License Plates Detection and Recognition Based on Semi-supervised Learning

Abstract

AbstractThe paper is devoted to the problem of vehicles license plates detection and recognition in real time based on machine learning methods. Proposed by the authors is the software system which combines a pseudo-labeling method for automated preparation of the dataset and two separate convolutional neural networks for license plates detection and symbols recognition, where the YoloV3 convolutional neural network is used for detection and a convolutional recurrent neural network is used for symbols recognition. The main feature of the proposed system is the semi-supervised learning method which allows obtaining a huge dataset for neural networks training containing 95,000 images with marked coordinates of license plates. Integration of the mentioned approaches into a single working flow proves its efficiency due to significant reduction of efforts on dataset preparation and high accuracy of the detection and recognition: the implemented system recognizes license plates with 94,8% accuracy on a test sample of 20,000 images.KeywordsLicense platesDataset preparationSemi-supervised learningConvolutional neural networkSymbols recognition