Underwater Image Enhancement Based On Cycle-Consistent Adversarial Network

Due to the complex underwater environment and the attenuation of light in the water,the quality of the image taken by the camera is seriously degraded,and degradation phenomena such as color cast,blur,and low contrast will occur.These degraded underwater images have certain limitations in display and analysis.For example,underwater images with high blurriness and serious color cast will greatly reduce the accuracy of marine life recognition.Therefore,in order to overcome the limitations of degraded underwater images and obtain high-quality underwater images,underwater image enhancement technology is particularly important.Based on this problem,this thesis proposes an improved generative confrontation network model based on the cyclic consistency confrontation network,which can fully learn the mapping between unpaired turbid underwater images and clear underwater images while preserving the image content.The turbid underwater image is converted into a clear underwater image to realize the enhancement of the underwater image.The main work content is as follows:(1)This article believes that an image has characteristic space,style space and structural space.Affected by different underwater operation scenes,different images have great differences in apparent style and structure.Therefore,underwater images in different domains are difficult to share the same style space and structural space.Under the influence of underwater targets,underwater images in different domains should share the same feature space.(2)This thesis designs a cyclic consistency confrontation network to achieve underwater image enhancement.The network improves the basic network for the three spaces.The specific method is: for the feature space,in order to better retain the features of the original underwater image,the feature extractor is used to extract the image features,and a pair of feature discriminators are added to identify and extract the features.Whether it is the feature of the input underwater image,and the process of feature loss constraint is adopted;for the style space,this thesis designs the color deviation loss based on the Shade of Gray to constrain the color migration effect between images in different domains;for the structural space,Increase the structural similarity loss to constrain the enhancement effect of the image on the structural details.(3)Because the network model proposed in this thesis has cyclic consistency,the model has the characteristics of two-way mapping.While achieving underwater image enhancement,its backward mapping process can convert clear underwater images into turbid underwater images,And then construct a turbid underwater image synthesis data set,which can be used in end-to-end deep learning training.Through a qualitative and quantitative comparison of the enhancement methods in this thesis,a large number of experimental results show that the underwater image enhanced by the method in this thesis obtains better visual quality than several other existing enhancement methods,and can effectively remove the turbid underwater image.Color cast,improve image contrast and reduce blur.In addition,in the feature point detection and matching experiment,it has also played a significant role in improving.