ES-CycleGAN: An Improved CycleGAN for VI-to-IR Translation

Infrared image has been used in many fields due to its strong environmental adaptability. But it has little data volume because of its prohibitive cost. In order to compensate for the deficiency of the infrared image dataset, CycleGAN is improved to be named ES-CycleGAN, which is used for translation of VI (visible images) to IR (infrared images). Converting asymmetric translation task to symmetric translation task by using visible single channel information instead of VI. At the same time, the image edge loss is calculated by the Canny operator, and the style loss is calculated by perceptual loss to ensure the content and style of the image. To enhance the translation of infrared images, extended cyclic consistency loss is used to replace the original cyclic consistency loss. In the Flir dataset, we implemented it and evaluated the results using PSNR and SSIM methods. The implementation results show that it effectively increases the quality of infrared image generation and better retains the image structure of the original data.