| Underwater images are an important carrier and representation for underwater information perception.However,due to the attenuation and scattering effects of light signals in water,underwater images suffer from degradation problems such as color distortion and fogging effects,which affect the applications in the field of underwater vision.In the past decades,recovery techniques and enhancement techniques have been proposed for underwater images as a way to improve the image vision quality.In this dissertation,extensive and systematic experiments are conducted to explore the characteristics of different image processing methods from two perspectives,enhancement and restoration,respectively,in studying the underwater image degradation problem.The specific research is divided into the following:(1)To solve the color distortion problem caused by the selective attenuation of light in underwater images,this dissertation proposes an image enhancement network based on dense residuals and attention mechanism.Deep dense residual blocks are used to fuse image features at different levels and representative feature channels are emphasized by non-equivalent weight assignment.The network to model mapping is accomplished by setting model parameters which preserve the physical interpretability of underwater imaging.The experimental results show that the enhancement method shows great performance in color reproduction which improves the color vividness and realism of the images.(2)To solve the problem of fogging effect and low contrast due to light scattering in underwater images.This dissertation proposes an underwater image restoration network based on backscattering prior and depth bilateral learning.The depth information is obtained using monocular depth estimation,the fogging effect of background light is removed using the backscattering prior,and finally the underwater transmission map is obtained using the depth bilateral affine network to recover the color and contrast of the image.The experimental results show that the recovery method can effectively remove the fogging effect for underwater images and improve the clarity and contrast of the images.(3)This dissertation combines and continues the experimental study of underwater image enhancement and image restoration,and proposes an enhancement algorithm that incorporates imaging models and convolutional networks,combining the properties of underwater prior knowledge and autonomous learning ability to jointly enhance color correction and defogging.In addition,this dissertation tests the application of the enhancement method in the fields of image stitching and target detection to further validate the feasibility of the method.This dissertation takes two perspectives of image enhancement and image restoration.We propose two different approaches to solve the problems of color distortion and fogging effect.And both are integrated with each other to solve the underwater image degradation problem together,showing better algorithm accuracy and stability.This is worthwhile for the development of underwater machine vision enhancement technology. |
