Generating pencil drawings from natural images is one of the popular areas within the numerous applications of computer technology. In recent years, end-to-end image translation using generative models like GAN have become more common with the development of deep neural network technology. In practical pencil drawing, there are often some special drawing methods used, such as abstraction and stroke style. Abstraction involves drawing only the salient parts with detail and purposefully omitting others to convey the artwork's theme. Altering the stroke's fineness, direction and density can accentuate details within the drawing. We suggest that by introducing these drawing methods to deep learning model, it is possible to generate more realistic pencil-style images than existing models. Hence, we propose StrokeGAN as a novel deep learning model based on CycleGAN, which simulates the abstraction techniques in pencil drawings by introducing saliency maps, while providing the user with local control of stroke styles. Results of user study show that the proposed algorithm outperforms the existing methods in terms of quality image and usability.