Depth-Assisted Light Field Super-Resolution in Layered Object Space

The captured light field may fail to reconstruct fine details of the scene due to under-sampling problem of lightfield acquisition devices. Therefore,super-resolution is required to restore high-frequency information from the light field and to improve the quality of therendered views. Conventionalsuper-resolution algorithms are not ideal for light field data, as they do not utilize the full potential of light field 4D structure, while existing light fieldsuper-resolution algorithms rely heavily on the accuracy of the estimated depth and perform complex sub-pixeldisparity estimation. In this paper, we propose a new light field super-resolution algorithm which can address depthuncertainty with a layered object space. First, a pixel-wise depth estimation is performed from the resampled views.Then we divide the depth range into finite layers and back-project pixels onto these layers in order to address the sub-pixel depth error. Finally, two super-resolution schemes: in-depth warping and cross-depth learning, are introduced tosuper-resolve the views from light field data redundancy. The algorithms is tested with extensive datasets, and theresults show that our method attains favorable results in both visual assessment and objective metrics compared toother light field super-resolution methods