稀疏光场的密集重建方法

摘要

Light field cameras usually sample sparsely in angular domain due to limited sensor resolution.To mitigate the problem, we proposed a method based on deep learning which can reconstruct 8×8 or even denser light fields from a 2×2 sparse light field.The method first uses CNN to estimate the depth information of the novel view from input views, and then obtain a synthesized image for the novel view by warping the input images using the estimated depth map.Finally, a high-quality synthesized image is generated by CNN with Resnet blocks.Experimental results show that our approach greatly improves the performance compared with the state-of-the-art methods, especially for the edges and the regions with rich textures.%针对光场技术中存在的角分辨率不足的问题,以稀疏光场为研究对象,提出了一种使用深度学习进行光场密集重建的方法,利用卷积神经网络从输入视图估计出新视角的深度信息,并融合输入图像颜色信息以及新视角的深度信息获得新视角的合成图;通过带有残差模块的卷积神经网络对新视角合成图进行高质量的修复重建,得到高质量的新视角合成图像.实验结果表明,所提方法较目前最新的方法有很大改进,尤其在图像的轮廓边缘、纹理丰富的地方重建得更好;可以从2×2的稀疏光场重建8×8甚至更加密集的光场.