Learned Fast HEVC Intra Coding.

Area of Science:

• Video Compression
• Machine Learning for Video Coding
• Digital Signal Processing

Background:

• High Efficiency Video Coding (HEVC) offers excellent rate-distortion performance through flexible coding unit (CU) partitioning and numerous intra-prediction modes.
• This performance comes at the cost of substantially increased computational complexity.

Purpose of the Study:

• To develop a configurable framework for fast HEVC intra coding (LFHI) that optimizes the tradeoff between coding performance and computational complexity.
• To address the high computational demands of HEVC intra coding.

Main Methods:

• A low-complexity asymmetric-kernel CNN (AK-CNN) was designed for efficient local directional texture feature extraction.
• The minimum number of RDO candidates (MNRC) concept was integrated for faster intra-mode selection using AK-CNN predictions.
• An evolution optimized threshold decision (EOTD) scheme was implemented for configurable complexity-efficiency tradeoffs.
• An interpolation-based prediction scheme was proposed to generalize the framework across all quantization parameters (QPs).

Main Results:

• The LFHI framework achieves high parallelism and a superior complexity-efficiency tradeoff compared to existing methods.
• Demonstrated up to 75.2% reduction in intra-mode encoding complexity.
• Negligible degradation in rate-distortion performance was observed.

Conclusions:

• The proposed LFHI framework effectively reduces HEVC intra coding complexity.
• It offers a configurable and superior tradeoff between encoding efficiency and computational cost.
• The framework generalizes well across different quantization parameters.