PID2Net: A neural network for joint underwater polarimetric images descattering and denoising

Underwater optical imaging is widely used in various marine exploration tasks. However, the scattering of light by suspended particles in water will result in poor contrast to underwater images. Underwater polarimetric imaging could effectively suppress the influence of scattered light. However, apart from scattering, the absorption and poor illumination will cause low signal to radio of images, and thus the influence of noise on image quality is also significant. Therefore, underwater images should jointly solve scattering and noise problems. In this paper, we propose a deep learning-based image enhancement method, called polarimetric image descattering and denoising neural network (PID2Net), to suppress the negative impact of both scattering and noise on underwater image quality. In particular, our network is based on frequency decomposition, which aims to remove scattered light at low frequency and suppress noise at high frequency by applying residual dense network. Experimental results show that our method can well enhance underwater images obtained in dense scattering media and severe noise environments, with performance superior to previous methods.