Infrared Image Generation Algorithm Based on GAN and contrastive learning

Abstract

For the task of converting dimly lit, low luminance nighttime visible to infrared images, we propose a Contrastive Visible-Infrared Image Translation Network (CVIIT). To better distinguish and translate objects such as pedestrians and vehicles, we introduce an attention module based on class activation map in the generator and discriminator of the Generative Adversarial Network (GAN), which captures richer context information in the images. In addition, we introduce contrastive learning to align the generated images with the visible images in terms of content. Qualitative and quantitative experiments on a publicly available visible-Infrared image pairing dataset (LLVIP) show that the proposed method generates infrared images of significantly higher quality than other state-of-the-art image-to-image translation (I2IT) methods.