Learned Image Compression Using Cross-Component Attention Mechanism.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Existing learned image compression methods primarily target RGB format, limiting their applicability to YUV420.
• YUV420 format presents unique challenges due to component variances, requiring specialized compression strategies.

Purpose of the Study:

• To propose an efficient and accurate image compression framework specifically for the YUV420 format.
• To leverage cross-component attention mechanisms for improved information preservation and compression performance.

Main Methods:

• Developed an information-guided compression framework featuring a dual-branch advanced information-preserving module (AIPM) with an information-guided unit (IGU) and feature attention block (FAB).
• Integrated an adaptive cross-channel enhancement module (ACEM) to utilize inter-component correlations for detail reconstruction, using the Y component for UV guidance.
• Introduced a quantization scheme for context models that avoids retraining and mitigates cross-platform decoding errors.

Main Results:

• The proposed framework achieves state-of-the-art performance in YUV420 image compression.
• Demonstrated an average BD-rate reduction of 8.37% compared to Versatile Video Coding (VVC) on common test conditions (CTC) sequences.
• The quantization scheme effectively addresses floating-point-related decoding errors, enabling cross-platform neural codec applications.

Conclusions:

• The novel information-guided framework offers superior image compression for YUV420 format.
• The cross-component attention mechanism and adaptive enhancement module significantly improve compression efficiency and detail preservation.
• The proposed quantization strategy facilitates the practical deployment of neural image codecs across diverse platforms.