End-to-End Underwater Image Restoration Based On Deep Learning

In recent years,underwater images and videos have played an increasingly important role in marine military,marine engineering,and marine research,so it is especially important to obtain high-quality images from the deep sea.However,poor underwater imaging environment and lighting conditions lead to underwater images often with color distortion,low contrast,texture blur and other quality degradation problems.For this topic,we conducted extensive literature research and learned that the current methods for underwater image restoration are mainly divided into traditional methods and deep learning methods based on Generative Adversarial Networks.Among them,traditional methods are limited by assumptions and prior knowledge,the models are inaccurate and color correction effects are not ideal,and existing methods based on Generative Adversarial Networks can correct the color of the underwater image to a certain extent,but they use the generative adversarial network to generate the dataset,ignoring the optical properties of the underwater imaging,so their robustness is not strong,and the image restoration effect is not ideal.This paper focuses on the above issues,our research content and main contributions include the following points.Firstly,aiming at the problem that the paired clear underwater images and the degraded underwater images dataset are difficult to obtain,a degraded underwater images simulation method based on the optical principle of underwater imaging is proposed.According to the physics principle of underwater imaging,that is,the light will propagate exponentially underwater,and the attenuation of light of different colors is different,a simulator to generate degraded underwater images is constructed.We use this simulator generates 17 different degraded underwater images for each clear underwater image and ultimately form a multi-pair clear and degraded underwater image dataset for subsequent training and model evaluation of deep learning networks.Secondly,in order to solve the problem that the existing underwater image restoration methods based on Generative Adversarial Networks cannot process degraded underwater images well,a deep learning network framework for underwater image restoration based on adversarial learning is proposed.First,the original Generation Adversarial Networks have the problems gradient disappearance,mode collapse and so on,so this paper introduces WGAN with gradient penalty term to guide the generator and discriminator to train.Second,since the original Generative Adversarial Network is unable to generate data directionally,we use a conditional Generative Adversarial Network to enable the generator to generate a clear underwater image that is consistent with the input degraded underwater image content.In addition,we add perceptual loss to the generator to further constrain the detail and semantic information of the generated image.In the discriminator,we use a fully convolutional block-based discriminator to learn the structural loss,rather than learning the entire image-level or pixel-level loss.Finally,the proposed method is applied to the task of underwater image restoration and the experimental results show that the method has better results in visual and quantitative indicators than other methods.Thirdly,since the effects of underwater image restoration methods on underwater image processing is not ideal,a network for underwater image restoration based on residual connection and scaled iterative network is proposed.We designed a encoder-decoder network based on residual blocks as the basic network,which includes some symmetric convolution and deconvolution layers to learn the end-to-end mapping of degraded underwater images to clear images.In order to strengthen the training ability of the network,we add a residual skip connection in the symmetric part of the convolution and deconvolution layer of the encoder-decoder network to make the network training more stable and obtain better results.In addition,in order to further improve the accuracy of the network,we propose a multi-scale iterative network training method,which restores the content and details of the image from coarse to fine,and the network parameters of image restoration under coarse and fine granularity are the same.This is equivalent to deepening the depth of the network without increasing network parameters.In addition,we designed a multi-scale loss function to train the network.Experiments show that the proposed method can effectively strengthen the network training,make the network parameters more optimized,and the recovery of degraded underwater images has better results in visual and quantitative indicators.