DCFNet: Infrared and Visible Image Fusion Network Based on Discrete Wavelet Transform and Convolutional Neural...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Current image fusion algorithms struggle with missing details, blurred target information, and poor visual quality.
• Infrared and visible-light image fusion is crucial for applications requiring comprehensive scene understanding.

Purpose of the Study:

• To develop an advanced infrared and visible-light image fusion algorithm addressing limitations of existing methods.
• To improve the clarity of target information and overall visual quality in fused images.

Main Methods:

• Proposed algorithm integrates discrete wavelet transform (DWT) and convolutional neural networks (CNNs) within an autoencoder backbone.
• DWT and inverse DWT (IDWT) layers optimize frequency-domain feature extraction and reconstruction.
• Incorporated bottleneck residual blocks and a coordinate attention mechanism to enhance feature characterization.
• Employed an l1-norm fusion strategy and a weighted loss function (pixel, gradient, structural loss) for network optimization.

Main Results:

• The proposed algorithm effectively fuses infrared and visible-light images, producing clearer results with enhanced scene information.
• Experimental evaluations on public datasets demonstrate superior performance in both subjective and objective metrics.
• Generalization experiments confirm the network's robust ability to adapt to diverse image data.

Conclusions:

• The developed DWT-CNN based fusion algorithm significantly improves image fusion quality by preserving detailed information and enhancing target visibility.
• The method offers a visually natural and harmonious fused image, validating its effectiveness and generalization capability.