Infrared and Visible Image Fusion Using Generative Adversarial Network with Feature Complement Block

This paper proposes a generative adversarial network with feature complement block for infrared and visible image fusion. In the generator network architecture, we adopt two different encoder paths to extract source image features. The two paths utilize the block to extract complementary gradient features from infrared images and complementary intensity features from visible images. It is the network that ensures rich features are extracted. Specifically, we propose a feature complement block which is able to exploit the shallow features from the source images and aggregate the complementary information. For the fusion rules, to gather more meaningful complementary features, we introduce feature consistency loss function to constrain the structural similarity between complementary feature maps and main feature maps. The results of qualitative and quantitative experiments demonstrate the superiority of our method over seven classic methods.