Learning-Based Just-Noticeable-Quantization- Distortion Modeling for Perceptual Video Coding.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Conventional video coding faces limitations in efficiency due to high computational complexity.
• Perceptual video coding (PVC) offers an alternative by minimizing perceptual redundancy using just-noticeable-distortion (JND).
• Existing JND models are inadequate for energy-reduced distortions.

Purpose of the Study:

• To propose a novel energy-reduced JND (ERJND) model suitable for JND-based PVC.
• To extend ERJND into learning-based just-noticeable-quantization-distortion (JNQD) models for preprocessing video encoders.
• To enable automatic adjustment of JND levels based on quantization step sizes.

Main Methods:

• Developed a discrete cosine transform-based ERJND model.
• Extended ERJND to two learning-based JNQD models: linear regression (LR-JNQD) and convolutional neural network (CNN-JNQD).
• Applied LR-JNQD and CNN-JNQD as preprocessing for High Efficiency Video Coding (HEVC).

Main Results:

• LR-JNQD achieved up to 38.51% (10.38% average) bitrate reduction.
• CNN-JNQD achieved up to 67.88% (24.91% average) bitrate reduction.
• Both models demonstrated minimal subjective video quality degradation compared to unprocessed input.

Conclusions:

• The proposed ERJND and JNQD models effectively enhance video coding efficiency.
• Learning-based JNQD models offer significant bitrate savings for video encoding.
• This work presents the first approach for automatically adjusting JND levels based on quantization step sizes for video encoder preprocessing.