DeepDuoHDR: A Low Complexity Two Exposure Algorithm for HDR Deghosting on Mobile Devices.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Growing demand for consumer-grade high dynamic range (HDR) imaging.
• Existing HDR deghosting algorithms often lack speed, memory efficiency, or robustness.
• Difficulty in achieving a balance between performance and computational cost.

Purpose of the Study:

• To develop a fast, memory-efficient, and robust HDR deghosting algorithm.
• To leverage neural networks and a conservative deghosting strategy for improved HDR image generation.
• To create an algorithm adaptable to various computational constraints, including mobile platforms.

Main Methods:

• Utilized attention mechanisms and U-Net-based neural representations.
• Employed a conservative deghosting strategy for image fusion.
• Generated HDR images by prioritizing well-exposed regions from high-exposure inputs and fusing aligned data otherwise.

Main Results:

• Achieved state-of-the-art performance in HDR deghosting across challenging scenarios.
• Demonstrated significantly lower computational complexity compared to existing methods.
• Validated performance using both visual and quantitative evaluations.
• Showcased effectiveness using only two bracketed exposures.

Conclusions:

• The proposed algorithm offers a robust and efficient solution for HDR deghosting.
• Its low computational requirements make it ideal for deployment on resource-constrained devices like smartphones.
• The method successfully balances image quality with computational efficiency, addressing a key challenge in HDR imaging.