Synthesizing Large-Scale Datasets for License Plate Detection and Recognition in the Wild

License Plate Detection and Recognition (LPDR) plays a key role in modern intelligent transportation systems. Recent state-of-the-art methods of LPDR are based on deep convolutional neural networks (DCNN), which require learning a huge number of parameters from a sufficient amount of labeled training images. However, collecting and manually annotating a large collection of diverse license plate images is tedious and challenging. In this paper, we propose to use a virtual world simulator to automatically generate realistic images with precise annotations, effectively avoiding the need for laborious image acquisition and labeling. Our method can generate any type of license plates, and simulate different scene backgrounds, illumination conditions, and viewpoints. In order to validate the effectiveness of synthesized images, we have generated a large collection of images with three types of annotations, including bounding box, corner points and plate numbers. In many real scenarios, the image regions of plates are rarely axis-aligned, and thus the axis-aligned bounding box is inappropriate for describing plate positions. Motivated by this observation, we propose to detect the four corners of the plate, and present a DCNN based method for plate corner detection. We conduct experiments on three tasks, including bounding box detection, corner detection and plate recognition, using the synthesized dataset combined with a real dataset. Experimental results show that our simulated image datasets can improve the performance clearly on all of the three tasks.