Avs3 Video Codec Technology

Block Partitioning: Sharpens Image Prediction Accuracy

• Block partitioning is the process of dividing a complex image into a plurality of rectangular blocks, and effectively encoding and decoding the image by block. In general, an object can be divided into multiple parts. Different parts are predicted using diverse partitioning techniques.

• including Quadtree plus Binary Tree (QTBT), Extended Quadtree (EQT), and Derived Tree (DT). The image can be flexibly divided into slices and coding tree units, according to specific scene. The asymmetric prediction unit in the DT mode effectively shortens the average prediction distance, greatly sharpening the prediction accuracy.

Intra-frame Prediction:
Enhances Prediction Efficiency

Intra-frame prediction is used to predict pixels on the basis of adjacent encoded pixels in the same frame, by utilizing the correlation in the video spatial domain. This effectively removes video temporal redundancy.

IPF

The intra prediction filter (IPF) re-filters the pixels at the boundaries of predicted blocks. It aims to optimize the boundaries of intra-frame blocks, and increase the spatial correlation between pixels, for improved prediction accuracy.

TSCPM

The two-step cross-component prediction (TSCPM) leverages the strong correlation between luminance and chrominance to remove linear redundancy between components.
Step 1: Perform linear scaling on a luma block to obtain the value of a corresponding chroma prediction block Pc′ (x,y). The coefficient α is related to four sample points adjacent to the luma block.
Step 2: Perform downsampling on the chroma prediction block to obtain a final value Pc.

Inter-frame Prediction: Increases Accuracy of Motion Vector Predictions

Inter-frame prediction is used to predict pixels on the basis of the encoded pixels in an adjacent image, by utilizing the correlation in the video time domain. This effectively removes video temporal redundancy.

AMVR

Adaptive motion vector resolution (AMVR) adaptively selects the motion vector resolution of 1/4 pixel, 1/2 pixel, 1 pixel, 2 pixels, and 4 pixels at the encoding end, according to the image attributes. Compared with the fixed resolution, AMVR reduces the data resulting from the encoding of the motion vector difference.

HMVP

History-based motion vector prediction (HMVP) adds an HMVP candidate list consisting of eight motion vectors, in addition to the spatial candidate list and temporal candidate list. The motion vectors are arranged in first input first output (FIFO) mode. Before a new candidate is added, the system checks whether the same MVP exists. HMVP adds a decoded block that is close but not adjacent to the current block from the candidate list, thereby increasing the efficiency of motion vector prediction.

UMVE

Ultimate motion vector expression (UMVE) fine-tunes the stride and direction of the basic motion vector in skip mode or direct mode, which conserves the bit rate and improves encoding quality. UMVE searches L0 and L1 in four directions and five distances, by using the position of basemv as the starting point.

Four-parameter AMC

Affine motion compensation (AMC) is classified into four-parameter and six-parameter AMC models, which encode motion vectors/motion vector differences of two or three control points respectively. The results of other points are obtained from calculations. AMC efficiently shows complex motions, such as scaling and rotation. The motion vector resolution is 1/16 pixel. Motion compensation based on 8 x 8 (unidirectional, bidirectional) or 4 x 4 (unidirectional) blocks is performed for coding units (CUs) greater than 16 x 16. A fixed size facilitates hardware implementation.