Realistic Nighttime Haze Image Generation with Glow Effect
Yading Zheng, Aizhong Mi, Yingxu Qiao, Yijiang Wang · 2022
Abstract: Haze rendering aims to generate realistic nighttime haze images from clear images. The results can be applied to various practical applications, such as nighttime image dehazing algorithms, game scene rendering, shooting filters, etc. We investigate two smaller but challenging problems in nighttime haze rendering, namely 1) how to accurately estimate the transmission map and air light from clear and haze images, respectively, in the absence of paired datasets? 2) How to render the characteristics of a realistic nighttime haze image: glow effect? For this purpose, we propose an unsupervised nighttime haze image rendering method called NHRM (Nighttime Haze Rendering Module). Precisely, the NHRM consists of three modules: 1) an illumination retention module based on Retinex theory. 2) a transmission map estimation network using multi-scale feature fusion and attention mechanism. 3) a glow generation module using tone mapping and convolution bloom technique. Compared with existing haze generation methods, NHRM can render realistic and controllable nighttime haze images by simulating the main features of nighttime haze scenes in an unsupervised manner. Numerous experimental results show that our method outperforms existing rendering methods with better visual effects and quantitative metrics.