Deep Learning Underwater Image Color Correction and Contrast Enhancement Based on Hue Preservation

Underwater Image suffers from serious color distortion and low contrast problems because of complex light propagation in the ocean. In view of computing constraints of underwater vehicles, we propose a high-efficiency deep-learning based framework based on hue preservation. The framework contains three convolutional neural networks for underwater image color restoration. At first, we use the first CNN to convert the input underwater image into the grayscale image. Next, we enhanced the grayscale underwater image by the second CNN. And then, we perform the color correction to the input underwater image by the third CNN. At last, we can obtain the color-corrected image by integrating the outputs of three CNNs based on the hue preservation. In our framework, that CNNs specialize on each work can be able to simplify each architecture of CNNs at most and improve the regression quality to achieve the low computing cost and high effeciency. However, the problem of the underwater CNNs is that the underwater training data is too few and without the corresponding ground truth. Thus, we use the unsupervised learning method CycleGAN to train the underwater CNNs. We design a training method as the combination of three CycleGANs that can train the three CNNs at the same time to share the regression status. This training method may let the three CNNs of our proposed framework support each other to avoid the training overfitting and without constraint. By the proposed framework and training method, our method can process the underwater images with high quality and low computing cost. The experimental results have demonstrated the correct colors and high image quality of the proposed method's results, compared with other related approaches.