Online dictionary learning for sparse coding
ICML '09 Proceedings of the 26th Annual International Conference on Machine Learning
| Hi-index | 0.00 |
Light field cameras, e.g. [Lytro 2012; Veeraraghavan et al. 2007], have ushered a new direction in photography allowing consumers to synthesize photographs with novel viewpoints or varying focus after the actual recording. Unfortunately, current light field camera designs impose a fixed trade-off between spatial and angular resolution --- spatial resolution is reduced to capture angular light variation on the sensor. We introduce a principled computational framework and a new camera design to acquire and reconstruct light fields at full spatial and angular resolution from a single exposure. Our framework introduces a high-dimensional sparse basis for light fields learned from millions of light fields patches. The same optimization procedure also allows for the synthesis of optimal mask patterns that are mounted at a slight offset in front of the sensor and optically attenuate the light field before it is recorded by the sensor. Finally, a weighted compressive sensing-style reconstruction is performed to recover the light field. We demonstrate, in theory and with simulations, how our compressive approach to light field photography outperforms state-of-art techniques.