in standard skip connections.
Experimental Results
The model was evaluated using a combination of simulated data (UC Merced Land Use and Virtual SAR datasets) and real-world SAR images, including Sentinel-1, Terra SAR-X, and Lynx airborne radar data.
Quantitative Performance
On simulated datasets, MFAENet achieved a Peak Signal-to-Noise Ratio (PSNR) of 23.95 dB, outperforming the classic SAR-BM3D algorithm by approximately 1 dB and significantly beating other deep learning benchmarks like FFDNet. Furthermore, the model maintained high edge preservation, with EPD-ROA metrics consistently exceeding 0.99, indicating that noise is removed without dulling sharp boundaries.
Computational Efficiency
A significant advantage of MFAENet is its inference speed. In comparative tests, MFAENet processed images in approximately 0.31 seconds, whereas the high-performing SAR-BM3D required 93.7 seconds. This efficiency makes the network suitable for near-real-time satellite data processing.
Conclusion
MFAENet provides a robust solution for SAR image despeckling by leveraging multiscale features and deformable convolutions. While the current architecture shows state-of-the-art performance, future work will focus on making the Feature Adaptive Mixup module more flexible for arbitrary image sizes and incorporating polarization-aware logic to further improve physical consistency in radar processing.
Summary updated