Deep Unrolled Low-Rank Tensor Completion for High Dynamic Range Imaging.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• High dynamic range (HDR) imaging for dynamic scenes faces challenges with ghosting artifacts due to motion or poor exposures.
• Deep learning methods offer high synthesis performance but lack interpretability and are data-dependent.
• Traditional model-based approaches are theoretically sound but underperform compared to learning-based algorithms.

Purpose of the Study:

• To develop a ghost-free HDR image synthesis algorithm that leverages the strengths of both model-based and deep learning approaches.
• To overcome the limitations of existing methods, including interpretability issues and sensitivity to training data diversity.

Main Methods:

• Formulating ghost-free HDR synthesis as a low-rank tensor completion problem using low dynamic range (LDR) images.
• Developing two regularization functions to address modeling inaccuracies by extracting hidden model information.
• Unrolling an iterative optimization algorithm into deep neural networks, updating variables with closed-form solutions and regularizers with learned networks.

Main Results:

• The proposed algorithm demonstrates superior ghost-free HDR image synthesis performance compared to state-of-the-art methods.
• The approach exhibits enhanced robustness across various datasets.
• Significantly fewer training samples are required compared to existing deep learning techniques.

Conclusions:

• The algorithm unrolling approach effectively combines model-based rigor with deep learning flexibility for HDR imaging.
• This method offers a robust and efficient solution for ghost-free HDR synthesis, particularly in challenging dynamic scenes.
• The technique shows promise for reducing reliance on extensive training data in deep learning for image synthesis.