Research On Underwater Image Enhancement Technology Based On Deep Learning

In exploring the ocean,high-quality underwater images can help researchers better analyze the marine environment.However,due to the different absorption and scattering of different wavelengths of light in the marine environment,as well as the influence of suspended particles and microorganisms in seawater,underwater images often suffer from blue-green bias,blurring and low contrast.Compared with traditional image processing methods that are only applicable to specific scenes,deep learning-based methods have better generalization in the field of underwater image enhancement,and their network types can be divided into convolutional neural networks and generative adversarial networks.However,the current relevant deep learning network models have the problems of large number of parameters and slow inference speed,so the lightweight underwater image enhancement model has important research significance,and the main work of this thesis is as follows for this direction.Firstly,to address the problems of large number of parameters and low real-time performance of current convolutional neural network models,we propose a parallel multi-scale underwater enhancement network using lightweight convolutional approach,and propose a multi-scale fusion strategy based on the scaled dot product attention mechanism and a multi-level residual module to enhance the feature extraction capability of the network,which can reduce the number of parameters and improve the speed of the network processing images while ensuring the quality of underwater image recovery..Secondly,to address the problem of insufficient feature extraction capability of the improved U-Net-based underwater image enhancement network in generative adversarial networks,a hybrid module fusing dense connections and residual connections is proposed to enhance the feature extraction capability of the U-Net bottleneck layer,while a hybrid pooling attention module is used to improve the feature fusion effect of jump connections in U-Net.Finally,we propose a lightweight "U"-shaped double downsampling module for the improved U-Net network,and replace the interpolation upsampling process with pixel reorganization to reduce the number of model parameters and verify the stability of the model in an unsupervised training environment.