Underwater Video Enhancement And Denoising

Underwater videos suffer from many problems such as color distortion,low contrast,loss of details due to the scattering of suspended particles and the absorption of light by water.Videos taken in the deep ocean are vulnerable to snow-like noise due to a large amount of suspended particles.Therefore,existing underwater video enhancement methods need to be improved in detail preservation,color correction and noise removal.Clear underwater video has an urgent application demand in many fields such as underwater cable and optical cable construction,underwater robot and marine biology.This thesis mainly focuses on the underwater video enhancement,and the main contributions are as follows:(1)We propose a new underwater video enhancement method based on structuretexture decomposition and optical flow estimation.This method consists of three modules: key frame selection,key frame enhancement and non-key frame enhancement.For key-frame selection,the video can be divided into key frames and non-key frames according to the change of color distribution between frames.For key-frame enhancement,the proposed method firstly performs color correction on the key frames to recover natural tone from greenish or blueish tone.Then quad-tree search strategy is introduced to obtain a more accurate background light by removing highlighted areas.After that,an image decomposition optimization model is established to decompose the key frames into low-frequency structure layers and high-frequency texture layers for enhancement and denoising respectively.For non-key frame enhancement,a transmission estimation method based on optical flow estimation is proposed to improve the temporal consistency of the enhanced video and reduce computational complexity.Experiments show that the proposed method achieves better enhancing results.(2)We propose a new snow-like noise removal method based on spatial-temporal low rank representation.The proposed method first characterizes the time-domain characteristics of snow-like noise and trains a random forest classifier to screen noise areas.Then we extract the image block centered on the highest priority pixel in noise areas according to the priority function,which consists of confidence term and data term.This image block and its similar blocks in adjacent frames form a matrix where the pixels to be repaired are missing elements.Finally,we restore pixels to be repaired by matrix filling method based on spatiotemporal low rank representation.Experiments show that the proposed method not only removes snow-like noise but also retains the details of video.