Polarization-based underwater image enhancement using the neural network of Mueller matrix images
Haoyuan Cheng, Jinkui Chu, Yongtai Chen, Jianying Liu, Wenzhe Gong
Abstract
In the paper, a method of underwater image enhancement based on the neural network of Mueller matrix images is presented. The novelty of our study lies in the employment of the neural network based on Mueller matrix imaging, which can provide complete information of the target, to conduct underwater image enhancement. To establish the dataset, we obtain the Mueller matrix images of different objects under different water turbidity. We utilize an improved neural network based on U-net and a loss function using the high-level feature extractor to enhance the underwater images. The method does not require the physical model of the underwater scene because of the use of deep learning. The enhancement of objects with different materials and textures is obvious, which illustrates that the proposed method is effective and robust. Our method is superior compared with the state-of-the-art methods in terms of quantitative measures and visual quality.