SSS-DPM: Diffusion Probabilistic Model for Speckle Noise Reduction in Side-Scan Sonar Images

The side-scan sonar (SSS) mounted on autonomous underwater vehicles (AUVs) enables efficient and precise seabed mapping as well as underwater exploration. However, SSS images are often impaired by seabed reverberation, environmental interference, and self-noise, with speckle noise from sub-resolution scatterers being a particularly significant issue. This signal-dependent multiplicative noise not only reduces image quality but also complicates feature extraction. To address these challenges, we introduce a diffusion probabilistic model (SSS-DPM) specifically designed to reduce speckle noise in SSS images. SSS-DPM employs a dual-branch forward diffusion process to accurately characterize the noise distribution in these images. During the reverse diffusion phase, it leverages a noise predictor based on Kullback-Leibler (KL) divergence to effectively suppress noise. Moreover, the model incorporates a multi-scale feature extraction network, allowing it to tackle complex noise patterns while preserving critical structural details. Experimental results demonstrate that SSS-DPM consistently surpasses state-of-the-art methods in peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and visual information fidelity (VIF) across diverse real-world datasets. Even in cases of severe degradation, SSS-DPM exhibits exceptional robustness and reliability.