Three-channel Cascade Network for Underwater Image Enhancement

In underwater scenes, the absorption and scattering of underwater light reduces the quality of underwater images, resulting in blurring and color distortion in underwater images, which affect the accuracy of underwater target detection, tracking, and navigation. In order to address these problems, we have built an underwater three-channel cascade convolutional neural network (UTC-Net) for underwater image enhancement in this paper. Specifically, the UTC-Net consists of a three-channel image restoration network (T-Net) and a cascade optimization network (C-Net). The T-Net recovers the gray images under different color channels respectively using an underwater optical physics imaging model and a dense connection structure, according to the attenuation ratio of light with different wavelengths in underwater images, and the C-Net consists of locally residual block and dense connection, which is used to better recover color and texture details. The UTC-Net has been quantitatively and qualitatively tested by Galdran2015 and RUIE datasets, the experimental results showed that the method achieves satisfactory results in enhancing images quality, such as color correcting, dehazing and feature restoring.