UIE-SFIFormer: Underwater Image Enhancement Based on Physical-Guided Spatial–Frequency Interaction Transformer
Yuan Zhou, Haiyong Xu, Gangyi Jiang, Mei Yu, Yeyao Chen, Ting Luo
Abstract
Light scattering and absorption can cause color distortion, blurring, noise, and other issues in underwater images, negatively impacting their quality and posing significant challenges for underwater research and exploration. To deal with the problem, a novel underwater image enhancement method, the UIE-SFIFormer, has been proposed by designing the physical-guided spatial–frequency interaction Transformer. Specifically, the proposed physical guidance fusion module (PGFM) is designed to fuse the dark channel inverse transmission map, incorporating prior knowledge, such as brightness and depth, with the raw image to enhance missing physical information. Subsequently, the spatial–frequency feature extraction module (SFFEM) is utilized for further feature extraction of the fused image. Within SFFEM, Transformer is employed for spatial and frequency domain feature extraction to address nonlocal degradation and excessive fuzzy noise in underwater images. Building upon this foundation, a spatial–frequency interaction block is constructed to combine dual features through spatial–frequency-domain hybrid cross-attention. Finally, experimental results on five underwater test data sets demonstrated that the proposed UIE-SFIFormer has a better performance in restoring and enhancing underwater images than other methods.