Reducing Complexity of HEVC: A Deep Learning Approach.

Area of Science:

• Video Compression Technologies
• Machine Learning for Signal Processing
• Digital Image and Video Analysis

Background:

• High Efficiency Video Coding (HEVC) offers reduced bit-rates compared to H.264 but suffers from high encoding complexity.
• The quad-tree partitioning of coding units (CUs) in HEVC contributes significantly to this complexity due to brute-force rate-distortion optimization (RDO).

Purpose of the Study:

• To propose a deep learning approach for predicting CU partitions in HEVC to reduce encoding complexity.
• To address complexity in both intra- and inter-prediction modes.

Main Methods:

• Development of a large-scale database for HEVC CU partition data.
• Representation of CU partitions as hierarchical CU partition maps (HCPM).
• Proposal of an early-terminated hierarchical CNN (ETH-CNN) for intra-mode prediction and an early-terminated hierarchical LSTM (ETH-LSTM) combined with ETH-CNN for inter-mode prediction.

Main Results:

• The proposed ETH-CNN effectively reduces HEVC intra-mode encoding complexity by replacing brute-force search.
• The combined ETH-LSTM and ETH-CNN approach significantly reduces HEVC inter-mode encoding complexity.
• Experimental results demonstrate superior performance over state-of-the-art methods in complexity reduction for both modes.

Conclusions:

• Deep learning, specifically CNN and LSTM networks, provides an effective solution for reducing HEVC encoding complexity.
• The proposed hierarchical prediction methods (ETH-CNN and ETH-LSTM) offer a promising direction for efficient video coding.