Deep Multi-Exposure Image Fusion for Dynamic Scenes.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Learning-based multi-exposure fusion (MEF) methods excel in static scenes but struggle with dynamic scenes, producing ghosting artifacts.
• Existing MEF methods lack datasets and solutions for dynamic scenes, hindering progress in this common scenario.

Purpose of the Study:

• To address the limitations of current MEF methods in dynamic scenes.
• To introduce the first benchmark dataset for multi-exposure fusion in dynamic scenes.
• To propose a novel deep learning framework for ghost-free image reconstruction from dynamic scenes.

Main Methods:

• Developed a 'static-for-dynamic' strategy to create a multi-exposure dataset of dynamic scenes with reference images.
• Proposed a deep dynamic MEF (DDMEF) framework utilizing pre-enhancement-based alignment and privilege-information-guided fusion.
• Implemented a privilege distillation scheme with an information attention transfer loss for enhanced deghosting.

Main Results:

• The proposed 'static-for-dynamic' strategy successfully generated a novel MEF dataset for dynamic scenes.
• The DDMEF framework effectively reconstructs high-quality, ghost-free images from two differently exposed images of dynamic scenes.
• Experimental results demonstrate that DDMEF outperforms state-of-the-art dynamic MEF methods in both qualitative and quantitative evaluations.

Conclusions:

• The creation of the first MEF dataset for dynamic scenes fills a critical gap in research.
• The proposed DDMEF framework offers a robust solution for ghost-free image reconstruction in dynamic environments.
• The study provides a valuable resource (dataset and code) for advancing research in dynamic multi-exposure fusion.