The event database contains all event-related information for the World of Photonics Congress.
Lecture Hall A1 SPIE Digital Optical Technologies > Digital Optical Technologies II > Digital Optics for Image Formation
13:50-14:10 h | Hall A1 Room 9 "Arthur Schawlow"
Subjects: Optics for Augmented and Virtual Reality Systems
Optical Character Recognition (OCR) is one of the most widely used digitization techniques, which recognizes the images of characters/texts and converts into machine editable format. Automatic License Plate Recognition (ALPR) is a technology designed to automatically read vehicle license plates. Traditional ALPR systems first detect the License Plate (LP) and then apply the OCR pipeline, which comprises LP image pre-processing, character segmentation, character classification and post-processing. This paper presents a novel real-time ALPR system which is developed and deployed by the VINCI Autoroutes company. An ALPR system developed with the traditional approaches often fails to provide acceptable results due to numerous challenging situations, particularly while a single system is used to read the LPs of different (a) origins or countries; (b) vehicle types e.g., cars and motorbikes have different license plates shapes and (c) image capturing conditions. These challenges significantly increase the variability of LPs and the characters to be classified. In order to overcome these challenges, recent methods use Convolutional Neural Network (CNN) models . However, many of these CNN based methods still exhibit vulnerabilities to environmental conditions: as a consequence, they generate an imprecise localization of the region of the characters’ sequence and fail at doing a proper segmentation. In order to overcome these limitations, we propose a novel method that take profit from the the saliency map within a CNN model. The key contribution is at the segmentation step where the characters are located via the saliency map. This map highlights the most probable position of a character based on the map computed by the object detector model. Then, the proposed system consists of the following modules: 1) LP-localization-CNN to detect the LP from the image captured at the toll system (large view). 2) Saliency-Map-CNN for segmentation to locate (segment) the LP characters. 3) Slim-Classification-CNN to efficiently classify the detected characters. The proposed ALPR system was evaluated on a proprietary dataset collected by the company VINCI Autoroutes (5000 images). This dataset includes various challenging cases such as were highlighted before. As the baseline, traditional pipeline was used, where character localization was achieved via local binarization method with considerable amount of preprocessing. The obtained results demonstrate that the proposed approach outperforms traditional pipelines in terms of accuracy and computation time. Additionally, we evaluated our method on the UFPR-ALPR  and the ReID  public datasets. Obtained results clearly show the efficiency of the proposed system and its suitability for real-world ALPR applications. References:  Gu et al. Recent advances in convolutional neural networks. In Pattern Recognition 2018.  Laroca et al. A Robust Real-Time Automatic License Plate Recognition based on the YOLO Detector, In IJCNN 2018.  Špaňhel et al. Holistic recognition of low-quality license plates by CNN using track annotated data, In AVSS 2017.