基于联合卷积自编码网络的多聚焦图像融合方法  

Multi-focus image fusion method based on joint convolution auto-encoder network

作　　者：罗晓清[1] 熊梦渔 张战成 LUO Xiao-qing;XIONG Meng-yu;ZHANG Zhan-cheng(School of Internet of Things,Jiangnan University,Wuxi 214122,China;School of Electronic and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]苏州科技大学电子与信息工程学院,江苏苏州215009

出　　处：《控制与决策》2020年第7期1651-1658,共8页Control and Decision

基　　金：国家自然科学基金项目(61772237);中央高校基本科研业务费专项资金项目(JUSRP51618B);苏州市重点产业技术创新项目(SYG201702)。

摘　　要：受相机景深的限制,单次成像无法对不同景深的内容全部清晰成像.多聚焦图像融合技术可以将不同聚焦层次的图像融合为一幅全聚焦的图像,其中如何得到准确的聚焦映射是多聚焦图像融合中的关键问题.对此,利用卷积神经网络强大的特征提取能力,设计具有公共分支和私有分支的联合卷积自编码网络以学习多源图像的特征,公共分支学习多幅图像之间的公共特征,每幅图像的私有分支学习该图像区别于其他图像的私有特征.基于私有特征计算图像的活动测度,得到图像聚焦区域映射,据此设计融合规则以融合两幅多聚焦图像,最终得到全聚焦的融合图像.在公开数据集上的对比实验结果显示:主观评测上,所提出的方法能够较好地融合聚焦区域,视觉效果自然清晰;客观指标上,该方法在多个评价指标上优于对比方法.Due to the limitation of focus depth of cameras,single focus imaging cannot capture clear scene for objects on different distances.Multi-focused image fusion technology can integrate images with different focus levels into a single focused image.How to get accurate focus map is a key point in a fusion process.Based on the powerful feature learning ability of convolution neural networks,a joint convolution auto-encoder network is proposed,which consists of common branches and private branches.With this hybrid structure,common branches tend to learn common features among multi-focused image pairs,and each private branch does for private features which distinguish an image from other images.Private features are used to calculate focus activity measurement,then focus map can be obtained.Guided by this map,a fusion rule is designed to fuse multi-focused images into a full focused image.Results on benchmark datasets demonstrate that compared with the state-of-art methods,the proposed method achieves more natural and clear focus effect.In terms of objective evaluation metrics,the proposed method also wins the first rank on most metrics.

关 键 词：联合卷积自编码 多聚焦图像 图像融合 聚焦映射 卷积神经网络 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]