MFAGAN: A multiscale feature-attention generative adversarial network for infrared and visible image fusion

Infrared images contain salient target information and visible images contain texture information. The fusion of infrared and visible images makes images express better visual understanding. Further improvement in this area is obtained by the application of the Generative Adversarial Network (GAN). But some fusion results based on GAN will lose details from the source images. Besides, sometimes the fusion results are inclined to certain source images. To force the fusion results to retain more source features and balance the information from source images, a novel GAN called multiscale feature-attention generative adversarial network (MFAGAN) is proposed. First, the infrared and visible source images are decomposed into images at different scales. Then, the multiscale images are encoded and fused at the corresponding scale. Finally, the decoder generates fused images. The game between the generator and discriminator can make the information distribution of the fusion results more reasonable, but we further propose a new generator loss function called feature-attention loss. Feature-attention loss creates a criterion that measures the similarity of high-dimensional features between fused images and source images at each scale. Extensive experiments performed on two commonly used datasets show that MFAGAN obtains good results and has some superiority over other existing methods.