Learning a Single Tucker Decomposition Network for Lossy Image Compression with Multiple Bits-Per-Pixel Rates.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Lossy image compression (LIC) is crucial for reducing image file sizes.
• Deep Convolutional Neural Networks (CNNs) have shown promise in LIC.
• Current CNN-based LIC methods require separate networks for each compression rate (bits-per-pixel, bpp).

Purpose of the Study:

• To develop a single CNN capable of performing LIC at multiple bpp rates.
• To overcome the limitations of the "one-network-per-bpp" approach in existing CNN-based LIC methods.
• To enhance the generality and flexibility of CNNs for practical LIC applications.

Main Methods:

• Proposed a Tucker Decomposition Network (TDNet) for LIC.
• Introduced a novel Tucker Decomposition Layer (TDL) to decompose latent image representations.
• Implemented an iterative non-uniform quantization scheme and a coarse-to-fine training strategy.

Main Results:

• TDNet successfully performs LIC at multiple bpp rates within a single network.
• The rank of the core tensor and its quantization allow for easy adjustment of the bpp rate.
• Achieved state-of-the-art compression performance, measured by PSNR and MS-SSIM.

Conclusions:

• TDNet offers a flexible and efficient solution for multi-rate lossy image compression.
• The proposed method demonstrates superior performance over existing CNN-based LIC techniques.
• This approach enhances the practical applicability of deep learning in image compression.