Efficient light field compression and reconstruction via view selection and directional view synthesis

Light field (LF) imaging captures both the intensity and direction of incoming light rays, enabling advanced applications such as refocusing, viewpoint generation, immersive visualization, virtual reality, and three-dimensional scene representation. However, the large amount of data generated by LF content, which is significantly greater than that of conventional images and videos, creates substantial challenges for storage, compression, transmission, and bandwidth utilization. Efficient methods for reducing LF data volume while preserving reconstruction quality are therefore essential for the widespread adoption of LF technology. This thesis investigates techniques for reducing the storage and transmission bandwidth requirements of LF content through view selection and view synthesis. Several LF view selection strategies are evaluated within an MV-HEVC-based compression framework. By selectively encoding, storing, and transmitting only a subset of LF views while reconstructing the remaining views at the receiver or decoder, the overall data volume can be significantly reduced. Experimental results demonstrate that the raster skip structure provides the most favorable balance between compression efficiency and reconstruction quality compared to alternative view arrangements. Objective evaluations further show that substantial storage savings and bandwidth reductions can be achieved while maintaining perceptually acceptable visual quality. Building upon these findings, this thesis investigates the impact of directional relationships on LF view synthesis using a Mamba-based architecture. Horizontal, vertical, and diagonal view correlations are independently analyzed to determine their influence on reconstruction performance. Direction-specific datasets are generated from LF content and used to train dedicated synthesis models. The resulting analysis demonstrates the importance of exploiting redundancies across all angular directions and the effectiveness of Mamba-based models in accurately reconstructing omitted LF views from a reduced set of encoded views, thereby further reducing storage requirements and transmission bandwidth while preserving visual fidelity. Overall, this thesis advances the state of the art in LF compression, storage optimization, and bandwidth-efficient transmission. The proposed view selection and synthesis techniques significantly reduce the amount of LF data that must be stored and transmitted while maintaining high reconstruction quality.