Generative adversarial network based frequency domain enhancement and color compensation underwater image enhancement

In complex underwater environments, due to the large number of suspended particles as well as the varying scattering and absorption characteristics of light in different waters, underwater images are subject to diverse forms of mixed attenuation. Such as color bias, poor contrast, and degradation of details. This greatly limits the operational efficiency of underwater systems. To this purpose, we propose a new generative adversarial network based frequency domain enhancement and color compensation underwater image enhancement method, which performs image enhancement simultaneously in both the frequency and spatial domains. Specifically, we designed a dual-encoder architecture with a structural encoder and a color compensation encoder in the generator. We embed a Multi-scale Dense Feature Aggregation (MDFA) module in the dual encoder, to make different encoders extract rich semantic and contextual information according to different task requirements. In the decoder, we designed a based Frequency-domain Fourier Enhancement Module (FFEM) and a Complementary-color Prior Color-compensation Module (CPCM). The FFEM conducts color correction and detail enhancement of the features captured by structural encoder within the frequency domain. In the spatial domain, the CPCM utilizes the color compensation information extracted by the color compensation encoder to adjust the enhancement results of the FFEM. Abundant experiments indicate that the suggested method significantly improves the degraded image quality, exhibits superior generalization performance, and outperforms the state-of-the-art methods in both quantitative and qualitative evaluations. Our code is available at https://github.com/LiJiaxin011/FCC-GAN.