Image Fusion via Domain and Feature Transfer

Abstract

The infrared and visible images fusion technique is designed to simultaneously preserve the thermal radiation information of the infrared image and the detailed texture information of the visible image. Therefore, the fused image with rich detail information and a clear target area in order to find it more efficiently. Inspired by the powerful GANs technology in recent years, we proposed a novel method for fusing infrared and visible images that based on domain and feature transfer. First, it is designated as an optimization problem with a latent encoding which can be mapped into a pixel intensity consistent image on the latent image. We employ generative adversarial network, which can capture content characteristics of one image dataset and figure out how these characteristics could be translated into the domain of another image dataset, to transfer the appearance of an image from visible image to infrared one. And then we use the feature consistent constrains to enhance the features of the fused image. Moreover, by adding gradient constraints to preserve the details of visible image, therefore we combine the fused image with a similar gradient to the visible image. Qualitative and quantitative comparisons on public datasets demonstrate our proposed method is superior to the state-of-the-art technology. The fused image we have generated is more like a high-resolution enhanced infrared image, which is more efficient for discovering the target and tracking.