Underwater Image Enhancement Based on Improved U-Net Convolutional Neural Network

Aiming at the problems of low contrast, dim and color distortion of underwater images caused by the attenuation and scattering of light in water propagation, an underwater image enhancement algorithm, which is based on a U-Net convolutional Neural Network is proposed. The algorithm consists of an image synthesis module and an image enhancement module. Firstly, a style transfer network based on transfer learning is built and employed to synthesize underwater images from clear images, image pairs are composed of these synthesized and original images for data expansion. Secondly, an underwater image enhancement network with a U-shaped convolutional variational autoencoder is constructed, and the image pairs serve as input to the second module for training. The qualitative and quantitative analysis results show that the proposed algorithm has good performs favorably in the color fidelity and detailed reservation of the target object.