Vehicle License Plate Detection Using Unigram Model and Difference-of-surf Bigram Model with Svm

One of the central issues in vehicle license recognition is designing a detector to localize license plates which are fast and accurate. We present a new approach for vehicle license plate detection using a fusion of unigram and bigram model with Speeded Up Robust Features (SURF). SURF descriptor is obtained from each keypoint to build the unigram model while difference-of-SURF for each doublet is obtained from a circular sub-window of the license plate region for building the bigram model. The unigram model determines if a particular keypoint is part of a license plate while the bigram model determines if a doublet is within a license plate region, within a nonlicense plate region or cross from background to license plate. Together, both models are learned using Support Vector Machine (SVM). During testing, a circle of initial radius r1 is scanned using a sliding windows-like approach and a binary classification based on the doublets obtained is performed for license plate detection. Experiments show that by incorporating the bigram model, the precision of the detector improves by almost 30%. Index Terms — n-gram model, support vector machine, local descriptor, Speed Up Robust Features, vehicle license plate detection