融合3D对极平面图像的光场角度超分辨重建  

Light-field angular super-resolution reconstruction via fusing 3D epipolar plane images

作　　者：陈纯毅[1] 范晓辉 胡小娟[1] 于海洋[1] CHEN Chunyi;FAN Xiaohui;HU Xiaojuan;YU Haiyang(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《光学精密工程》2023年第21期3167-3177,共11页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.U19A2063)。

摘　　要：针对光场成像中因硬件限制而造成的光场图像角度分辨率低的问题,提出一种融合3D对极平面图像的光场角度超分辨重建方法。该方法首先将输入图像按不同的视差方向排列分别进行特征提取,以充分利用输入图像的视差信息,提高深度估计的准确性。利用深度图将输入图像映射到新视角位置,生成初始合成光场。为了使重建光场图像能够保持更好的细节信息及几何一致性,先通过水平3D对极平面图像融合重建分支和垂直3D对极平面图像融合重建分支,分别对初始合成光场进行水平融合重建和垂直融合重建,再将两个结果进行混合重建,生成最终的高角度分辨率光场图像。实验结果表明:相比于现有方法,本文方法在合成光场数据集和真实光场数据集上的重建效果均得到了提高,峰值信噪比的提升幅度最高达1.99%,有效地提高了重建光场的质量。Light-field(LF)imaging can capture spatial and angular information of light rays in a scene.Compared to traditional 2D/3D images,LF images provide a more comprehensive description of the scene.To address the problem of low angular resolution in LF images caused by hardware constraints,a LF angular super-resolution reconstruction method based on the fusion of 3D epipolar plane images(EPIs)is proposed.First,to make full use of the parallax information of the input images and improve the accura-cy of depth estimation,the input images are arranged in varying parallax directions,and their features are extracted individually.Then,initial synthetic LF images are generated by transforming the input images to match the new viewpoint location using the corresponding depth maps.Finally,to ensure that the recon-structed LF image retains better detail information and geometric consistency,the LF is reconstructed hori-zontally and vertically via the horizontal and vertical 3D EPI fusion reconstruction branches,respectively.These two reconstruction results are then fused to produce the final high-angular-resolution LF image.Ex-perimental results show that,compared to existing methods,the proposed method achieves an improved reconstruction quality across both synthetic and real-world LF image datasets,and the maximum increase in the peak signal-to-noise ratio reaches 1.99%.Thus,the proposed method can effectively improve the quality of the reconstructed LF.

关 键 词：光场 超分辨重建 3D对极平面图像 卷积神经网络 

分 类 号：O439[机械工程—光学工程] TP391.41[理学—光学]