Research On Underwater Image Enhancement Method Based On GAN Networks

Underwater optical images are widely used in fields such as underwater resource exploration and underwater biological perception.However,in underwater environments,particularly in marine settings,light scattering and attenuation cause underwater images to suffer from blurring,low contrast,and color shifts,severely affecting tasks such as underwater target recognition.Therefore,it is of great significance to conduct in-depth research on underwater image enhancement methods to improve the quality of underwater images.Traditional image enhancement methods suffer from poor generalization and the need for prior knowledge,making them insufficient to meet the current demands of underwater image enhancement.Deep learning,especially generative adversarial networks(GANs),provides an effective tool for underwater image enhancement by capturing diverse features and improving model accuracy through adversarial training.This thesis aims to address the issues of color shifts,blurring,and low contrast in underwater images by investigating image enhancement methods based on generative adversarial networks.The main research contributions of this thesis are as follows:(1)This thesis provides a comprehensive analysis of the principles of underwater optical imaging and explores the effectiveness of traditional methods such as white balance,gamma correction,and sharpening for underwater image enhancement.Multiple deep convolutional neural networks and generative adversarial networks are compared for underwater image processing,with a focus on analyzing the principles and structural designs of U-Net,UWGAN,and the fused generative adversarial network FGAN.This provides theoretical and methodological support for designing deep learning networks in underwater image enhancement.(2)To address the severe color distortion in underwater images,an improved generative adversarial network called UIGAN is proposed,which incorporates a color balance fusion method.The method first introduces an image pre-processing module,including techniques such as white balance,gamma correction,and sharpening,to mitigate the effects of color distortion.It then designs a multi-branch feature extraction module that combines high-level and low-level features through convolution and skip connections to enhance image contrast.Finally,multiple loss functions are employed to train and optimize the network.The proposed method is compared with seven other image enhancement methods on three publicly available datasets,namely EUVP,UFO-120,and UIEB.The results demonstrate that the method achieves visually pleasing effects in terms of overall color saturation and contrast enhancement.It also outperforms other deep learning methods by over 4%in color-related evaluation metrics and achieves above-average results in metrics related to image blur.(3)To further address the issues of image blurring and low contrast in underwater images,a generative adversarial network called DRSGAN is constructed based on a dense residual block structure.To avoid the problem of gradient vanishing during backpropagation,residual convolutional modules are designed,and skip connections are incorporated into different levels of convolutional modules,enhancing the network’s ability to extract detailed information from images.The network also incorporates adversarial,content-aware,and global representation loss functions to improve its learning capacity for image structure,details,and colors.In comparison with typical deep learning enhancement methods on public datasets,DRSGAN achieves significant improvements subjectively in terms of color,contrast,and blur.From a quantitative perspective,it achieves over 5%improvement in comprehensive evaluations of blur,contrast,and color.