Reconhecimento automático de placas veiculares brasileiras incorporando Max-Margin Object Detection e redes neurais convolucionais

Abstract

Automatic License Plate Recognition (ALPR) consists in identifying and reading license plate texts of a least one of vehicles in a image. The ALPR systems are present in various applications, such as traffic fiscalization and parking lot access control. In general, an ALPR system is composed by three stages: license plate detection, character segmentation and segmented characters recognition. The solutions involving systems that deals with brazillian license plate particularities may show more especific results or use high computational cost techniques in order to hanle less controlled environments. Thus, this work proposes of a system that seeks to be compatible with real world data using simple image processing techniques and also more complex but not too costly techniques, such as Max-Margin Object Detection and simple convolutional neural networks. The proposed system shows promising results, specially in license plate detection stage and character classification, finding its biggest dificulties in character segmentation stage.