Deep Coupled Feedback Network for Joint Exposure Fusion and Image Super-Resolution

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Photographic images often exhibit low dynamic range (LDR) and low resolution (LR), deviating from natural scenes due to camera limitations.
• Multi-exposure image fusion (MEF) and image super-resolution (SR) are common techniques, but typically addressed independently.

Purpose of the Study:

• To propose a unified deep learning approach for simultaneous multi-exposure image fusion and super-resolution.
• To develop a Coupled Feedback Network (CF-Net) capable of generating high dynamic range (HDR) and high-resolution (HR) images from LDR inputs.

Main Methods:

• A deep Coupled Feedback Network (CF-Net) architecture is proposed, comprising two coupled recursive sub-networks.
• Each sub-network utilizes feature extraction blocks (FEB), super-resolution blocks (SRB), and coupled feedback blocks (CFB) for progressive refinement.
• The network takes a pair of low-resolution, multi-exposure LDR images as input to generate a high-resolution HDR image.

Main Results:

• The CF-Net successfully achieves simultaneous MEF and SR, producing images with both enhanced dynamic range and resolution.
• Experimental results demonstrate that the CF-Net significantly outperforms state-of-the-art methods in SR accuracy and fusion performance.
• The proposed method effectively refines the fused high-resolution HDR image through a series of coupled feedback blocks.

Conclusions:

• The CF-Net offers a novel and effective solution for simultaneously addressing low dynamic range and low resolution in images.
• This unified approach surpasses existing independent methods for image fusion and super-resolution.
• The developed network provides a significant advancement in generating natural-looking, high-quality images from limited-quality inputs.