HAMamba-UIE:融合频域-空间域建模的混合架构水下图像增强OA

Underwater imaging often suffers from issues such as detail blurring,color distortion,and reduced contrast due to the absorption and scattering of light by water,severely limiting the effectiveness of visual perception and high-level vision tasks.To address these challenges,this paper proposes HAMamba-UIE,a hybrid architecture for underwater image enhancement that integrates frequency-domain and spatial-domain modeling.Based on a revised underwater image forma-tion model,the framework decomposes and reconstructs a degraded image into four physical parameters in a feedforward manner:scene radiance(estimated by the core network J-Net),direct transmission map(estimated by TD-Net),backward transmission map(estimated by TB-Net),and global background light(estimated by GBL module),enabling more accu-rate modeling of the underwater light propagation process.As the enhancement backbone,J-Net incorporates a Mamba(selective state space model)-based multi-branch fusion module(MFFM)to establish long-range dependencies,thereby improving the modeling of global degradation characteristics in underwater scenes.The frequency-spatial domain module(FSDM)employs multi-level Haar wavelet decomposition to exponentially expand the receptive field,enhancing the perception and preservation of low-frequency information for superior global color correction.The Mamba convolutional fusion module(MCFM)synergizes the global modeling capability of selective scan for 2D data(SS2D)with the local feature extraction of convolutions,ultimately outputting the high-quality enhanced image J(x)from J-Net.Extensive experiments on multiple public datasets demonstrate that the proposed method outperforms or is comparable to state-of-the-art approaches in terms of both subjective visual quality and objective metrics,including peak signal-to-noise ratio(PSNR),structural similarity index(SSIM),underwater image quality measure(UIQM),and underwater color image quality evaluation(UCIQE).On the LSUI dataset,the proposed method outperforms all mainstream existing methods in both PSNR and SSIM metrics.Compared with the state-of-the-art method LiteEnhance,it achieves a performance gain of approximately 22.67%in PSNR and 12.38%in SSIM.Comprehensive ablation experiment results further validate the effectiveness of each component,confirming the strong generalization capability of the model.