Underwater image enhancement based on dual U-net

Abstract

As land resources have continually decreased, ocean exploration by humans has steadily grown. Underwater imaging is one of the most intuitive means to reflect the internal conditions of the ocean. However, due to the complex imaging environment of the ocean and light scattering in the sea, underwater images exhibit severe degradation, making it difficult to distinguish effective information. Thus, underwater imaging must be enhanced. Compared with traditional methods (e.g. histogram equalization method) and modeling methods, deep learning has been well applied in the field of computer vision. The key points are the acquisition of the training set and the generalization ability of the convolution model. Because the model-based method often needs to measure prior data manually in advance, it will cause inevitable errors; and direct generalization of the neural network will also cause image blurring. In this paper, we design a double U-Net for underwater image enhancement with strong generalization ability, in combination with modeling and deep learning methods. The gray image of the input image is processed with the attention mechanism in advance, and the relevant transmittance information is obtained using a U-Net. Then, the input image is processed with the information output from each layer of the previous U-Net. The final result is obtained by dividing the two U-Net results by pixels. The proposed network is trained using a paired training set generated by CycleGAN. Through quantitative and qualitative analysis, our method is proved to be more effective than the methods in recent papers in the field of underwater image enhancement.