FMFNet: Frequency-Aware multi-scale attention with fine-grained feature fusion for sonar image semantic segmentation

Sonar image semantic segmentation is a challenging task due to low contrast, severe speckle noise, and blurred object boundaries, which constrain the representation capability of purely spatial-domain convolutional features. This paper proposes FMFNet (Frequency-Aware Multi-Scale Attention and Fine-Grained Feature Fusion Network) for robust sonar image perception. The network integrates frequency-domain prior knowledge, multi-scale contextual modeling, and detail-preserving feature fusion to address the inherent limitations of sonar images. Specifically, the frequency-domain feature extraction module maps intermediate spatial features to a real-valued 2-D DCT transform domain, employs learnable band-wise reweighting to suppress noisy interference and preserve structural feature components, and adaptively embeds the optimized frequency features into the spatial feature stream. The embedded multi-scale attention mechanism aggregates cross-resolution contextual information from diverse receptive fields and generates joint scale-channel-spatial attention maps to highlight effective sonar echo features. Furthermore, the fine-grained feature fusion module aligns and adaptively fuses high-resolution detailed features with low-level semantic features. Experiments on the evaluated SAS dataset show that FMFNet achieves higher aggregate mIoU, OA, and mF1 than representative convolutional and transformer-based segmentation baselines under the reported single-dataset split. The current evaluation is therefore interpreted within this SAS setting rather than as evidence of cross-dataset generalization.