Research On Underwater Image Enhancement Based On Transformer Texture Sharpening And Color Recovery

In the last decade,the unexplored marine resources have been gradually brought to the attention of human beings,and the ocean as a huge treasure trove of natural resources will be gradually developed in the future.As the main carrier of ocean information,the underwater images collected from the ocean can visually reflect the ocean environment,ocean resources and other related ocean hydrographic parameters.Due to the complex marine underwater environment and lighting conditions,the collected underwater images are usually degraded by wavelength-related absorption and scattering effects and the corresponding marine dust,which cause low contrast,poor detail performance and color bias in the collected marine underwater images,limiting the use of underwater images and videos in marine environment monitoring,marine artifact exploration and salvage,marine resource detection,marine bionic These factors limit the practical application of underwater images and videos in marine development strategies such as marine environmental monitoring,marine artifact exploration and salvage,marine resource exploration,marine bionic detection,and marine target detection.To address these problems,in order to improve the quality of the acquired underwater images while enriching the marine information they carry,researchers have conducted experimental studies in three branches: non-physical model-based image enhancement methods,physical model-based image enhancement methods,and deep learning-based enhancement methods to restore the images.In recent years,deep learning-based methods have been widely used by researchers for various image enhancement tasks because of their superior feature learning capability over the other two methods,and the results are better than traditional methods.However,because the underwater environment is limited by the latitude,longitude and hydrological conditions,the deep learning-based image enhancement methods still have some room for optimization in terms of color and detail restoration of degraded images,so this paper proposes a multi-input Trans Former and convolutional neural network CNN-based underwater image restoration method that can effectively improve the color and detail rendering of images.The research of the text is experimented in two aspects as follows:In the first part,a multi-input underwater image recovery method based on Trans Former and convolutional neural network CNN is proposed to solve the problems of color distortion,low contrast and loss of image details due to light attenuation,absorption and scattering by seawater and impurities in seawater during underwater imaging.First,the method will use multiple inputs,divided into the original degraded image,the ACE image obtained by automatic color equalization and the texture image extracted by Relative Total Variatio(RTV)method.The three input images will be scaled to 256×256×3 pixel size to facilitate subsequent processing.Secondly,this method uses Trans Former and Relative Total Variatio to construct a depth feature extraction module,which fuses the texture image extracted by Relative Total Variatio with the image information extracted by Trans Former to provide reliable feature information for subsequent neural network recovery,thus effectively enhancing the detail features of the image.Again,the method in this paper uses automatic color equalization and Lab color space to construct a color correction module,which contains two color spaces and an ACE image,and can achieve to enhance the image contrast and correct the color at the same time.Then,the method uses multiple loss functions to constrain the network convergence,which contains global similarity loss,structural similarity error loss and content perception loss,and obtains the enhanced clear underwater image under the combined effect of multiple loss functions.Moreover,in order to improve the generalization performance of the method,a dataset UIPD with many different underwater environment features is constructed,and the method is trained with this dataset and has good results in enhancing images with different degradation features.Finally,the quantitative and qualitative comparative analysis of this paper’s method with other methods on the test set is carried out,and four metrics,PSNR,UIQM,SSIM and UCIQE,are used for the quantitative.The experimental results index values of 23.2175,0.8979,3.4177 and 0.5856 indicate that the images processed by the method in this paper are better than other comparison methods in terms of sharpness,contrast,color representation and texture information.In the second part,in order to improve the utilization efficiency of the model,this paper supports the subsequent application of the method by lightening the model of the method proposed in the first part,and compressing the parameter scale of the model as much as possible while ensuring the recovery effect of the method on the underwater images.The experimental results show that the optimized method is basically consistent with the method proposed in Part I in terms of indexes,and the method effectively corrects color deviation,improves contrast,sharpens detailed textures,and the enhanced image is more colorful,brighter,and overall clearer than the original image.At the same time,in order to make the method of this paper clearer and easier to be applied to solve specific problems,and to ensure that low-basic or non-related practitioners can quickly apply the results of this paper,this paper developed a set of visual operation software using the Python language implementation of the Qt framework,namely the PyQt5 GUI library.The software includes features such as automatic loading of deep learning models,information prompting,custom saving of results and multi-threading to speed up the model running,which can achieve low latency and real-time visualization of results.