Infrared Image Generation From RGB Images Using CycleGAN

Thermal imaging is more robust than optical imaging against the illumination related issues. Therefore, it is preferred or utilized with RGB data in some of the essential problems such as surveillance, environmental monitoring and so on. Deep learning has been used in various fields including the problems in the scope of thermal imaging and proved its ability to solve lots of problems. However, due to the requirement of large datasets in deep learning and the lack of public thermal data because of the constrains of thermal imaging, deep learning can not be used much in thermal imaging. In this study, we mainly investigate whether using CycleGAN on paired images rather than impaired ones increases the success rate and the effect of our electromagnetic spectrum based normalization approach. Evaluations on public data sets show that our approach has potential to increase the success rate of CycleGAN, but further study is required.