基于多模态退化特征学习的水下图像增强

打开文本图片集

中图分类号：TP391文献标志码：A

Multi-modal Degradation Feature Learning for Underwater Image Enhancement

XIONG Qingbo 1 ，CHEN Lei 1 ， LIANG Xiaoli 1 ， LIU Tianxu²

（1. School of Software，Henan University，Kaifeng 45OO46，Henan，China; 2.Henan Provincial Transportation Dispatching Command Center，Zhengzhou 45Oo16，Henan，China）

Abstract：Toaddress the lack of generalizationand flexibilityin traditional underwater image enhancementmodels，a multi-modal degraded contrastive language-image pre-training（MD-CLIP）model was proposed.MD-CLIP model was trainedusingcontrastive learning toencodetheimage featuresand textfeaturesof low-qualityunderwaterimages into multi-modaldegraded features.Across-atentionmechanismand prompt embedding wereused to integrate themultimodal degraded featurespredictedbyMD-CLIP modelintotheunderwaterimageenhancementmodel，adjustingthe model's performance and generalization.Ablation and comparativeexperiments were conducted to validate the ffectivenessof themulti-modal degraded features.Theresultsshow that the multi-modal degraded featurespredicted by MD-CLIP model were embed into theunderwater image enhancement modelbyusing cross-atention mechanism，the image enhancement performanceand generalization performance of the model are significantlyimproved.MD-CLIP model can be added to other image enhancement models as a universal enhancement module.

Keywords：underwaterimage enancement；multi-modaldegradation feature;；contrastivelearning；cross-attentionmechanism

近年来，随着海洋资源开发的兴起，水下图像增强技术备受关注。（剩余15082字）