License plate recognition based on edge histogram analysis and classifier ensemble

In this paper, a new approach for Iranian vehicle license plate recognition (LPR) is proposed. The proposed algorithm contains four main steps including plate localization, segmentation, recognition, and fusion of multiple recognition results. The license plate localization is begun with some preprocessing for down-sampling, denoising and histogram equalization. Then, histogram of vertical edges is used for detection of candidate lines expected to contain the license plate. In this step, we design a filter in order to reduce the number of false line candidates. The candidate plates are then extracted using vertical projection histogram of edges and aspect ratio characteristic. The binary image of these candidates obtained by locally adaptive thresholding is transmitted to the segmentation and recognition modules. Our recognition method is accomplished using a classifier ensemble with mixture of experts architecture. Using a feedback from the recognition result of candidate plates, the true candidate is detected. To improve the recognition accuracy and robustness, we apply the proposed LPR on three consecutive frames of a vehicle captured by different exposure times and then combine their recognition outputs. The experimental results in practical situations of day and night show that the proposed approach leads to 95.39% accuracy in plate localization and 92.45% overall accuracy after recognition.