Ghost-Free Deep High-Dynamic-Range Imaging Using Focus Pixels for Complex Motion Scenes.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Multi-exposure image fusion often results in ghost artifacts due to registration inaccuracies.
• Existing deep learning methods for HDR fusion struggle with misaligned inputs, leading to artifacts in saturated regions.

Purpose of the Study:

• To develop a deep learning technique for seamless fusion of multi-exposed low dynamic range (LDR) images into high dynamic range (HDR) images.
• To leverage focus-pixel sensor data for artifact reduction and improved HDR image reconstruction.

Main Methods:

• Utilized focus-pixel sensor's inherently aligned and less saturated L/R luminance images.
• Developed separate sub-networks for luminance and chrominance fusion.
• Employed joint learning in a luminance recovery network fusing focus-pixel and overexposed LDR images.
• Applied a chrominance network to fuse color components for the final HDR image.

Main Results:

• The proposed method successfully reconstructs missing luminance using aligned, unsaturated focus pixel data.
• Achieved accurate color fusion guided by the recovered luminance.
• Demonstrated restoration of fine details in saturated areas.
• Produced high-quality, ghost-free HDR images without requiring pre-alignment.

Conclusions:

• The novel deep learning approach effectively mitigates ghost artifacts in multi-exposure image fusion.
• Focus-pixel sensor data offers significant advantages for HDR imaging, particularly in handling saturation and alignment.
• The method provides a robust solution for generating artifact-free HDR images from LDR inputs.