Generating infrared image from visible image using Generative Adversarial Networks

Abstract

Infrared images have properties that are unaffected by illumination compared to visible images, object can be clearly recognized at day or night. Therefore, it is a better choice to use infrared images when training deep learning algorithms in unmanned systems. However, how to obtain a large number of infrared images makes it possible to train deep learning algorithm, which is the focus of our attention. In contrast to difficulties in direct shooting, computer simulations can generate large infrared images from visible images. But models that are trained on simulated data usually do not translate well to real scenarios. To bridge the domain gap between simulated and real infrared images, we exploit recent advances in paired image-to-image translation. We extend Pixel2Pixel to generate infrared image from visible image and propose a novel way of multilayer semantic information fusion, which significantly improves the quality of generated images. Finally, the image generated by our proposed PixelPro network is almost consistent with the real image distribution, achieved the purpose of expanding the dataset.