Joint Quality Assessment and Example-Guided Tone Mapping by Disentangling Picture Appearance From Content.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep learning significantly impacts low-level image processing tasks like style transfer, enhancement, and quality assessment.
• These tasks, though often separate, share a common goal: manipulating image appearance without altering underlying content.

Purpose of the Study:

• To develop a novel disentangled representation learning method separating image content and appearance.
• To utilize these learned features for a new image quality prediction model (DisQUE).
• To demonstrate the versatility of the learned features in various image processing applications.

Main Methods:

• A self-supervised disentangled representation learning approach to decompose images into content and appearance features.
• Development and training of the DisQUE model using the learned features for quality prediction.
• Evaluation of DisQUE across diverse quality prediction tasks and distortion types.

Main Results:

• DisQUE achieves state-of-the-art accuracy in image quality prediction.
• The learned disentangled features are effective for various image processing tasks, including HDR tone mapping.
• The model demonstrates the ability to tune output characteristics using example input-output pairs.

Conclusions:

• A novel disentangled representation learning method effectively separates image content and appearance.
• The learned features provide a powerful foundation for both image quality assessment and diverse image processing applications.
• This approach offers a unified framework for enhancing image appearance while preserving content integrity.