Deep Guided Learning for Fast Multi-Exposure Image Fusion.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Multi-exposure image fusion (MEF) is crucial for enhancing image quality by combining information from multiple exposures.
• Existing MEF methods often struggle with computational efficiency and handling arbitrary image resolutions.
• The perceptual quality of fused images is a key metric for evaluating MEF techniques.

Purpose of the Study:

• To develop a fast and effective multi-exposure image fusion (MEF) method for static image sequences.
• To improve the visual quality of fused images compared to state-of-the-art methods.
• To achieve significant speed enhancements in the MEF process.

Main Methods:

• A fully convolutional network processes low-resolution input sequences for weight map prediction.
• Guided filtering is employed for joint upsampling of predicted weight maps.
• End-to-end training optimizes the perceptually calibrated MEF structural similarity (MEF-SSIM) index.

Main Results:

• MEF-Net demonstrates consistent improvement in visual quality across a majority of test sequences.
• The proposed method achieves a speedup of 10 to 1000 times compared to existing state-of-the-art MEF techniques.
• The method is applicable to static image sequences with arbitrary spatial resolutions and exposure numbers.

Conclusions:

• MEF-Net offers a computationally efficient and high-quality solution for multi-exposure image fusion.
• The end-to-end optimization using MEF-SSIM leads to superior perceptual quality.
• The public availability of the code facilitates further research and application in image fusion.