Abstract
One embodiment provides a computer-implemented technique that predicts, for each input image, a codec operating point expected to meet a target perceptual quality metric with respect to a reference image. A machine-learned model is trained using labels derived from a sparse sweep of operating points per training image, followed by interpolation of a rate-quality curve to identify a crossing point for the target quality. During serving, features are extracted from the source image using one of multiple selectable feature tiers, and the trained model outputs a predicted operating point. A calibrated safety margin is applied asymmetrically in the direction of increased quality based on a residual distribution, and the resulting operating point is clamped to codec limits before encoding the image, thereby reducing wasted bits for easy images while maintaining a quality constraint for most images without an encode-measure search loop.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Machine-Learned Per-Image Codec Operating-Point Prediction with Calibrated Quality Margin", Technical Disclosure Commons, (October 28, 2026)
https://www.tdcommons.org/dpubs_series/12522