With the rise of variety and available access to the super-high volume data in multiple dimensions, several applications in image and video processing can benefit from such kind of data with consideration of the storage and the computation speed. We focus on the research of super-high volume image and video processing on three specific applications: 4D light-field, super-resolution (SR), and 3D reconstruction.

We propose a synthetic 4D light-field data generation algorithm from a single RGB-D input image. The algorithm solves the problem of the hardware limit of the light-field camera in trade-offs of spatial and angular resolutions by simulating the imaging process of the light-field camera. The synthetic light-field data is validated to be feasible for light-field rendering w...[ Read more ]

With the rise of variety and available access to the super-high volume data in multiple dimensions, several applications in image and video processing can benefit from such kind of data with consideration of the storage and the computation speed. We focus on the research of super-high volume image and video processing on three specific applications: 4D light-field, super-resolution (SR), and 3D reconstruction.

We propose a synthetic 4D light-field data generation algorithm from a single RGB-D input image. The algorithm solves the problem of the hardware limit of the light-field camera in trade-offs of spatial and angular resolutions by simulating the imaging process of the light-field camera. The synthetic light-field data is validated to be feasible for light-field rendering with super-high spatial and angular resolution.

We propose to utilize the video data captured by a high-speed camera to do the super-resolution (SR) over the license plate part of a fast-moving vehicle. The high temporal resolution provided by the high-speed camera contributes a lot to the spatial resolution in our SR algorithm. The experimental results are evaluated subjectively and objectively by the recognition results and confidence index of the license plate recognition system, which shows a big performance improvement.

We propose to accelerate the pre-processing of images taken by cryo-electron microscopy (cryo-EM) and improve the Fourier-Slice-Projection-Theorem based 3D reconstruction by introducing the learning-based initial 3D model from the proposed simplified 3D Resnet. The reconstructed 3D structures have competitive resolutions as compared to the state-of-art cryo-EM 3D reconstruction algorithms.