Learned Image Compression With Gaussian-Laplacian-Logistic Mixture Model and Concatenated Residual Modules

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep learning methods increasingly surpass traditional image compression techniques.
• Current learned image compression models use single entropy models, which is suboptimal for diverse image content.
• Existing methods lack adaptability to variations within and across images.

Purpose of the Study:

• To develop a more flexible and accurate entropy model for latent representations in learned image compression.
• To enhance the network architecture for improved learning capabilities in image compression.
• To achieve state-of-the-art compression performance exceeding current standards.

Main Methods:

• Proposing a discretized Gaussian-Laplacian-Logistic Mixture Model (GLLMM) for adaptive entropy modeling.
• Introducing Concatenated Residual Blocks (CRB) with enhanced shortcut connections for network architecture.
• Evaluating the proposed scheme on standard datasets (Kodak, Tecnick-100, Tecnick-40).

Main Results:

• The GLLMM-CRB scheme demonstrates superior performance compared to leading learning-based methods and VVC intra coding.
• The model achieves better results in both Peak Signal-to-Noise Ratio (PSNR) and Multi-Scale Structural Similarity (MS-SSIM) metrics.
• The proposed approach offers improved accuracy and efficiency for image compression.

Conclusions:

• The developed GLLMM and CRB significantly advance learned image compression.
• The proposed method provides a more adaptable and efficient solution for compressing diverse image content.
• This work sets a new benchmark in image compression performance.