SonoLF-ASPPNet: A Low-Frequency Focused Multiscale Network for Speckle Denoising in Forward-Looking Sonar Images

Two-dimensional forward-looking sonar images are often severely degraded by speckle noise, impairing human interpretation and computer vision tasks. Existing denoising methods, including traditional multiplicative noise models and deep learning approaches, suffer from over-suppression of high-frequency information, leading to residual artifacts, incomplete noise removal, and blurred edges. To address these challenges, this letter proposes SonoLF-ASPPNet, a novel sonar image denoising network, along with a new simulated dataset construction method. The network incorporates three key contributions: the low-frequency structure-edge focusing module combines the traditional SRAD with a Laplacian kernel to enable simultaneous low-frequency denoising and high-frequency edge preservation; the low-frequency attention module incorporates an atrous spatial pyramid pooling module to enhance contextual awareness and suppress high-frequency interference; The training utilizes a composite loss function to enhance denoising in smooth areas, suppress color spots, and homogenize backgrounds. The experimental results indicate that on the simulated dataset, SonoLF-ASPPNet achieved PSNR and SSIM values of 27.76dB and 0.817, respectively, along with excellent visual performance. On the real forward-looking sonar dataset, the model outperformed comparative algorithms in terms of visual outcomes, demonstrating superior noise suppression and texture preservation. Moreover, it attained competitive quantitative scores in ENL, EPI, and M-score.