LAMSNN: Learnable adaptive modulation for artifact suppression in spiking underwater image enhancement networks

Underwater Image Enhancement (UIE) is critical for marine applications but is severely hampered by complex optical degradation. Spiking Neural Networks (SNNs) present an energy-efficient alternative to conventional Artificial Neural Networks (ANNs) for UIE; however, their efficacy in generative tasks is often compromised by spike-induced artifacts and a discernible performance gap. This paper introduces LAMSNN (Learnable Adaptive Modulation for Spiking Neural Networks), a novel framework meticulously engineered to substantially elevate the quality of SNN-based UIE by effectively suppressing these artifacts. The core innovation of LAMSNN resides in a learnable adaptive modulation strategy that dynamically transforms the input signal, diverging from direct output prediction, which is principally realized through two synergistic contributions: (1) Adaptive Input Modulation (AIM), which leverages network-derived, spatially-variant features (K, B) to apply a learnable affine transformation to the original input, thereby proactively mitigating signal degradation and anticipated artifacts; and (2) the integration of Dynamic Tanh (DyT) within AIM, which introduces a learnable scaling factor (α) to the constraining activation, enabling fine-grained, adaptive control over the modulation strength and range. We instantiate the LAMSNN framework with two bespoke SNN architectures, SpikingResformer and SpikingFilmNet, specifically tailored for generative UIE. Comprehensive experimental results demonstrate that LAMSNN significantly diminishes spike-induced artifacts, achieves image enhancement quality that is competitive with state-of-the-art ANN-based methodologies, and upholds the inherent computational advantages of SNNs. This research underscores the profound potential of adaptive modulation strategies to unlock and augment the capabilities of SNNs for complex generative vision tasks. The code will be made publicly available at https://github.com/Jinxinshao/LAMSNN.