Research On Underwater Image Enhancement Based On Semantic Guidance And Contrastive Learning

Underwater image enhancement is an indispensable part of underwater image processing,which has a wide application prospect in underwater environment perception,marine environment monitoring,underwater engineering monitoring and other fields.With the rapid development of deep learning in the field of computer vision,underwater image enhancement algorithms have made great progress.However,there are still many challenges in the field of underwater image enhancement due to the scarcity of paired data sets in real scenes,the quality of reference images and the lack of effective supervision signals in underwater image enhancement networks.To solve the above problems,this thesis investigates and analyzes the existing underwater image enhancement algorithms,and proposes an underwater image enhancement algorithm based on semantic guidance and contrast learning to improve the performance of image restoration.This thesis mainly includes the following research contents:(1)Considering the lack of more effective supervisory signals for training underwater image enhancement networks,this thesis attempts to add semantic information as an additional supervision signal to the underwater image enhancement network,and guide the enhancement network to adopt consistency enhancement for semantically related areas to improve image edge blur.In addition,for some rare degradation types but semantically related in image training set,semantic information provides the network with prior knowledge,which improves model performance and enhances model generalization ability.In this thesis,the feature attention fusion mechanism is introduced to avoid the loss of context information caused by the direct fusion of cross-domain information,so as to better combine semantic information and reconstruct features,and maximize the guiding role of semantic information.Considering the existence of uneven degradation in underwater images,through the combination of spatial attention and channel attention,the network assigns more reasonable weights to the seriously degraded areas and improve the enhancement effect of images.(2)Considering the global visual representation learned by the existing contrastive learning framework is not enough for low-level visual tasks requiring rich texture details,and the existing data enhancement methods generating positive and negative samples will destroy the pixel correspondence,which is not suitable for low-level visual tasks,the contrastive learning framework in this thesis adopts GAN-like training mode.In the training process,the degradation sensing ability of the contrastive discriminator is improved,and the contrastive discriminator is used as the feature embedding network to measure the feature distance.By adding color bias to the enhanced reference image,a series of underwater images containing random color bias are generated as negative samples so as to improve the color perception ability of the network.The enhanced reference images are selected as positive samples.The two-way constraints of positive and negative samples on the enhanced network further improve the network performance.The underwater image enhancement network proposed in this thesis based on semantic guidance and contrastive learning performs well on multiple real underwater datasets and accords with human visual perception.The underwater image enhancement network of this thesis achieves 24.30 and 0.9144 in PSNR and SSIM metrics respectively.In addition,the contrastive learning framework presented in this thesis also be applied to other underwater image enhancement networks to improve the performance of network enhancement.After the introduction of the contrastive learning framework,the networks’ PSNR and SSIM metrics maximumly increased by 0.16 and 0.0017 respectively.