An automated system to detect and recognize vehicle license plates of Bangladesh

Bangladesh is a country in South Asia that uses Retro Reflective license plates. The plate has two lines with words, letters, and digits. An automated system to detect and recognize these plates is presented in this paper. The system is divided into four parts: plate detection, extraction, character segmentation and recognition. At first, the input image is enhanced using CLAHE and a matched filter specially designed for license plates with two lines is applied. Then tilt correction using Radon transformation, binarization and cleaning are performed. For character segmentation, mean intensity based horizontal and vertical projection is used. In recognition, we have used two different Convolutional Neural Network (CNN) to classify digits and letters. Tesseract OCR is used for district names. We have developed a dataset of over 400 images of different vehicles (e.g., private car, bus, truck etc.) taken at different times of day (including nights). The plates in this dataset are in different angles, including blurry, worn-out and muddy ones. On this dataset, the proposed system achieved a success rate of 96.8% in detection, 89.5% in extraction, 98.6% in segmentation and 98.0% in character recognition.