Underwater Joint Depth Estimation And Color Correction Based On Deep Learning

As a basic task of underwater computer vision,accurate depth estimation can improve the level of underwater vehicle ranging,positioning,target capture and other tasks.High quality color correction is of great significance to the development of underwater target detection and recognition.Underwater depth estimation is to estimate the distance of each pixel in the underwater RGB image relative to the shooting source.Color correction solves the problem of light attenuation and backscattering caused by water environment in image.However,compared with the advantages of large amount of data and mature technology research and development in the onshore environment,uneven illumination and low visibility lead to problems such as image blurriness and many impurities in underwater imaging,which bring a lot of uncertainties for accurate annotation of underwater images and efficient and highquality image processing.Therefore,it is necessary to first synthesize a large number of underwater data with the help of a reliable image conversion model,so as to alleviate the problem of lack of supervision information in real underwater data.And,the depth map is closely related to the formation of ambient light attenuation and backscattering in the water,underwater image encoded in a lot of distance information,and the color of the underwater image distortion is associated with distance in the scene,so jointing the underwater depth estimation tasks and color correction task,implementing multi-task learning,help to improve the results.Aiming at the above problems and solutions,this paper designs an end-to-end unsupervised network structure.The network synthesizes a large number of effective underwater image datasets through style transformation without a pair of underwater image training sets.After that,a large number of generated underwater images were used to conduct multi-task learning of depth estimation and color correction at the same time,so as to promote the performance of each other with the help of each other's feature information.Finally,the feature domain adaptation strategy is adopted to narrow the domain difference between the synthesized underwater data and the real underwater data,and then improve the adaptability of the network to the real underwater scene.The main contributions of this paper are as follows:(1)Multi-domain underwater image style transformation framework based on generative adversarial network.According to different underwater scenes,the framework can use the effective domain label information to transform the style of a large number of land datasets and synthesize underwater image datasets.(2)A multi-task learning framework for depth estimation and color correction of underwater images.Both of the two subtasks adopt encoder-decoder structure,and in the learning process of the two subtasks,the feature graphs of each layer of the decoder are interacted,so that the two subtasks can learn more useful information and achieve mutual benefit.(3)Feature domain adaptation strategy of teacher-student structure.Both the teacher network and the student network are complete task network structure.The student network is responsible for modeling the input-output relationship in the synthesize image domain,while the teacher network guides the student network to migrate to the real underwater image domain,so as to improve the adaptability to the real underwater scene.In order to verify the effectiveness of the proposed joint depth estimation and color correction network for underwater images,this paper carried out training and testing on multiple public data sets,and compared with the previous methods on objective and subjective indicators such as RMSE,PSNR and SSIM,all of which achieved excellent performance.