MixRformer: Dual-Branch Network for Underwater Image Enhancement in Wavelet Domain

This paper proposes an underwater image enhancement model MixRformer that combines the wavelet transform and a hybrid architecture. To address the problems of insufficient global modeling in existing CNN models, weak local feature extraction of Transformer and high computational complexity, multi-resolution feature decomposition is performed through a discrete wavelet transform (IWT/DWT) in which low-frequency components retain structure and texture, and high-frequency components capture detail features. An innovative dual-branch feature capture module (DFCB) is designed as follows: (1) the surface information extraction block combines convolution and position encoding to enhance local modeling; (2) the rectangular window gated Transformer expands the receptive field through the convolution gating mechanism to achieve efficient global relationship modeling. Experiments show that the model outperforms mainstream methods in color restoration and detail enhancement, while optimizing computational efficiency.