Adaptive Spatial Resampling for Machine Vision Encoding

The patent changes the compression parameters from human-perception-based metrics to machine-vision-based metrics. Specifically, it uses spatial information features (such as edge density, texture complexity, and frequency content) as the basis for compression, rather than traditional human visual system models. This parameter change enables the compression algorithm to preserve features that are critical for machine vision tasks like object detection and recognition, while still achieving efficient compression.

The patent implements dynamic spatial resampling by computing spatial information features for each video block or region and adapting the resampling strategy accordingly. Regions with high spatial information (complex textures, edges, or details) are resampled at lower rates, while regions with low spatial information (smooth areas) are resampled at higher rates. This dynamic adaptation ensures that compression efficiency is maintained across diverse video content with varying spatial complexities, preventing performance degradation on any particular type of video.