PS-GAN: Pseudo Supervised Generative Adversarial Network With Single Scale Retinex Embedding for Infrared and Visible Image Fusion

Abstract

Currently, ground truth fusion image, fused image contrast, and naturalness are rarely considered in existing infrared and visible image fusion (IVF) methods. In this article, we proposed a pseudosupervised generative adversarial network (GAN) with single scale retinex (SSR) embedding for IVF. First, a pseudoground truth fusion image conception and its computation method was proposed to solve ground truth fusion image shortage problem. Second, a novel SSR module embedded residual GAN was designed to improve fusion image contrast and naturalness. Finally, a special dense and mixed modal inputting strategy was also proposed for better modal mixed feature extraction. Extensive experimental results on public IVF datasets verified the superior performance of our proposed approach over other representative methods. It was demonstrated that the fused image details, contrast, and naturalness were significantly improved.