We proposed an improved CycleGAN framework for translating short-wavelength infrared facial images and visible-light facial images. Based on the CycleGAN framework, a loss function calculation path was added and a new loss function was designed. A dataset was established, and the model parameters were adjusted based on experiments to improve the translation effect of the proposed model on the facial images. It effectively overcame the differences in images caused by different spectral characteristics so that the images could be easily recognized. The experimental verification was performed with a self-built dataset. The subjective evaluation, FID(Fréchet inception distance), and recognition accuracy were used to compare the proposed framework with several other frameworks. The results show that the improvement of the proposed framework is obvious and the structural features of the original target are better maintained, which effectively improves the observability and recognition accuracy of image translation results.