모바일 장비를 위한 광학문자 인식과 다단계 합성곱 신경망 접근법에 위한 차량 번호판 검출

Abstract

Existing number plate detections are suffering from light invariances and partial occlusions. The majority of number plate detection algorithms are using image preprocessing and geometrical rules. The ratio between the width and height of a possible number plate region is an example of a geometrical rule. Nevertheless, applying geometrical rules for number plate detection, within a real-time and real-world environment, results in misclassification of number plates due to partial occlusion or lightning invariance etc. Classifiers based on neural networks on the other hand, are capable to recognize an ever-changing scene according to the amount and kind of training data. Classifiers based on neural networks require sliding windows to scan the complete scene, which requires a higher computation time than algorithms based on geometrical rules. Sliding windows-based number plate detection algorithms were not suited for mobile devices in the past, because the devices did not have sufficient computation power. Nevertheless, recent research for number plate detection shows that convolutional neural network (CNN) based classification leads to higher detection results [1, 2]. A method is proposed to achieve improved number plate detection for mobile devices by applying a multi CNN approach while reducing the total amount of classification steps. The proposed algorithm processes supervised CNN-verified car detection first. In the second step, the detected car regions are applied to the next supervised CNN-verifier for number plate detection. In the final step, the detected number plate regions are verified through optical character recognition by another CNN-verifier. Since mobile devices are limited in computation power, we propose a fast method to recognize number plates. It is expected to be used in the field of intelligent transportation systems (ITS). Monitoring vehicles results in gathering data of vehicle movement and traffic flow.