Images of road scenes in low-light situations are lack of details which could increase crash risk of connected autonomous vehicles (CAVs). Therefore, an effective and efficient image enhancement model for low-light images is necessary for safe CAV driving. Though some efforts have been made, image enhancement still cannot be well addressed especially in extremely low light situations (e.g., in rural areas at night without street light). To address this problem, we developed a light enhancement net (LE-net) based on the convolutional neural network. Firstly, we proposed a generation pipeline to transform daytime images to low-light images, and then used them to construct image pairs for model development. Our proposed LE-net was then trained and validated on the generated low-light images. Finally, we examined the effectiveness of our LE-net in real night situations at various low-light levels. Results showed that our LE-net was superior to the compared models, both qualitatively and quantitatively.