AV1 Chroma Compression via Cross-Component Correlation

This article provides an overview of how the AV1 video codec leverages cross-component correlation to compress chrominance (color) channels. By utilizing an advanced intra-prediction tool known as Chroma from Luma (CfL), AV1 models the relationship between brightness and color information to remove redundant data, resulting in higher visual fidelity at reduced bitrates.

Understanding Cross-Component Redundancy

In standard digital video, content is typically represented in the YCbCr color space. The luma channel (\(Y\)) contains the black-and-white structural details and edges, while the chroma channels (\(Cb\) and \(Cr\)) carry the color differences. In natural scenes, the color components generally share the same edge orientations, textures, and gradients as the brightness component.

Traditional video codecs traditionally predicted chroma blocks solely using neighboring reconstructed chroma samples. AV1 overcomes this limitation by recognizing that the reconstructed luma block already contains rich structural information that can directly predict the chroma block.

The Chroma from Luma (CfL) Mechanism

AV1 implements cross-component prediction through its Chroma from Luma (CfL) prediction mode. Because chroma channels are often subsampled (most commonly in 4:2:0 format, where chroma resolution is half the width and half the height of luma), the process follows several precise steps:

  1. Luma Downsampling: If the input uses chroma subsampling, the reconstructed luma pixels corresponding to the current block are downsampled to match the dimensions of the target chroma block.
  2. AC Component Extraction: AV1 computes the average value (DC component) of the downsampled luma block and subtracts it from each downsampled luma pixel. This isolates the high-frequency structural details (the AC component) from the baseline brightness.
  3. Linear Scaling: The zero-mean luma signal is scaled by a factor (\(\alpha\)) to approximate the variation in the chroma channel.
  4. DC Restoration: The intra-predicted chroma DC value (derived from adjacent chroma pixels) is added to the scaled luma signal to produce the final predicted chroma block.

The mathematical model for each chroma sample is expressed as:

\[\text{Chroma}_{\text{pred}} = \alpha \times (\text{Luma}_{\text{sub}} - \text{DC}_{\text{Luma}}) + \text{DC}_{\text{Chroma}}\]

Parameter Optimization and Signaling

Rather than computing the scaling factor \(\alpha\) implicitly at the decoder, the AV1 encoder evaluates the optimal \(\alpha\) parameters for both the \(Cb\) and \(Cr\) channels independently during rate-distortion optimization.

The encoder tests multiple values of \(\alpha\) to minimize prediction error and transmits the chosen parameter directly in the bitstream. AV1 uses an efficient signaling scheme with a limited set of discrete fractions and integer values (ranging from negative to positive values) to represent the scaling parameter, keeping the signaling overhead minimal.

Coding Efficiency Gains

By utilizing the already decoded luma channel as a reference, CfL eliminates the need to independently signal complex structural and directional information for the chroma channels. This significantly reduces the residual energy (the difference between the original and predicted pixels) that must be transformed and quantized, producing notable coding gains in textured and sharp-edged regions without increasing color artifacts.