Dual-path attention and transformer-enhanced disentanglement network for underwater image restoration

Abstract

Due to the complex underwater environment, underwater imaging is affected by the absorption and scattering of atmospheric light on water, resulting in problems such as the main subject not being prominent, the background being blurry, the outline being unclear, and the overall image presenting a bluish-green tone. Currently, many popular underwater image enhancement methods can only restore some details, with poor overall effects. Therefore, this paper proposes a DAT-MDNet network for underwater image restoration. It can decompose the degraded underwater image into the object's own light, medium transmission effect, and global background light by simulating the physical imaging model. The network also has a dual-path attention mechanism: one path focuses on texture optimization and local detail processing, and the other path handles the overall information relationship. A lightweight Transformer module first extracts global information and then fuses local details to achieve the integration of information at different scales. During the training process, we use a composite loss function to ensure that the decomposed images look natural and clear. Compared with other popular algorithms, the method proposed in this paper achieves more accurate colors, higher contrast, and clearer details in terms of visual perception.