A JND-Based Pixel-Domain Algorithm and Hardware Architecture for Perceptual Image Coding.

Area of Science:

• Computer Vision
• Image Processing
• Hardware Architecture

Background:

• Traditional image coding methods often struggle to balance compression efficiency with perceptual quality.
• Existing just-noticeable difference (JND) models can be computationally intensive, limiting their real-time hardware implementation.
• Perceptual coding requires accurate estimation of visual redundancies for optimal performance.

Purpose of the Study:

• To develop a hardware-efficient pixel-domain just-noticeable difference (JND) model.
• To integrate the JND model into a low-complexity perceptual image coding architecture.
• To achieve high visual quality and improved rate-distortion performance in image compression.

Main Methods:

• Implementation of a pixel-domain JND model and its hardware architecture on an FPGA.
• Development of a perceptual image coding architecture using adaptive downsampling and predictive coding.
• Identification of regions-of-interest (ROIs) based on JND visibility thresholds for adaptive downsampling.

Main Results:

• The proposed JND model demonstrates improved accuracy in estimating visual redundancies compared to classic models.
• Experimental results show enhanced rate-distortion performance and visual quality over JPEG-LS.
• Compression experiments indicate reduced bit rates compared to JPEG 2000 at similar peak signal-to-perceptible-noise ratios (PSPNR).

Conclusions:

• The developed hardware-efficient JND model and perceptual encoder offer a viable solution for high-quality, low-complexity image compression.
• The adaptive downsampling strategy effectively preserves visual quality by considering JND thresholds.
• FPGA implementation results confirm moderate hardware requirements and high throughput (over 100 Megapixel/s).