Reference-Free Enhancement of Forward-Looking Sonar Images: Bridging Cross-Modal Degradation Gaps Through Deformable Wavelet Scattering Transform and Multi-Frame Fusion

Enhancing forward-looking sonar (FLS) imagery is crucial for underwater target detection. Existing deep learning approaches typically rely on supervised training with simulated data, but their generalization is constrained by the scarcity of real-world paired datasets. Recent reference-free methods can denoise single frames, yet their reliance on local statistics makes them fragile at low signal-to-noise ratios (SNRs) and prone to losing global structural consistency. Multispectral/hyperspectral remote-sensing (RS) and FLS imagery share speckle-like textures and range-dependent contrast, so large-scalesupervisedRS priors are a natural candidate for transfer; however, the cross-modal degradation gap still causes naïve feature transfer to over-smooth targets and compress brightness. To exploit these priors safely, we follow a cross-modalalign-before-usestrategy based on a Deformable Wavelet Scattering Transform (DWST) Feature Bridge. The bridge learns small, regularized scale-orientation offsets and maps sonar images into a stable, translation-invariant feature space aligned with supervised high-/low-resolution RS models. On top of this space, a HighRes-Net-style multi-frame fusion backbone is fine-tuned end-to-end in a fully reference-free manner on noisy sonar sequences, leveraging inter-frame complementarity for speckle suppression and contrast/resolution gains while preserving key sonar imaging characteristics. Experiments on an in-house C900-II pool dataset, the public DEBRIS forward-looking sonar benchmark, and the UATD underwater acoustic target detection dataset show that the proposed DWST-augmented fusion framework substantially outperforms existing reference-free methods in noise suppression, edge preservation, and brightness enhancement, and consistently improves downstream object detection, underscoring its practical effectiveness for underwater applications.