Underwater Image Restoration Via Scene Depth and Minimum Intensity Prior of Red Channel

Underwater images often suffer from hazy and blur due to absorption and scattering effects. To address these problems, we propose a variational framework combining scene depth map and minimum intensity prior of red channel. Depending on the complete underwater image formation model, an adaptive weighted strategy is first designed to estimate the scene depth by fusing the information of brightness, gradient and chromatic aberration. After that, the transmittance of three color channels can be further precisely calculated. Then, in the proposed variational model, the data fidelity term is established according to the forward scattering, and the minimum intensity prior of red channel is regarded as the regular term. Using image pyramid, blur kernel estimation is further carried out by one stepwise optimization strategy from coarse to fine scales. Finally, the alternating direction multiplier method is employed to speed up the non-smooth optimization of the proposed variational model. Qualitative and quantitative experiment is conducted on the UIEB dataset, comparing with several state-of-the-art methods, the proposed method obtains increase by 15% across three quality evaluation metrics in terms of UCIQE, FADE and CPBD, and the restored images have greater clarity and richer edge information.