Color Held Illumination Map Estimation using GAN for Low-light Image Enhancement

Recently, museum learning has attracted considerable attention. Guide robots that detect exhibits are being used to enhance learning effects. However, detecting exhibits in the dark areas of the museum is difficult. Therefore, it is necessary to enhance low-light images. In this study, we present an image enhancement method, color held illumination map estimation (CHIME) using generative adversarial networks (GAN) (CHIMEGAN), which can be trained without paired data. Our method estimates a channel-wise illumination map for optimal image enhancement. Furthermore, CHIMEGAN can preserve the optimal color information of images by comparing the reflectance obtained from an elaborate illumination map. Experimental results show that our proposed method outperformed state-of-the-art methods and that it can be used to enhance low-light images of dark areas.