Learning to measure infrared properties of street views from visible images

With the rapid development of infrared cameras, infrared images which have the unique advantage of all-weather detection ability are more and more incorporated into practical applications. However, it is hard to develop infrared-domain artificial intelligence with the shortage of labeled infrared data. Traditional infrared image simulation techniques face the problem of strict parameters, slow solution speed, and limited scenarios. In this paper, we tackle the problem of simulating infrared image proprieties with generative networks via reference to visible images. We propose a novel infrared image generation model that complied with the infrared imaging principle, the V2IR-GAN, consisting of spontaneous emission, reflected radiation and transmission coefficient modules. Extensive experiments and analyses are conducted to demonstrate the high stimulative performance of the V2IR-GAN, showing that it achieves the best SSIM of 88.2% against the state-of-the-art generative models. We also provide a specific application to validate the practicability of the V2IR-GAN.