Sensitivity Decouple Learning for Image Compression Artifacts Reduction.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning has advanced image compression artifact reduction.
• Current methods often neglect intrinsic image attributes, hindering downstream task performance.
• Decoupling intrinsic attributes is crucial for improved artifact reduction and feature representation.

Purpose of the Study:

• To propose a novel Dual Awareness Guidance Network (DAGN) for image compression artifact reduction.
• To decouple intrinsic image attributes into compression-insensitive and compression-sensitive features.
• To enhance the performance of downstream machine vision tasks by improving artifact reduction.

Main Methods:

• Utilized adversarial training to retain high-level semantics in encoded features.
• Developed a compression quality-aware feature encoder for compression-sensitive features.
• Introduced a cross-feature fusion module within DAGN to maintain feature consistency.

Main Results:

• Achieved an average 2.06 dB PSNR gain on the BSD500 dataset.
• Demonstrated superior performance compared to state-of-the-art methods.
• Processed images in 29.7 ms on BSD500, indicating high efficiency.

Conclusions:

• The proposed DAGN effectively reduces image compression artifacts.
• DAGN enhances both visual quality and performance in downstream machine vision tasks.
• The method shows significant efficiency and effectiveness across multiple datasets.