License Plate Detection And Recognition Based On And-Or Graph

In this paper, we propose a model based on And-Or Graph for license plate detection and recognition under various application environments. The traditional methods contain three stages: license plate detection, character segmentation and recognition. The relationship between the three stages is linear therefore the latter stage only depends on the former stage and do not provide feedback. In order to deal with the chanllenges exists is this area such as various illuminations and viewpoints, we unify the three stages together and be able to give a global solution for the position and label of the license plate.We model the license plate with a two-layers And-Or Graph. The and-nodes indicate the different configuration of license plate, and or-nodes indicate the selection between configurations and the terminal-nodes represent the HOG templates. There are spatial constraints between or-nodes and can let all the characters move together in a subspace of affine transformation. The latent-structural SVM is used to describe our model and CCCP procedure and cutting plane method are used for training. We propose a coarse-to-fine cascade algorithm for detection. First the coarse plate template and character template are applied to locate the license plate and find the viewpoint roughly, and then we use the precise character templates to determine the positions and labels of characters.In the experiments, four datasets are collected to evaluate the performance and speed of our algorithm. The overall precisions in the datasets Highway and Crossroad achieve higher than 99%, and in the dataset Bridge with extreme tilt viewpoint can achieve 92.9%, and in the dataset Night can achieve 98.5%. The results demonstrate that our algorithm is robust for various viewpoint to some extent and the performance is almost unaffected by illumination. The detection time is 0.4s in an 800×600 image, and satisfies most requirements in actual applications.