Adaptive Quantization based on an activity mask.

The human eye is more tolerant towards errors in areas of high activity and is quick to find out errors in areas of lower activity. To leverage this psychovisual characteristic, quantization can be made adaptive based on the activity mask. Activity masking will be implemented in two phases:

  1. Biasing the RDO based on the activity at a specific region.
  2. Varying the quantizer offsets across segments based on activity.

Optimizing the quantization algorithm using Trellis Quantization:

Using Trellis Quantization passing the activity measurements as the weights to the trellis. The output of this Activity masked Trellis is used for quantization. This helps the quantization perform better in PSNR metrics while increasing perceptual image quality significantly. This is a feature that has proved to perform better in the case of x264 and is expected to yield similar results at rav1e too.

Organization

Student

Shreevari SP

Mentors

  • Nathan Egge
close

2019