DNN-Based Lossy Compression System for Standard Format Compatibility

The patent segments the compression process into two distinct parts: a DNN-based importance assessment module that analyzes multi-dimensional data and determines region importance, and a separate standardized compression module (using AVC/HEVC/VVC) that performs the actual compression. This segmentation allows the system to leverage the high efficiency of DNN for importance evaluation while maintaining compatibility with standard formats through the separate encoding stage, thus resolving the contradiction between compression efficiency and format compatibility.

The patent introduces compression level information as an intermediary between the DNN importance assessment and the standardized compression process. This intermediary component translates the DNN's importance evaluation into compression parameters (such as quantization parameters, bitrate allocation, or quality factors) that can be applied within standard compression frameworks. By using this intermediary, the system achieves both DNN-driven efficiency and standard format compatibility without direct conflict between the two approaches.

The patent implements a dynamic configuration where the DNN model can be selectively retrained only when necessary (i.e., when important object definitions actually change). The system includes a configuration unit that manages model training status and allows flexible adjustment of training data and parameters. This dynamic approach enables the system to adapt to changing requirements while minimizing unnecessary retraining operations, thus balancing adaptability with time efficiency.

The patent applies local quality by allowing different regions or aspects of the DNN model to be updated independently based on changing requirements. Rather than requiring complete model retraining, the system can selectively retrain specific components or use transfer learning techniques to update only the portions of the model affected by changes in important object definitions. This localized update strategy reduces overall relearning time while maintaining necessary adaptability.

The patent applies local quality by using the DNN to assess importance at different spatial locations and applying different compression levels accordingly. Important regions (such as areas containing important objects) are preserved with higher quality and less compression, while less important regions are compressed more aggressively. This location-dependent quality control maintains overall image quality where needed while enabling faster compression of less critical areas, thus resolving the contradiction between image quality and compression speed.

The patent implements partial action by focusing computational resources on assessing and preserving only the important portions of the data rather than uniformly processing the entire high-resolution dataset. The DNN-based importance assessment identifies and prioritizes critical regions, allowing the compression process to concentrate on maintaining quality where it matters most while applying more aggressive compression elsewhere. This selective approach improves compression speed without sacrificing necessary image quality in important areas.