Tone Image Classification and Weighted Learning for Visible and NIR Image Fusion

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Rule-based visible and near-infrared (NIR) image synthesis methods suffer from slow processing speeds.
• Existing learning-based methods often compromise image quality or visibility.

Purpose of the Study:

• To develop a fast and effective image fusion method for synthesizing visible and NIR images.
• To improve upon the processing speed of traditional rule-based methods using convolutional neural networks (CNNs).

Main Methods:

• A novel image fusion method utilizing DenseFuse, a CNN-based architecture, is proposed.
• The method incorporates a raster scan algorithm for dataset preparation and a classification technique based on luminance and variance.
• Feature map synthesis within a specific fusion layer is investigated and compared to other fusion layer approaches.

Main Results:

• The proposed DenseFuse method achieves superior image quality and visibility compared to existing learning-based methods.
• The synthesized images retain the high visual quality characteristic of rule-based methods.
• Processing time is reduced by at least three times compared to the rule-based image synthesis method.

Conclusions:

• The DenseFuse-based image fusion method offers a significant improvement in processing speed for visible and NIR image synthesis.
• This approach successfully balances enhanced speed with high-quality image synthesis, outperforming existing learning-based techniques.