MFSonar: Multiscale Frequency-Domain Contextual Denoising for Forward-Looking Sonar Image Semantic Segmentation

Semantic segmentation of forward-looking sonar (FLS) images is crucial for enhancing situational awareness in marine environments. However, FLS images are often degraded by environmental noise, similarity noise, and shading effects, which result in low resolution, poor signal-to-noise ratio, and suboptimal image quality. These issues significantly hinder the accuracy of semantic segmentation in FLS images. To address these challenges, we propose a novel method called MFSonar, which is based on the Transformer-Mamba architecture. MFSonar incorporates a context channel denoising module (CCDM) that exploits the similarity characteristics of local and global features to effectively suppress similarity noise and enhance target features. Additionally, the Multiscale Frequency-Domain Decoding Module integrates multiscale frequency-domain convolution with visual state-space (VSS) blocks, leveraging frequency-domain characteristics to mitigate environmental noise and occlusion shadows. Furthermore, our approach prioritizes local features before global features to achieve effective fusion and enhancement of global semantic features and multiscale local visual information. Extensive comparative experiments across multiple datasets demonstrate that MFSonar achieves state-of-the-art performance. Moreover, ablation studies and visual comparisons on a primary dataset validate the superiority, effectiveness, and uniqueness of our approach. Our implementation is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/NWPUFranklee/PVSonar.git</uri>.