Research On Underwater Image Restoration And Enhancement Algorithms

With the increasingly strategic position of ocean development in China,underwater images which are essential to obtain oceanic information,have attracted increasing attention.Absorption and scattering of light with water,make the quality of the underwater image degradation,such as hazing,color bias and contrast reduction.Based on color-line model and deep learning network,this paper studies underwater image restoration and enhancement to improve the quality of underwater images.The concrete research content is as follows:Firstly,based on atmospheric and underwater image-forming principle,optical imaging models and underwater optical attenuation characteristics,an underwater optical imaging model is constructed with the imaging distance and global background light of RGB channels.The defuzziness equation based on the color line model is constructed with fuzzy evaluation characteristics of the color-line model,color-line generating strategy and screening conditions.Simulation shows the dehazing effect of the model.Secondly,quadtree segmentation and unary gray entropy are applied to estimate the candidate regions of global background light,improving accuracy estimation of global background light.In the study of transmission estimation with color-line model,a color-line generation strategy is proposed based on similar region merging and major pixel collection extraction,saving the time of color-line generation.Meanwhile,an augmented lagrange equation is constructed to estimate the transmission and reduce the influence of the intersection error which is caused by the approximate intersection of color line and global background light vector,improving the applicability of the algorithm in multiple position states of vector and line.Ablation experiments demonstrate that the proposed algorithm can effectively restore low quality underwater images.Finally,a multi-scale conditional generative adversarial network is proposed based on deformable convolution and fuzzy feature index to enhance underwater images in real time.A global-local nesting generator with deformable convolution and a multi-scale discriminator are designed to learn global information and local detailsmore sophisticated.enhancing the learning ability of network,improving the efficiency of feature extraction and the effect of feature learning,The fuzzy evaluation index is proposed to improve the deblurring ability of the network according to the fuzzy evaluation characteristics of the Color-line model.Simulation results show that the proposed algorithm can meet the real-time requirements,generate clear underwater images with rich details,and perform well in color correction.