License Plate Extraction for Moving Vehicles

Abstract

In this paper, a license plate extraction algorithm is proposed, which can be used in moving vehicles. First, the Haar cascade classifier was used to find candidate regions. Then, a DoG filter was used to detect the edges and connected component labeling was applied to obtain the candidate blocks. The license plate color characteristics were used to eliminate irrelevant blocks using histogram comparison and color quantization. The Bhattacharyya distance and the correlation metric were used to compare the histograms. Experiments with real data showed good performance. The dataset consists of various road and weather conditions including expressway, downtown, sunny days and rainy days. For our dataset, the recall was 0.72, the precision was 0.88 and the F-score was 0.79. For the Caltech dataset, the recall was 0.86, the precision was 0.96 and the F-score was 0.91.