A Two-To-One Deep Learning General Framework for Image Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image fusion algorithms are crucial for enhancing scene descriptions in computer vision.
• Existing multi-modal image fusion methods face challenges in generalization.
• Convolutional neural networks (CNNs) have shown success but require specialized architectures.

Purpose of the Study:

• To propose a general image fusion framework addressing the generalization challenges in multi-modal image fusion.
• To develop an improved CNN-based approach for robust and stable image fusion.
• To enhance the performance of image fusion for both human recognition and automatic detection systems.

Main Methods:

• A novel framework utilizing multiple feature extraction layers to capture input image information.
• Feature maps are stacked channel-wise to create a feature fusion map.
• Skip connections and convolution filtering are employed for high-dimensional feature map reconstruction.

Main Results:

• The proposed model demonstrates generality and stability across different datasets.
• Achieved superior subjective visualization and objective evaluation compared to existing methods.
• The average running time is at least 94% faster than reference neural network algorithms.

Conclusions:

• The developed CNN-based framework offers a general and efficient solution for multi-modal image fusion.
• The approach significantly improves fusion quality and processing speed.
• This work advances the application of deep learning in computer vision for enhanced image analysis.