Deep HDR Imaging via A Non-local Network.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Reconstructing high dynamic range (HDR) images from multiple low dynamic range (LDR) inputs is challenging due to ghosting artifacts from object motion.
• Existing methods struggle with significant object motion, failing to adequately suppress ghosting artifacts.

Purpose of the Study:

• To present a novel deep framework, NHDRRnet, for effective ghosting artifact removal in HDR image reconstruction, particularly for large object motions.
• To exploit non-local correlations within LDR inputs to improve HDR image quality.

Main Methods:

• NHDRRnet utilizes a Unet architecture for fusing LDR inputs and mapping them to a deep feature space.
• A global non-local module reconstructs pixels by weighted averaging based on feature correspondences.
• A triple-pass residual module is incorporated to enhance local feature extraction.

Main Results:

• NHDRRnet demonstrates superior performance in suppressing ghosting artifacts compared to existing methods.
• The framework is particularly effective in scenarios with large object motions across LDR inputs.
• Experiments on benchmark datasets validate the efficacy of the proposed approach.

Conclusions:

• NHDRRnet successfully addresses the challenge of ghosting artifacts in HDR reconstruction, especially under significant motion.
• The method's ability to leverage non-local correlations and local features offers a robust solution for complex imaging conditions.