纹理感知联合颜色直方图特征的水下图像增强  

Underwater image enhancement using joint texture perception and color histogram features

作　　者：袁国铭 刘海军 李晓丽 张瑞蕾 单维锋 YUAN Guoming;LIU Haijun;LI Xiaoli;ZHANG Ruilei;SHAN Weifeng(Department of emergency management,Institute of Disaster Prevention,Sanhe 065201,China)

机构地区：[1]防灾科技学院应急管理学院,河北三河065201

出　　处：《光学精密工程》2024年第13期2112-2127,共16页Optics and Precision Engineering

基　　金：中央高校基金资助项目(No.ZY20180223);河北省自然科学基金资助项目(No.D202251200)。

摘　　要：基于深度学习的水下图像增强方法在视觉效果上表现出色,但已有的端到端网络较少针对水下场景中常见的颜色失真和纹理模糊来精心构建网络结构。为了提高网络的性能,提出了纹理感知联合颜色直方图特征的水下图像增强网络,它包含纹理感知网络,颜色直方图特征提取网络以及颜色纹理融合的水下图像增强网络。纹理感知网络的构建引入变形Transformer模块,该模块利用空间感知的变形卷积设计Transformer模块的多头注意力机制,有效地将变形卷积的几何感知能力与Transformer模块的全局语义捕获能力相结合来提取纹理特征。颜色直方图特征提取网络认为直方图中包含丰富的颜色信息,利用真实水下图像的直方图来监督该网络提取颜色特征。最后,利用所提的颜色纹理融合模块将前两个网络提取的纹理和颜色特征融合,并注入到后续的颜色纹理融合的水下图像增强网络中,以实施水下图像增强,该操作不但有效地保留了纹理结构,校正了失真的颜色,还使得纹理和颜色信息保存一致。实验结果表明,本文算法较已有的水下图像增强算法,具有更好的视觉增强效果,水下图像质量测量指标(Underwater Image Quality Metric,UIQM)提高10%,单张图像的运行时间缩短了9%,仅为0.051 s,满足工程实践中水下视觉增强任务要求。While deep learning methods show promising visual results,current end-to-end networks often lack tailored architectures to address common issues like color distortion and texture blurriness.To im⁃prove their effectiveness,we propose an underwater image enhancement network that utilizes joint texture perception and color histogram features.The network comprises a texture-aware module,a color histo⁃gram extraction module,and a color-texture fusion enhancement module.The texture-aware network in⁃corporates a deformable transformer module,leveraging spatially aware deformable convolution to en⁃hance multi-head attention and extract texture features.The color histogram extraction module harnesses histograms from real underwater images to compute the loss function.Subsequently,the color-texture fu⁃sion module merges the color and texture features,which are then processed by the enhancement network to produce the final results.This approach effectively preserves texture structures,corrects color distor⁃tions,and maintains information consistency.Extensive experiments demonstrate that our method surpass⁃es existing underwater image enhancement algorithms,achieving a 10%increase in the UIQM metric and reducing processing time to just 0.051 s per image.Our model successfully meets the demands of underwa⁃ter visual enhancement tasks.

关 键 词：图像增强 多头注意力机制 纹理感知 卷积神经网络 深度学习 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]