基于视点图像与EPI特征融合的光场超分辨率

光场（Light Field， LF）成像能同时捕获场景中光线的空间信息和角度信息，应用广泛。然而，它的分辨率受到成像设备硬件以及空间和角度分辨率之间制衡的限制。过低的空间分辨率严重影响了光场图像的质量及其应用。因此，本文充分利用光场特性增强图像细节，提出一种基于视点图像（Viewpoint Image， VI）和极平面图像（Epipolar Plane Image， EPI）特征融合的端到端光场超分辨率方法，能够同时超分辨率所有视点图像。本方法将低分辨率光场图像按照水平和垂直EPI方向堆叠排列，利用三维视点图像堆栈中包含EPI信息的特点，采用双分支结构的3D递减卷积网络处理输入的四维光场数据。这样能够同时对视点图像和EPI信息进行特征提取和融合，充分探索光场的纹理信息及几何一致性。在真实和合成光场数据集上的实验结果均表明，该方法相比现有主流方法不仅在客观指标上具有更好的表现，主观质量上也能保持更好的几何一致性，同时还具有更少的模型参数和更快的执行速度。 

Light Field Super-Resolution Based on Viewpoint Image and EPI Feature Fusion

? ?Light field （LF） has a wide range of applications since it can capture spatial and angular information of light rays simultaneously. However， the resolution is limited by the hardware of the imaging device and the trade-off between spatial and angular resolution. The low spatial resolution greatly affects the quality and applications of LF images. In this paper， by making full use of the characteristics of LF to enhance the image details， we proposed an end-to-end LF super-resolution method based on the feature fusion of viewpoint image （VI） and epipolar plane image （EPI）， which was capable of super-resolving all VIs at the same time. Specifically， in this method， the low-resolution LF images were stacked and arranged in the horizontal and vertical EPI directions. With the property that EPI information was contained in 3D VI stacks， a 3D decreasing convolutional network with dual-branch structure was used to process the input 4D LF data， which was able to extract and fuse the features of the VI and EPI information simultaneously and fully explore the texture information and geometric consistency of the LF. Experimental results on both real-world and synthetic LF datasets showed that the proposed method not only has better visual performance， but also maintains better geometric consistency in subjective quality than other state-of-the-art methods. It also had fewer model parameters and faster execution speed.