Research On Underwater Scene Enhancement Method Based On Attention

With the growing demand for land resources,the ocean has become an attractive option for exploration and exploitation due to its abundant natural resources.The quality of underwater imaging is often degraded due to factors such as color deviation,scattering blur,and low contrast.As a result,improving the visual quality of underwater images has become a significant area of research.To meet the visual needs of the changing underwater environment,more and more researchers are applying deep learning techniques to enhance the quality of underwater images.Recently,Transformer-based methods have shown the same excellent performance as convolutional neural networks,but the large number of parameters required for these methods hinder their practical deployment.In addition,for most networks with encoder/decoder structure,the direct skip connection ignores the differences between pre and post features.The main research content of this thesis is as follows:(1)This thesis proposes a novel adaptive group attention,which can effectively reduce the number of attention parameters by dynamically selecting visually complementary channels based on their dependencies.This adaptive group attention is applied in the Swin Transformer module to design an end-to-end underwater image enhancement network.The network also incorporates the multiscale cascade module and the channel attention mechanism to improve its performance.This thesis conducted ablation study,visual analysis,and metrics evaluation.The results show that the application of adaptive group attention significantly compresses the model size while ensuring performance,and other application components have significant gain on the network.Compared with other advanced methods,the network in this thesis has outstanding performance.(2)This thesis proposes an underwater image enhancement network focusing on prepost differences,which uses a multiscale input and output structure.In this thesis,the cross-wise Transformer module is designed to capture the dependencies between multiscale features,which guides the interactive learning of features in different periods.In addition,feature supplement module is designed at the initial stage of each scale to fuse the multiscale input features,and efficient channel attention and residual learning modules are introduced to highlight the important features and ensure gradient information of the network.The experiments on several datasets show that the adopted modules lead to significant improvements in network performance,and the proposed network has outstanding performance in visual comparison and quantitative metrics.