Research On Underwater Image Restoration

Underwater imaging technology has been widely used in many fields,such as marine life research,underwater target detection.In underwater imaging,due to the serious absorption and scattering of light by sea water,the acquired underwater optical images are reduced in sharpness and show obvious color deviation.It not only seriously affects the image quality of underwater image,but also brings adverse effects to the subsequent detection and recognition work.Therefore,how to remove the interference of seawater effectively is a very valuable and challenging work.Aiming at the shortcomings of the existing underwater image processing methods based on deep learning in training data and network structure,two underwater image restoration method based on convolution neural network are proposed in this paper.By improving the accuracy of estimating unknown parameters of underwater images,the problems of low contrast and obvious color deviation of underwater images can be better solved.The contributions of this thesis are summarized as follows:(1)In order to solve the problem of lacking in more diversified data sets at present,we first propose an underwater image synthesis method based on underwater imaging model.With the help of the existing depth map data sets,the synthesized underwater images are calculated by underwater imaging model under the prior that the attenuation of red channels is more serious than that of blue and green channels.This method can obtain underwater synthetic images with different blue-green deviation,different blurring degree and different scenes.It can also be used to synthesize underwater images under artificial light source.The method realizes the diversity of the data set,and ensure the correctness of the data.(2)In order to solve the problem that the estimation of transmission and ambient light of existing underwater image restoration methods are not accurate enough,this thesis proposes a method of underwater image restoration based on deep learning.A parallel parameter estimation network for underwater image is designed,which includes the branch of underwater image transmission estimation and the branch of underwater image ambient light estimation.In particular,the structure of cross-layer connection and multi-scale estimation is designed in the underwater image transmission estimation network,whichcan effectively preserve edge features and prevent halo artifacts in the restoration results.The transmission and ambient light of the underwater image are estimated by the convolution neural network,and then the restored image is obtained by inversion of the underwater imaging model.It can be proved by comparative experiments that the proposed method can estimate ambient light and transmission more accurate.And it can enhance the contrast of underwater degraded images and correct its color deviation,the results of restoration is more natural.(3)Considering the fact that artificial light sources are often used in underwater photography to compensate the brightness of the scene,in order to eliminate the influence of compensation light on image restoration,an underwater image restoration method for both natural and artificial light sources is proposed in this paper.An underwater imaging model and data synthesis method under artificial light environment are constructed,and the intensity estimation network of underwater image artificial light is designed.In addition,the accuracy of ambient light estimation is improved by adding transmission map to the input of the ambient light estimation network.The method can further improve the application scope and processing effect of the method,and can adjust the image brightness while enhancing the image contrast and correcting the color cast of the image.