Ghost-Free HDR Imaging in Dynamic Scenes via High-Low-Frequency Decomposition.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• High-dynamic-range (HDR) image generation in dynamic scenes is challenging.
• Transformers show promise for motion handling but struggle with high-frequency details and complexity.
• Existing methods often compromise structural detail recovery.

Purpose of the Study:

• To develop a ghost-free HDR reconstruction network for dynamic scenes.
• To address the limitations of Transformers and CNNs in capturing high-frequency details and managing computational complexity.
• To improve the quality of HDR images by effectively processing high- and low-frequency components.

Main Methods:

• Proposed HL-HDR network, decomposing features into high- and low-frequency components.
• Introduced Frequency Alignment Module (FAM) for motion capture and detail refinement.
• Utilized Frequency Decomposition Processing Block (FDPB) to fuse frequency information for precise reconstruction.

Main Results:

• HL-HDR effectively overcomes limitations of existing Transformer and CNN-based methods.
• The network demonstrates superior performance in capturing large-scale motion and refining local details.
• Achieved state-of-the-art results on five public HDR datasets.

Conclusions:

• HL-HDR offers an effective solution for ghost-free HDR reconstruction in dynamic scenes.
• The proposed frequency decomposition approach enhances detail recovery and reduces computational complexity.
• HL-HDR represents a significant advancement in HDR imaging technology for dynamic environments.