关键词: 无监督学习, 残差密集网络, 生成对抗网络, 红外图像, 可见光图像, 图像融合

Abstract: Most of the known infrared and visible image fusion models based on convolutional neural networks make little use of the hierarchical features from visible images， thus resulting in insufficient texture details of the fused image. Inspired by the residual network and dense network， an image fusion algorithm is proposed based on unsupervised deep learning to solve the problem of insufficient texture information of fused images. The residual dense block has a continuous storage mechanism to retain the feature information of each layer to the maximum extent. The design of local residual fusion and global residual fusion is conducive to learning the structural texture in the image. In addition， to better preserve the detailed texture in visible images， the generative adversarial network is introduced to perform unsupervised learning on the dataset. Subjective and objective experiments show that the proposed algorithm achieves not only a good visual fusion effect， but also more edge texture information of the fused image. Compared with that of the existing state-of-the-art algorithms， the objective evaluation index of the method proposed is also greatly improved.

Key words: unsupervised learning, residual dense network, generative adversarial network, infrared image, visible image, image fusion