QNet: An Adaptive Quantization Table Generator Based on Convolutional Neural Network.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• JPEG compression relies heavily on quantization tables for performance.
• Optimizing these tables is crucial for efficient lossy image compression.

Purpose of the Study:

• To develop an image-adaptive quantization table generation method using Convolutional Neural Networks (CNNs).
• To improve JPEG compression performance and computational efficiency through adaptive techniques.

Main Methods:

• A dataset of over 10,000 images was created, with optimal quantization tables generated using a genetic algorithm.
• A CNN-based regression network was trained to generate adaptive quantization tables by fusing frequency and spatial domain information.
• A CNN-based classification network was trained to select optimal quantization tables for enhanced performance.

Main Results:

• The proposed CNN methods significantly outperform state-of-the-art techniques.
• Average Peak Signal-to-Noise Ratio (PSNR) gains of 1.2 dB (regression) and 1.4 dB (classification) were achieved at 1.0 bpp.
• Structural Similarity Index Measurement (SSIM) improvements of 0.4% and 0.54% were observed, respectively.
• Competitive computational efficiency was demonstrated, with processing times of 15 ms (regression) and 6.25 ms (classification) per image.

Conclusions:

• CNN-based adaptive quantization table generation offers superior JPEG compression performance.
• The proposed methods provide significant gains in image quality and compression efficiency.
• The approach is computationally efficient, making it practical for real-world applications.