A dual-branch feature decomposition fusion network with data augmentation for infrared and visible image fusion

Image fusion is a vision enhancement technology that aims to combine images obtained from different types of sensors to generate a fused image rich in scene information. Addressing the challenge of insufficient extraction of modality-specific information in traditional infrared and visible image fusion methods, this paper proposes a fusion algorithm based on multi-dimensional feature decomposition. By integrating wavelet transformation, this algorithm employs feature decomposition at both pixel and frequency dimensions. The modules then extract and reuse modality information across various domains and scales. In the feature fusion layer, the introduced affine transformation and KAN are designed to enhance the quality of feature fusion, while the residual dilated convolution block is utilized to capture wavelet high-frequency information. Additionally, the algorithm employs random intensity-based augmentation to improve model representation capabilities. Experimental results on several public datasets demonstrate that our algorithm exhibits superior fusion performance compared to seven state-of-the-art methods.