摘要
针对水下图像质量退化严重的问题,提出一种端到端的基于融合逆透射率图的水下图像增强算法.将原始RGB图像和基于传统方法得到的逆透射率图分别输入到双流卷积神经网络的2个编码器中;通过跨模态特征融合模块使得2种图像信息充分融合互补,让网络更好地学习到水下光学成像的特点;通过特征增强模块,增强特征的表达能力;通过残差解码模块连接解码器和编码器,以补充和丰富RGB特征.通过逆透射率图的水下图像增强算法以及跨模态跨尺度的信息融合,由粗到细地进行逐级处理,最终输出增强后的RGB图像.实验结果表明,所提算法能够有效地提升水下图像视觉质量.综合主观评价和客观评价,所提算法优于所对比的白平衡WB、直方图均衡化HE、Water-Net、UGAN、UWCNN、Ucolor 6种算法.
An end-to-end underwater image enhancement algorithm via fusing reverse medium transmission map was proposed to resolve the problem of quality degradation in the underwater images.The original RGB map and the reverse medium transmission map obtained by the traditional method were fed into two encoders of the proposed two-stream convolutional neural network.The information of two images was fully integrated through the crossmodality feature fusion module,and the network could learn the characteristics of underwater optical images well.The feature expression capabilities could be further strengthened through the feature enhancement modules.The decoder and encoder were connected via a residual decoding module to supply the RGB features.The underwater image enhancement algorithm via fusing reverse medium transmission map processes information from coarse to fine through cross-modality and cross-scale information fusion,and finally outputs an enhanced RGB image.The experimental results showed that the proposed algorithm could effectively improve the visual quality of underwater images.Taking into account both subjective evaluation and objective evaluation,the proposed algorithm was better than the six competing algorithms of white balance WB and histogram equalization HE,Water-Net,UGAN,UWCNN,Ucolor.
作者
张剑钊
郭继昌
汪昱东
ZHANG Jian-zhao;GUO Ji-chang;WANG Yu-dong(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2023年第5期921-929,共9页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(62171315).
关键词
逆透射率图
水下图像增强
水下光学成像
双流卷积神经网络
特征融合
reverse medium transmission map
underwater image enhancement
underwater optical imaging
two-stream convolutional neural network
feature fusion
