CycleGAN-Based Image Translation for Near-Infrared Camera-Trap Image Recognition

Abstract

Due to its invisibility, NIR (Near-infrared) flash has been widely used to capture the images of wild animals in the night. Although the animals can be captured without notice, the gray NIR images are short of color and texture information and thus is difficult to analyze, for both human and machine. In this paper, we propose to use CycleGAN (Generative Adversarial Networks) to translate NIR image to the incandescent domain for visual quality enhancement. Example translations show that both color and texture can be well recovered by the proposed CycleGAN model. The recognition performance of a SSD based detector on the translated incandescent images is also significantly better than that on the original NIR images. Taking Wildebeest and Zebra for example, an increase of 16 % and 8 % in recognition accuracy has been observed.