Research On Underwater Image Enhancement Algorithm Combining Multi-Color Space And Contrastive Learning

With the increasing demand of human society for marine resources,underwater vision tasks play an increasingly important role in acquiring marine information.However,due to the absorption and scattering of light,the underwater image has problems such as blurring,low contrast,and color distortion,which is not conducive to the extraction of underwater image information and the development of subsequent advanced visual tasks.With the rapid development of deep learning and powerful feature learning ability,it is widely used in underwater image enhancement tasks.However,the current research still has the phenomenon of partial loss of details and oversaturation.At the same time,deep learning is based on data-driven,and the lack of data sets is also a problem that limits the further improvement of image quality.Aiming at the above problems,this paper studies the algorithm based on multi-color space and contrastive learning.By expanding the data set,the detailed information of the image is enhanced while improving the image quality.The main work of this paper is as follows:(1)In view of the small number of paired data sets,this paper adopts the method of style transfer to migrate the characteristics of underwater images to high-definition images in the air,and form paired data through high-definition air images and air images with underwater characteristics.Combine some public data sets to generate your own data set,increase the amount of training data,and thus improve the generalization ability of the model.Then,for the problem of insufficient extraction ability of the network,this paper proposes an enhanced context residual block,which improves the feature extraction ability of the module by starting from the width of the network.The four branches use different receptive fields,interact global information with local information,and allocate key regions and connections between channels through the channel spatial attention mechanism,further improving the feature extraction ability of the network.Finally,this paper proposes an underwater image enhancement algorithm based on the enhanced context residual block.Through end-to-end network learning,combined with the powerful extraction ability of the enhanced context residual block,to achieve the purpose of color correction and deblurring.(2)In order to further improve the detailed information of the generated image,this paper proposes an underwater image enhancement algorithm that combines brightness information and contrast learning.The input image is converted into a Lab color space image,and the semantic information and the detailed information are better interacted through the codec structure and the interactive connection of static features.At the same time,a contrastive learning loss is introduced to make the generated image closer to the high-quality image and deviate from the low-quality image,which improves the image generation quality.Experimental results on real underwater datasets show that compared with other underwater image enhancement methods,the proposed method achieves the better results on Peak Signal-to-Noise Ratio,Structural Similarity and Underwater Image Quality Measure.