Wide-Band Color Imagery Restoration for RGB-NIR Single Sensor Images

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Multi-spectral RGB-Near-Infrared (NIR) sensors capture visible and NIR light simultaneously.
• NIR bands (700-1100 nm) can interfere with visible bands (400-700 nm), compromising RGB color accuracy.
• Existing methods for color correction are often insufficient.

Purpose of the Study:

• To develop and evaluate deep learning architectures for restoring accurate RGB color images.
• To remove near-infrared (NIR) spectral information that contaminates visible bands.
• To improve the state-of-the-art in color image restoration for multi-spectral imagery.

Main Methods:

• Proposed two novel deep learning architectures based on convolutional neural networks (CNNs).
• These architectures directly restore high-resolution RGB images from compromised multi-spectral data.
• Evaluated the models using diverse sets of outdoor images.

Main Results:

• Both proposed deep learning architectures achieved similar, high performance levels.
• The models demonstrated effectiveness across various scenarios and similarity metrics.
• The developed methods significantly outperformed existing state-of-the-art approaches.

Conclusions:

• Deep learning offers a powerful solution for correcting RGB color distortion in multi-spectral imaging.
• The proposed CNN-based architectures successfully restore accurate RGB color by mitigating NIR crosstalk.
• These findings advance the field of image processing for applications using RGB-NIR sensors.