Improving the Visibility at Night for Advanced Driver Assistance Systems Using Night-to-Day Image Translation

Automobile manufacturers are targeting to increase the safety of drivers and passengers by incorporating different Advanced Driver Assistance Systems (ADAS). Most of these ADAS are vision-based and in order to operate them properly, these systems require a clear vision which is challenging to acquire during the night. Considering this limitation, the study presented explores the possibility of translating night-time images to clear and visible day-time images which can be used for ADAS instead of poor-quality night-time images. Even though there exist many deep-learning-based techniques to transform images between two domains, most of them highly depend on pixel-to-pixel paired datasets during training. It is challenging to develop such a dataset, particularly in dynamic roadside environments. Hence, this study proposes unsupervised deep learning with the popular Cycle-GAN model to cater the problem. Another challenging task is accessing the quality of the Cycle-GAN generated images. Since there do not exist pixel-to-pixel paired images, to compare the quality of the regenerated images, Blind Referenceless Image Spatial Quality Evaluator (BRISQUE), a referenceless image quality evaluation technique, is utilised to evaluate the performance of the model. The synthesized output of the trained Cycle-GAN indicated an average BRISQUE score of 28.0416, whereas that of the original day-time images was 26.2156. This exhibits that the Cycle-GAN was able to generate synthesised day-time images with unpaired night images with significant similarity to the actual day-time images. The source code along with the dataset of this study is publicly available at https://www.github.com/isurushanaka/Unpaired-N2D.