In the State of Mexico, millions of vehicles circulate every day, so the identification of their license plates through CCTV cameras is an arduous job that requires trained personnel to capture important scenes. However, automatic plate recognition has factors that prevent correct detection. These can be problems such as weather, rusty plates, lighting, etc. A method of detecting plates located in any part of the vehicle is proposed and by means of a convolutional neural networks (CNN) algorithm using Tensor Flow. This is located, and then the text is segmented with the use of a threshold. The proposed model consists of a set of 355 individually obtained license plate images of the State of Mexico, which will serve to train the model. The main contribution of this work is the detection of vehicle plates anywhere, in addition to the implementation of a threshold with a value of 155 as the optimal value. The methodology was validated with the use of 50 new images of car plates registered in the vehicle registry of the State of Mexico, reaching an accuracy of 92.56% effectiveness.
TopIntroduction
In the State of Mexico, currently 2 million vehicles circulate per day, these present road problems such as reports of theft, collisions between vehicles, etc. So the traffic department of the state government has a staff assigned to control this large number of incidents. That is why strategies have been implemented to detect the plates of these vehicles, such as the installation of closed-circuit television cameras (CCTV), but the personnel assigned to identify these problems in the cameras is not enough, also observing problems of fatigue or staff distractions. that monitors these cameras, with which they let events pass where there may be events such as vehicular mishaps, among others.
In this paper we present a technique based on convolutional neural network using the TensorFlow platform, to subsequently apply the segmentation technique based on the threshold for the detection of the characters of the vehicle license plate. Once threshold technique is applied, the binarized characters contain noise, so it is necessary to implement techniques based on mathematical morphology, to leave only the characters on the plate.
Preliminaries
The recognition of vehicle license plates has been studied in the state of the art, have served to search for techniques that help each time to detect vehicle license plates in spite of the mentioned problems. The most relevant work being done in relation to license plate detection is mentioned below.
In (Lubna, Mufti, & Ali Shah, 2021), a detailed review of current technologies and advances in Automatic Number Plate Recognition (ANPR) systems is presented. With a complete comparison of the performance of different algorithms tested and simulated in real time. ANPR technology can detect the identification of vehicles through their license plates using identification techniques. Successful implementation of the ANPR system may require additional hardware to increase the accuracy of the system. License plate status, non-standard format, complex scene, camera quality, camera installation position, distortion tolerance, image quality, etc. Camera stabilization position, distortion tolerance, motion blur, contrast issues, reflections, display and processing limitations and memory, environmental conditions, indoor/outdoor or day/night photography, software tools. Their study aims to improve the state of the art in Intelligent Transportation Systems (ANPR) based on computational vision algorithms.
On the other hand, in (R.C. Chen et al, 2019), they designed a prototype called Automatic Plate Recognition (ALPR), based on a deep learning framework using YOLO, where they occupy 7 convolutional layers for plate recognition with the use of a sliding window that detects the digits of the plate. Finally, an accuracy of 96.94% is obtained in the recognition of the characters on the plate. On the other hand (W Puarungroj & N. Boonsirisumpun, 2018), carried out motorcycle license plate detection, such license plate recognition system combines 2 subsystems of license plate detection and recognition by deep learning (MobileNets and Inception-v3), presenting an accuracy of 96.94% in line recognition and 91.76% in character recognition. However (Z. Selmi, M.B. Halima, U. Pal, & M.A. Alimi, 2020), design automatic license plate recognition (ALPR) systems, used in very complicated environments, are based on convolutional neural networks (CNN) and tested on four license plate datasets with different characteristics such as: orientation, low quality images, blurred and license plates with complex backgrounds. In the end, an accuracy of 99.3% in license plate detection is reported.
Characteristics of the License Plates of Automobiles in the State of Mexico
The configuration of texts and images that contains a vehicle license plate, varies in its characteristics in any state of the Mexican Republic, as well as in different countries of the world. In this work, we focused on type of license plates used in the State of Mexico, which are made up of three letters and four numbers that allow the control of vehicles. According to the traffic regulations of the State of Mexico, automobile identification plates must be installed in the place of the vehicle designated for this purpose, so that one is on the front and the other on the back. The license plates must be kept in good condition and free of objects, marks, signs, opaque micas and bends that may hinder or impede their reading.
Figure 1. Example of vehicle license plate of the State of Mexico with the characteristics that identify it