Continuous and infinite video zoom using hierarchical and lorentzian autoencoders

Abstract

A system and method for compressing and restoring data using hierarchical autoencoders and Lorentzian autoencoders for video processing. For general data, the system employs hierarchical autoencoders operating at multiple scales. For video data, Lorentzian autoencoders preserve three-dimensional tensor structure where spatial and temporal relationships remain intact throughout compression and decompression. A correlation network, trained on cross-correlated data sets, enhances restoration by leveraging relationships between compressed representations, recovering information lost during compression. The Lorentzian approach enables advanced video features including temporal prediction and infinite zoom, where users can examine regions beyond original resolution with synthesized yet plausible details. This architecture achieves higher compression ratios while maintaining or improving data quality, particularly for video content where spatiotemporal coherence is critical, applicable to surveillance, medical imaging, entertainment, and remote sensing.