Abstract

In this thesis, different aspects of video and light field reconstruction are considered such as super-resolution, inpainting, segmentation and codecs. For this purpose, each of these strategies are analyzed based on a specific goal and a specific database. Accordingly, databases which are relevant to film industry, sport videos, light fields and hyperspectral videos are used for the sake of improvement.

This thesis is constructed around six related manuscripts, in which several approaches are proposed for multi-frame reconstruction. Initially, a novel multi-frame reconstruction strategy is proposed for lightfield super-resolution in which graph-based regularization is applied along with edge preserving filtering for improving the spatio-angular quality of lightfield. Second, a novel video reconstruction is proposed which is built based on compressive sensing (CS), Gaussian mixture models (GMM) and sparse 3D transform-domain block matching. The motivation of the proposed technique is the improvement in visual quality performance of the video frames and decreasing the reconstruction error in comparison with the former video reconstruction methods. In the next approach, student-t mixture models and edge preserving filtering are applied for the purpose of video super-resolution. Student-t mixture model has a heavy tail which makes it robust and suitable as a video frame patch prior and rich in terms of log likelihood for information retrieval. In another approach, a hyperspectral video database is considered, and a Bayesian dictionary learning process is used for hyperspectral video super-resolution. To that end, Beta process is used in Bayesian dictionary learning and a sparse coding is generated regarding the hyperspectral video super-resolution. The spatial super-resolution is followed by a spectral video restoration strategy, and the whole process leveraged two different dictionary learnings, in which the first one is trained for spatial super-resolution and the second one is trained for the spectral restoration.

Furthermore, in another approach, a novel framework is proposed for replacing advertisement contents in soccer videos in an automatic way by using deep learning strategies. For this purpose, a UNET architecture is applied (an image segmentation convolutional neural network technique) for content segmentation and detection. Subsequently, after reconstructing the segmented content in the video frames (considering the apparent loss in detection), the unwanted content is replaced by new one using a homography mapping procedure.

In addition, in another research work, a novel video compression framework is presented using autoencoder networks that encode and decode videos by using less chroma information than luma information. For this purpose, instead of converting Y'CbCr 4:2:2/4:2:0 videos to and from RGB 4:4:4, the video is kept in Y'CbCr 4:2:2/4:2:0 and merged the luma and chroma channels after the luma is downsampled to match the chroma size. An inverse function is performed for the decoder. The performance of these models is evaluated by using CPSNR, MS-SSIM, and VMAF metrics. The experiments reveal that, as compared to video compression involving conversion to and from RGB 4:4:4, the proposed method increases the video quality by about 5.5% for Y'CbCr 4:2:2 and 8.3% for Y'CbCr 4:2:0 while reducing the amount of computation by nearly 37% for Y'CbCr 4:2:2 and 40% for Y'CbCr 4:2:0.

The thread that ties these approaches together is reconstruction of the video and light field frames based on different aspects of problems such as having loss of information, blur in the frames, existing noise after reconstruction, existing unpleasant content, excessive size of information and high computational overhead. In three of the proposed approaches, we have used Plug-and-Play ADMM model for the first time regarding reconstruction of videos and light fields in order to address both information retrieval in the frames and tackling noise/blur at the same time. In two of the proposed models, we applied sparse dictionary learning to reduce the data dimension and demonstrate them as an efficient linear combination of basis frame patches. Two of the proposed approaches are developed in collaboration with industry, in which deep learning frameworks are used to handle large set of features and to learn high-level features from the data.