SonarGAN: A Progressive GAN Framework for Sonar Image Denoising under Multi-Type Noises
Zhangrui Hu, Yunxuan Feng, Binyu Nie, Lei Yan, Wenjie Lu, Liang Hu
AI summary
Problem
Forward-looking sonar images suffer from severe speckle, sidelobe, and structural noise that blur features and degrade downstream tasks, while existing denoising methods either over-smooth details or require impractical paired datasets.
Approach
The framework progressively denoises images in three stages: unpaired preliminary suppression, paired structural noise removal, and constrained joint refinement, preserving target geometry without needing paired training data.
Key results
- Eliminates reliance on costly paired sonar datasets
- Achieves state-of-the-art PSNR and SSIM across synthetic and real sonar types
- Significantly improves completeness and geometric accuracy of downstream 3D reconstructions
- Operates at 20 FPS, meeting real-time underwater robotic requirements
Why it matters
Provides a practical, real-time denoising solution for underwater robots and AUVs operating in turbid environments where optical sensors fail.
Abstract
Forward-looking sonar is essential for underwater perception especially in turbid waters, yet its images are often strongly degraded by various noises, including speckle, sidelobe, and structural noises, which severely hinder downstream tasks such as underwater reconstruction, positioning, and navigation. Most conventional sonar denoising methods reduce the noise at the expense of loss of fine image features or blurred image, while modern supervised learning methods demand large paired datasets that are impractical to obtain in real underwater conditions. In this paper, we propose SonarGAN, a progressive Generative Adversarial Networks (GAN) based framework that denoises sonar images under multi-type noises in one go. Unlike traditional supervised methods, SonarGAN avoids the need for costly paired datasets by combining un- paired real and simulated images, synthetic noisy–clean pairs, and joint refinement for comprehensive denoising. Extensive experiments across multiple types of sonar and underwater environments demonstrate the effectiveness of SonarGAN and its generalization in real-world conditions. We further show that SonarGAN provides high-quality inputs for downstream 3D reconstruction, significantly improving both the completeness and geometric accuracy of the reconstructed models. Our code and dataset are available at https://github.com/ Amarantos12/SonarGAN.