基于深度引导与自学习的高动态成像算法  

Method for High Dynamic Range Imaging Based on Deep Guided and Self-learning

作　　者：张俊超 杨飞帆 时伟[1] 陈溅来 赵党军[1] 杨德贵[1] ZHANG Junchao;YANG Feifan;SHI Wei;CHEN Jianlai;ZHAO Dangjun;YANG Degui(School of Aeronautics and Astronautics,Central South University,Changsha 410083,China;Research Institute for Frontier Science,Beihang University,Beijing 100191,China)

机构地区：[1]中南大学航空航天学院,长沙410083 [2]北京航空航天大学前沿科学技术创新研究院,北京100191

出　　处：《电子与信息学报》2023年第1期291-299,共9页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(62105372,61901531);国防科技重点实验室基金(6142401200301);湖南省自然科学基金(2021JJ40794,2021JJ40781)。

摘　　要：多曝光图像融合是将同一场景不同曝光度的图像进行融合,是当前高动态场景成像的主流方法。为了获得更自然的融合效果,该文提出基于深度引导与自学习的多曝光图像融合网络(MEF-Net)。该网络是以端到端的方式融合任意数量的不同曝光度图像,无监督地输出最优的融合结果。在损失函数方面,通过引入强度保真约束项和加权的多曝光图像融合结构相似度(MEF-SSIM),提升融合效果。此外,针对两幅极度曝光情况下的图像融合,该文采用自学习的方式,基于预训练的模型进行参数微调与优化,减弱光晕现象。基于大量测试数据,实验结果表明,该文所提算法在定量指标和视觉融合效果方面均优于现有主流算法。Multi-exposure image fusion aims to fuse a series of images with different exposures for the same scene, and it is the main-stream method for high dynamic range imaging. To obtain more realistic results, a Multi-Exposure image Fusion Network(MEF-Net) based on deep guided and self-learning is proposed. This network is designed to fuse any number of images with different exposures in an end-to-end way, and generate the best-fused results in an unsupervised way. In terms of the loss function, an intensity fidelity constraint term and the weighted Multi-Exposure image Fusion Structural SIMilarity(MEF-SSIM) are introduced to improve the fusion quality. Moreover, a self-learning method is adopted to fine-tune and optimize the pre-learned model,considering the fusion problem of two images under extreme exposure to mitigate the halo phenomenon generated by fusion. Based on abundant testing data, experimental results show that the proposed algorithm outperforms other mainstream methods in terms of both quantitative measurement and visual fused quality.

关 键 词：多曝光图像融合 高动态成像 强度保真约束 自学习 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]