Transparent composite model for DCT coefficients: design and analysis.

Area of Science:

• Image processing and computer vision
• Statistical modeling
• Data compression

Background:

• Discrete Cosine Transform (DCT) coefficients are crucial for image compression.
• Heavy-tailed distributions are common in DCT coefficients, posing modeling challenges.
• Existing models may not fully capture the complex distribution of DCT coefficients.

Purpose of the Study:

• To propose a novel Transparent Composite Model (TCM) for DCT coefficients.
• To enhance modeling accuracy and achieve data reduction.
• To investigate the effectiveness of TCM for real-valued and discrete DCT coefficients.

Main Methods:

• Developed a TCM that separates DCT coefficient sequences into a heavy tail and main body.
• Modeled the tail with a uniform distribution and the main body with a parametric distribution.
• Employed maximum likelihood estimation for parameter estimation and proposed efficient online algorithms.

Main Results:

• For real-valued coefficients, TCM with truncated Laplacian showed the best accuracy-complexity tradeoff.
• For discrete coefficients, the geometric mixture TCM (GMTCM) outperformed Laplacian and generalized Gaussian models.
• GMTCM effectively identified outliers, enabling data reduction and feature extraction.

Conclusions:

• TCM provides accurate modeling of DCT coefficient distributions.
• GMTCM is a practical and effective model for discrete DCT coefficients in image/video applications.
• The data reduction capability of GMTCM aids in feature extraction and outlier image identification.