AVS Intra Prediction Algorithm: Implementation and Optimization Techniques

The AVS intra prediction algorithm serves as a core technology in video coding standards, primarily designed to predict current coding blocks for spatial redundancy reduction. This algorithm analyzes information from adjacent encoded blocks to generate predictive values for the current block, thereby enhancing coding efficiency. Common prediction modes include Horizontal Prediction, Vertical Prediction, DC (Mean) Prediction, Diagonal Down-Left Prediction, and Diagonal Down-Right Prediction.

Horizontal Prediction Horizontal prediction utilizes reference pixel values from the left side of the current block, executing prediction along the horizontal direction. Each row of pixels in the current block is filled using values from the left reference pixels. This mode is suitable for images containing vertical edges or textures. In MATLAB implementation, this typically involves creating a matrix where each row replicates the left reference pixel values using repmat() or similar broadcasting functions.

Vertical Prediction Vertical prediction operates based on reference pixel values from the top, performing prediction along the vertical direction. Each column of pixels in the current block adopts values from the top reference pixels. This mode applies well to images with horizontal edges or textures. Code implementation generally requires transposing the reference pixel array and assigning it column-wise to the prediction block matrix.

DC Prediction DC prediction calculates the average value of reference pixels and fills the entire prediction block with this mean value. This mode works effectively for flat regions, significantly reducing prediction errors. The MATLAB implementation involves computing the mean using mean() function and generating a constant matrix of the same dimensions as the prediction block.

Diagonal Down-Left Prediction Diagonal Down-Left prediction performs interpolation along the diagonal direction (from top-right to bottom-left), suitable for images with diagonal textures. This mode utilizes reference pixels from the top and right sides, employing interpolation calculations to derive predictive values. Implementation typically requires weighted averaging of reference pixels using bilinear interpolation algorithms with specific angle coefficients.

Diagonal Down-Right Prediction Diagonal Down-Right prediction operates along the opposite diagonal direction (from top-left to bottom-right), also effective for diagonal textures. This mode bases its predictions on left and top reference pixels through interpolation to generate results. The coding approach involves similar interpolation techniques but with modified weighting factors corresponding to the different diagonal orientation.

In MATLAB implementation, the process begins with extracting reference pixels for the current block, followed by calculating predictive values according to different prediction modes. The final step compares these predictions with the actual block to determine the optimal prediction mode. Through optimized prediction mode selection—often implemented using rate-distortion optimization (RDO) algorithms—AVS significantly improves compression efficiency in video encoding. Key functions typically include pixel extraction routines, mode-specific prediction calculators, and SAD (Sum of Absolute Differences) or SATD (Sum of Absolute Transformed Differences) comparators for mode decision.