S2LIC: Learned image compression with the SwinV2 block, Adaptive Channel-wise and Global-inter attention Context

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Deep learning significantly improves image compression, but effective entropy models are crucial for latent representation probability estimation.
• Current entropy models often overlook multi-dimensional correlations, focusing mainly on 1D channel and spatial information.

Purpose of the Study:

• To propose an Adaptive Channel-wise and Global-inter attention Context (ACGC) entropy model for advanced image compression.
• To enhance rate-distortion performance and processing speed in learned image compression.

Main Methods:

• The ACGC model aggregates features in inter-slice and intra-slice contexts using a parallel checkerboard approach.
• Deformable attention is employed for dynamic weight refinement in global-inter slices context.
• A Residual SwinV2 Transformer and dense block network are integrated for global feature capture and nonlinear representation enhancement.

Main Results:

• The proposed method achieves encoding and decoding speeds of 0.31s and 0.38s, respectively.
• Outperforms VTM-17.1 and recent learned methods in PSNR metrics.
• Achieves BD-Rate reductions of 8.87%, 10.15%, and 7.48% on Kodak, Tecnick, and CLIC Pro datasets.

Conclusions:

• The ACGC entropy model offers superior rate-distortion performance and efficiency for learned image compression.
• The integration of attention mechanisms and advanced transformer models contributes to significant performance gains.