Unsupervised Infrared and Visible Image Fusion with Pixel Self-attention

In this paper, we propose a convolutional neural network (CNN) based unsupervised infrared and visible image fusion method. The proposed method optimizes both network structure and loss functions to obtain better fused images. Specifically, an effective pixel self-attention module is applied to emphasize the importance of different pixel locations of the feature map, which enables the network to better integrate the salient information in infrared images and the detail information in visible images. As to the loss function, we adopt the perceptual loss and texture loss to preserve the detail information as well as improve the visual perception of the fused image. Experimental results demonstrate that our method can achieve a superior performance compared with other fusion methods in both subjective and objective assessments.