Research On Underwater Optical Image Enhancement Method Based On Physical Model And Deep Learning

Underwater optical imaging is one of the important ways for humans to acquire marine information.In the complex underwater imaging environments,light absorption,refraction,and scattering effects caused by a large number of dissolved matter and particles often greatly reduce the imaging quality,resulting in low contrast,color distortion,blurred details,etc.,which increases the difficulty of computer vision tasks.Therefore,how to effectively improve the optical image quality in complex water environments is the core problem faced by underwater optical imaging technology.The research content of this thesis focuses on the key issues in the field of underwater optical image enhancement.This thesis is aiming at the limitation of the underwater imaging physical model,the ill-posed problem of scene prior assumptions,the ill-conditioned estimation of model parameters,and the incomplete removal of backscattered light.Based on the mechanism and characteristics of the underwater optical imaging system,an in-depth study on underwater optical image enhancement methods is carried out based on physical model and deep learning.(1)Aiming at the problem that the traditional underwater imaging model does not fully describe the underwater image degradation process,this thesis establishes a three-dimensional coordinate system of underwater optical imaging,and constructs optical imaging models for natural illumination and complex illumination by analyzing the multiple light attenuation stages related to imaging under different light conditions.It based on the Jaffe-Mc Glamery underwater optical imaging model and the underwater radiation transfer equation.Combined with the absorption and scattering properties of real water bodies,this thesis carries out underwater image simulation experiments to simulate the underwater imaging process of different light conditions,different water bodies and different water depths.Simulation for the model shows that the underwater imaging models established in this thesis can describe the degradation process of underwater optical image more systematically than traditional imaging model.(2)Aiming at the limitation of ill-posed prior assumptions under complex illumination,this thesis analyzes the non-uniform color deviation caused by the attenuation of artificial light and natural light in the incident stage based on the multi-stage light attenuation based underwater imaging model,and proposes a local white balance method to improve image color and suppress the over-enhancement of highlight areas.By studying the relationship between transmission and scene depth,this thesis transforms transmission estimation problem into a normalized scene depth estimation problem,and proposes an image descattering method based on total variational scene depth estimation.The objective function of scene depth optimization is designed according to the prior knowledge of scene depth,and solved by the alternating direction multiplier method.Combined with the background light at infinity solved by the quadtree hierarchical search method,the backscattering effect in the scene-camera stage is removed.Experimental results show that this method can effectively solve the low contrast problem caused by backscattering and the non-uniform color distortion problem caused by complex illumination.(3)Aiming at the problem of ill-conditioned estimation of model parameters caused by different attenuation properties of different types of water environments,this thesis presents an underwater image enhancement method based on multi-attenuation parameter joint-learning generative adversarial networks by combining multi-stage attenuation based underwater imaging model with deep neural networks.It establishes multi-parameter learning tasks for different light attenuation stages,improves the stability of parameter estimation through deep neural networks,and improves the visual quality of underwater images in terms of color,contrast,and details.This thesis combines the multi-stage attenuation based underwater imaging model with the absorption and scattering properties of various real water environments to establish a multi-parameter general synthetic underwater image dataset of diverse water bodies.It provides training data for data-driven underwater image enhancement methods and can meet multi-parameter and end-to-end training requirements.Experimental results show that this method can effectively improve the image visual quality in diverse attenuated water environments.(4)Aiming at the incomplete removal of backscattered light in high turbidity water environments with strong scattering,this thesis proposes an underwater image enhancement method based on multi-polarization state information fusion based generative adversarial networks by combining the light polarization characteristics with deep neural network.It establishes a mapping function between multi-polarization state information and scene radiation,and use the polarization state difference between the scene reflected light and the backscattered light to suppress the backscattered light.It enables end-to-end underwater image enhancement and avoids the bias caused by traditional parameter estimation and background region selection.This thesis establishes an underwater polarized image dataset covering different turbidity in various water environments,which can provide data support for data-driven polarization image enhancement model training and validation.Experimental results show that this method can effectively eliminate the strong backscattering effect in high turbidity water environments.Aiming at the key problems of underwater optical image enhancement methods,this thesis designs several underwater optical image enhancement methods based on physical models and deep learning.They solve the degradation problems of underwater optical images in complex water environments,and improve image visual quality to support underwater operation tasks based on optical vision.The research content of this thesis has important theoretical significance and practical application value.