摘要
传统的图像融合去雾算法通过统计大量图像特征来设置融合权重,不仅耗时费力,还易出现误差.为此,本文提出了图像融合的循环神经网络来解决该问题.首先,从初始雾图中推导其衍生的白平衡图像、对比度增强图像、伽马校正图像,作为去雾所需的融合图像;随后,构建编码解码网络的去雾模型,并将三张融合图像与初始雾图相串联,共同作为网络的输入.利用此网络学习和生成融合图像对应的权重图,以融合信息估计无雾图像,从而解决传统图像融合去雾算法中权重计算耗时费力,易产生误差的问题;最后,为了进一步优化去雾结果,在编码解码的网络模型中嵌入循环单元,构建循环编码解码网络,即将上次循环时网络输出作为下次循环时网络输入,同时使循环单元中的隐藏状态也随之传递,以便更好优化去雾结果.实验结果表明,在合成和真实图像的测试下,本文算法都具有较高去雾精度,与已有算法相比,其去雾精度提高了19%,能有效用于工程实践中.
Conventional image dehazing methods via image fusion strategy set fusion weights depending on static features of images,which not only cost time but also easily occur error.In order to solve these problems,we propose a recurrent network based on images fusion scheme.Firstly,we derives the white balance image,contrast enhanced image and gamma correction image from the hazy image,as the fused images.Secondly,these images concatenated with the hazy image serves as the input of our proposed encoder-decoder network.By learning and generating fusion weights for three derived images,this network estimates the clear image effectively.Lastly,we embed the recurrent unit into the encoder-decoder network for constructing the recurrent network and optimize the dehazing result iteratively.Specifically,when we pass the output of the network back to input,the hidden status which is saved in the recurrent unit is also passed to the corresponding module in next iteration.Experiments demonstrate that our approach achieve superior performance on both synthetic and real images.Comparing with existing methods,the accuracy of our method has been improved 19%.Hence,it can be used in engineering practice.
作者
任敏敏
REN Min-min(Information Engineering Department,Engineering University of PAP,Xi'an 710086,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第7期1513-1518,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(51975462)资助。
关键词
图像融合
循环神经网络
编码解码
图像去雾
image fusion
recurrent neural network
encoder decoder
image dehazing
