Hierarchical Predictive Coding-Based JND Estimation for Image Compression.

Area of Science:

• Computer Vision
• Neuroscience
• Image Processing

Background:

• The human visual system (HVS) processes information hierarchically.
• Hierarchical predictive coding theory offers a framework for understanding visual perception stages.
• Existing methods for estimating just noticeable distortion (JND) can be improved.

Purpose of the Study:

• To model the HVS as a three-level communication system.
• To propose a novel JND estimation scheme based on hierarchical predictive coding.
• To integrate the JND model into the Versatile Video Coding (VVC) codec for image compression.

Main Methods:

• Modeled the HVS using a three-level communication system based on hierarchical predictive coding.
• Developed a surprise-based measurement for positive and negative perception effects using residue self-information and local Shannon entropy.
• Estimated JND thresholds for each stage individually and combined them using non-linear superposition.

Main Results:

• The proposed JND estimation scheme was successfully incorporated into the VVC codec.
• Experimental results demonstrated superior performance compared to existing JND models.
• Achieved over 16% bit rate reduction without compromising perceptual quality.

Conclusions:

• The novel surprise-based JND estimation scheme effectively models visual perception.
• The proposed method offers significant improvements in image compression efficiency.
• This approach enhances the performance of video codecs like VVC.