In the past few years, license plate recognition systems have been widely used in parking lots. In order to identify license plates easily, traditional license plate recognition systems used in the parking lot have a fixed light source and a shooting angle. For particularly tilting angles, such as license plate images taken with super wide-Angle lenses or fisheye lenses, the deformation of the license plate can be particularly severe, resulting in poor recognition of traditional license plate recognition systems. In this paper, we propose a three-stage license plate recognition system based on Mask R-CNN that can be used for various shooting angles and more oblique images. Experimental results show that the proposed architecture can identify license plates with bevel angles over 0∼60 degrees and achieve mAP rates of up to 91%. Compared with the approach using YOLOv2 model, the proposed method with Mask R-CNN has made significant progress in identifying characters that are inclined above 45 degrees. The experimental results also show that the proposed method is superior to other methods in the open Taiwan license plate data set (called AOLP data set).