Multi-Attention Learning and Exposure Guidance Toward Ghost-Free High Dynamic Range Light Field Imaging

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Light field (LF) images often have low dynamic range and poor exposure due to sensor limitations.
• Existing high dynamic range (HDR) LF imaging methods struggle with ghosting artifacts and parallax distortion in dynamic scenes.

Purpose of the Study:

• To propose a novel ghost-free HDR LF imaging method for dynamic scenes.
• To enhance spatial-angular quality consistency in reconstructed HDR LF images.
• To improve the performance of downstream LF applications like depth estimation.

Main Methods:

• A multi-attention learning framework with exposure guidance is developed.
• Multi-scale cross-attention for efficient multi-exposure LF feature alignment.
• Dual self-attention Transformer blocks for geometric information extraction and feature fusion.
• Exposure masks and a local compensation module guide feature fusion and refine details.
• A multi-objective reconstruction strategy restores high-quality HDR LF images.

Main Results:

• The proposed method effectively suppresses ghosting artifacts in dynamic scenes.
• Achieves high spatial-angular quality consistency in HDR LF images.
• Outperforms state-of-the-art methods in quantitative and qualitative evaluations.
• Demonstrates improved performance for LF applications, including depth estimation.

Conclusions:

• The novel ghost-free HDR LF imaging method successfully addresses limitations of existing techniques.
• The multi-attention and exposure guidance approach yields superior image reconstruction quality.
• This method offers a robust solution for capturing high-quality HDR LF data in challenging dynamic scenarios.