Fast and High-Performance Learned Image Compression With Improved Checkerboard Context Model, Deformable Residual...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep learning-based image compression shows promise but suffers from slow serial processing and high network complexity.
• Existing methods often struggle to balance rate-distortion performance with computational efficiency.

Purpose of the Study:

• To develop novel techniques for deep learning image compression that enhance speed and reduce complexity without compromising performance.
• To address the limitations of serial context-adaptive entropy models and high computational demands in current learned image codecs.

Main Methods:

• Introduced a deformable residual module for improved redundancy removal.
• Designed an improved checkerboard context model enabling parallel decoding.
• Implemented a three-pass knowledge distillation scheme to reduce decoder complexity.
• Applied L1 regularization for sparse latent representations, encoding only non-zero channels.

Main Results:

• Achieved encoding speeds approximately 20x faster and decoding speeds 70-90x faster than state-of-the-art learned methods.
• Demonstrated a 2.3% improvement in rate-distortion performance.
• Outperformed classical codecs like H.266/VVC-intra and recent learned methods on standard datasets (Kodak, Tecnick-40) using PSNR and MS-SSIM.

Conclusions:

• The proposed techniques effectively balance rate-distortion performance and computational complexity in deep learning image compression.
• The method offers significant speedups in both encoding and decoding, making it suitable for practical applications.
• This approach sets a new benchmark for efficient and high-performance learned image coding.