Research On Underwater Image Enhancement Methods Based On Generative Adversarial Network

Due to the limited and decreasing reserves of available resources on land,human beings are increasingly dependent on the abundant resources in the ocean,especially in terms of food and renewable energy.Countries that master the technology of efficient development of marine resources are bound to be in a leading position in the future economic and social development.Underwater images,as a significant medium for conveying marine information,are capable of intuitively depicting the seabed world and serve as a critical element in advancing the development of marine resources as well as detecting the underwater environment.The underwater imaging environment is complex and light propagating in water is subject to selective absorption and scattering effects,resulting in significant degradation of the quality of collected underwater images.The degraded underwater image limits its application in practical scenarios such as marine resource exploration,underwater target detection and marine archaeology.Therefore,it is of great significance to design an effective underwater image enhancement method to improve the efficiency of human development and utilization of marine resources.After years of development and improvement,the deep learning algorithm has made great breakthroughs in many fields,and underwater image enhancement has become an important research field in deep learning.Using deep learning algorithm to solve the degradation problem of underwater images can not only solve the problem of poor enhancement effect of traditional methods,but also effectively reduce the inference time of the algorithm.Despite the advances in deep learning-based underwater image enhancement methods,the complexity and diversity of underwater scenes present significant challenges.Therefore,there is still room for improvement in this field.In order to tackle this problem and enhance the visual quality of underwater images,two underwater image enhancement models with robust generalization and scalability are proposed in this paper,addressing the complexity and diversity of underwater scenes.The main research objectives of this paper are outlined below:(1)To tackle the issues of color distortion,low definition,and low brightness in underwater images,this study presents an underwater image enhancement technique based on a multi-scale residual generative adversarial network.First,the data used in training needs to be scaled to a uniform size.Then,a multi-scale residual module is introduced into the generative network to make full use of multi-scale fusion and residual ideas to improve network performance and overcome the problems of gradient disappearance and detail loss during network training.Then the spectral normalization method is used to stably discriminate the training of the network.Finally,the global similarity loss function is added to the standard adversarial network loss function to focus on the image features of the reference image.The experimental results demonstrate that the proposed method outperforms the comparison method,with average values of PSNR,SSIM,UCIQE,and UIQM being 22.0745,0.7458,0.3825,and 2.9613,respectively,which are 1.5 %,4.6 %,1.7 % and 0.5 % higher than the second place.It is proved that the proposed method performs well in correcting color deviation,improving brightness and clarity,and can effectively improve the visual quality of underwater images.(2)In order to solve the problem that the existing underwater image enhancement methods fail to simultaneously consider the inconsistent attenuation of underwater images in different color channels and spatial regions,an underwater image enhancement method combining Transformer and generative adversarial network is proposed.By global feature modeling Transformer module based on spatial mechanism and channel-level multiscale feature fusion Transformer module integrated in generative network,the overall network’s attention to color channels and spatial regions with more severe attenuation is enhanced.In addition,a multiple loss function combining RGB and LAB color spaces is designed to make the color performance of the generated image better.The experimental results indicate that the proposed method outperforms the comparison method,with average values of PSNR,SSIM,UCIQE,and UIQM being 22.0745,0.7458,0.3825,and 2.9613,respectively,which are 8.5 %,3.7 %,2.4 % and 3.6 % higher than the second one.It is proved that the proposed method can effectively correct the color deviation in underwater image enhancement,uniformly improve the clarity,and improve the visual perception effect of the image.