| Underwater environments contain impurities that cause light scattering and absorption,leading to color deviations,blurriness,and distortion in underwater images.Therefore,research on underwater image restoration has become essential.Deep learning-based underwater image restoration faces several challenges.Firstly,the difficulty in data collection results in a lack of large-scale underwater image datasets for training image restoration models,and the visual quality of the Ground-Truth generated by existing image rendering algorithms is also suboptimal.Secondly,existing underwater image restoration models generally suffer from a lack of ability to extract global information from feature maps and capture local contextual details.To address these issues,this study conducted the following research:Firstly,a underwater image rendering model based on image style transfer,named UWST network,is proposed.In response to the limitations of traditional physically-based methods in considering various physical priors and the susceptibility of GANs to mode collapse,the UWST network is designed based on style transfer.It adaptsively learns the style of real underwater images and renders them into synthesized underwater images representing various underwater scenes.UWST first employs a deep prediction module to calculate the depth map of the clean air domain image and integrates it with the clean air domain image.Then,in the Content Encoder and Style Encoder stages,a spatial and channel attention fusion module is embedded to enhance the network’s ability to extract crucial features.Finally,in the decoder stage,AdaIN is introduced to adjust the mean and variance of the target image,enabling the migration and fusion of the Style map into the Content map during the upsampling process,thus achieving style transfer on the target image.Experimental results demonstrate that compared to various underwater restoration networks,UWST network renders underwater images that are closer to the style of real underwater images.Training underwater image restoration networks using the UWST-rendered UWSTD underwater image dataset shows better restoration performance compared to other typical datasets.Secondly,a underwater image restoration model,SwinTUNet-UIR,based on ConvsSwin Transformer is proposed.It addresses the limitations of UNet based on pure convolution in learning global information and interaction of distant semantic information,as well as the problem of Swin Transformer’s limited ability to extract local information from feature maps.A channel and spatial attention fusion module is introduced,which consists of two convolutional layers responsible for channel attention and spatial attention,respectively,replacing the linear layers in the original Swin Transformer’s W-MSA and SW-MSA multi-head attention structures.By constructing a Convs-Swin Transformer based on the novel Convs-W-MSA and Convs-SW-MSA convolutional multi-head attention modules and embedding it into the UNet model,the ability of the model to extract features is further enhanced.To address the issue of chessboard artifacts that often occur during restoration,a double upsampling method,combining Bilinear and PixelShuffle,namely Double Up-sample,is used,replacing the original Patch Expanding upsampling method.By conducting comparative and ablation experiments on the UIEB,UFO-120,and UWSTD datasets,the effectiveness of the improved algorithm is verified,and the positive contribution of each key module to restoration performance is demonstrated. |
