Speeding up VP9 Intra Encoder with Hierarchical Deep Learning Based Partition Prediction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Video Compression

Background:

• VP9 video codec uses computationally intensive rate-distortion optimization (RDO) for superblock partitioning.
• The combinatorial search space in RDO leads to significant encoding time.

Purpose of the Study:

• To develop a deep learning framework for predicting intra-mode superblock partitions in VP9 video encoding.
• To accelerate the encoding process by reducing the reliance on RDO.

Main Methods:

• A hierarchical fully convolutional network (H-FCN) was designed to predict four-level partition trees.
• A large dataset of VP9 superblocks and partitions was created for training the H-FCN model.
• The trained H-FCN model was integrated into the VP9 encoder.

Main Results:

• The proposed H-FCN approach achieved an average intra-mode encoding speedup of 69.7%.
• This speedup came at the cost of a 1.71% increase in Bjøntegaard-Delta bitrate (BD-rate).
• The model outperformed the fastest built-in speed level of the reference VP9 encoder for good quality intra encoding.

Conclusions:

• Deep learning offers an effective method to accelerate VP9 video encoding.
• The H-FCN framework provides a significant speed improvement with a small trade-off in compression efficiency.
• This approach enhances VP9 encoding performance, particularly for intra-mode compression.