Study for License Plate Detection

Published Online December 2019 in MECS

For the smart cities, intelligence transportation system is important for promoting the applications that used in the surveillance system of highway road, electronic toll gate system, traffic control system, and many others. License Plate Recognition (LPR) system is one of the main drivers of intelligent transportation systems. LPR recognizes the characters that exist over the license plates to identify the vehicle by applying image processing and computer vision techniques. LPR system composed of two main parts: License Plate Detection (LPD), and Character Segmentation and Recognition. LPD is the basic part of LPR system and the accuracy and the processing speed of LPR system are dependent on the accuracy of LPD and the efficiency of running time. So,

LPD system is the important step or the heart of LPR System.

The most common useful detection methods for license plate [1, 2, 3, 4, 5] are edge detection, morphological operation, color and texture based detection, and deep learning based detection. License plates have many different characteristics [6]:

• •    Color : the type of the license plate can distinguish with the color of license plate. The type is mostly difference on the background color of license plate.

• •    Texture: the license plate has a continuous regular border, and edge features are more abundant in the vertical direction than the horizontal direction.

• •    Text : the number of license plate is composed of numbers and characters. Many cities’ license plate is written with the English letters and numbers.

• •    Style : the style of license plate is different based on the shape, area, space of character and distance between the characters and borders.

The main objective of comparative research is the identification of similarities and differences between social entities. In our community, traffic must be monitored and vehicle registration can be achieved. LPD system serves to track incoming and exiting traffic by municipalities. The purpose of this study is to understand an effective, automated identification scheme for the approved car using the number plate for the car. It aims to draw conclusions about past events which may help to anticipate or explain future events.

This paper is organized with the following section. Section 2 introduces about the critical reviews that have been recently. Section 3 describes the comparison of previous methods. Section 4 presents the general steps of license plate detection system and section 5 finally concludes the paper and then follows the references

• II.    Critical Reviews

Rabbani et al. proposed the License Plate Detection and Recognition System for Bangladeshi License by utilizing Morphological operation and convolutional neural network. Their system composed with four modules: license plate area detection, license plate extraction, characters and words segmentation, and characters and words recognition. Image preprocessing works image resizing, image enhancement by using haze removal techniques, and noise removal by applying wiener 2-D filter. The aspect ratio between 1.9 and 2.1 is used to detect the area of license plate. After detection, horizontal line rescaling and vertical line rescaling are used in the corners of license plate. Connected component analysis is used for character segmentation and convolutional neural network is used to recognize the character over license plate. The authors implemented the system on the customized dataset for digits and characters of Bangla. All accuracies of detecting license plate, segmenting and recognition the characters are over 90%. The accuracy will be loss when the license plate’s characters are separate. This system used small dataset, and need high resolution image to detect broken characters for the characters segmentation. [7]

Yuan et al. proposed the robust and efficient techniques for the real time detection of license plate in complex scenes by applying novel line density filter (LDF) for candidate extraction and cascaded license plate classifier (CLPC) for candidate verification. Downsampling and gray scale conversion of image are used in the pre-processing step. Before Line Density Filter for Candidate Extraction, edge detection and binarization are processed by using Sobel operator and adaptive thresholding (AT) respectively. After extracting the candidate region, finding the candidate with connected-component labeling (CCL) and removing the un-exhibited area of the license plate image is operating. Finally, the license plates are identified from among the candidate regions that detected from the image by applying a cascaded license plate classifier (CLCP) and used the saliency features of two color channels: HSV and RGB. The system is implemented on the two datasets: Caltech vehicle dataset and their collected PKU dataset. Average accuracy of the system is 96.62% and the average run time is less than other approaches. Different scenes are the limitation of the system and the authors suggested interesting the MSER or Hough transform approach for the further study. [8]

Character-specific extremal regions (ERs) and hybrid discriminative restricted Boltzmann machines (HDRBMs) are presented for a vehicle license plate recognition by Gou et al.. There are four main steps for vehicle license plate recognition: license plates’ coarse detection, character region extraction, license plate detection, and license plate recognition. In the first step, background noises are removed with top-hat transformation, vertical edges are detected with Sobel filter, and back spaces are removed by applying Closing that is one of morphological operations. Real AdaBoost classifier is applied to select character-specific ERs as the region of characters in the second step. For the license plate detection step, license plate location and character segmentation are processed by inferring geometrical character attributes. In the last step, histogram of oriented gradients (HOG) descriptors is used to extract the features of character regions and HDRBM are used to recognize the character as an offline trained classifier. The system is experimented on the dataset extended the conference version dataset with real traffic monitoring scenes under various illumination conditions. The average performance percentages of License Plate Detection Rate (LDR), Character Recognition Rate (CRR), Overall Performance (OVR1 and OVR2) are 95.9, 98.2, 91.9 and 94.1, respectively. Further work may be deep architectures for location and recognition. [9]

Davis et al. introduced the faster license plate detection mechanism by using vertical edge detection algorithm (VEDA). The gray image conversion and adaptive thresholding are used in the preprocessing step. After preprocessing, vertical edges detection step is processed by applying unwanted lines estimation algorithm (ULEA) and vertical edge detection algorithm (VEDA). And then, License Plate Extraction are operated by implementing highlight desired details (HDD), candidate region extraction (CRE), plate region selection (PRS) and eliminating unwanted regions (EUR). For working in the global system, the system experiments on Indian vehicle image and also on different countries license plate images. The VEDA for vertical edge detection can more detect than histogram equalization and morphology operations. License plate detection system based on VEDA takes less computation time and it can detect other text blocks. However, it cannot detect in the highly blur image. [10]

Fomani and his colleague proposed the license plate detection algorithm by implementing adaptive morphological closing to locate the regions of gray level image, local adaptive thresholding for image smoothing, and morphological opening to separate the license plate region and other regions. Local histogram equalization (LHE) is used to preprocess the image before the main steps of the system. Adaptive morphological closing is composed of dilation, constructed structuring elements and erosion. Local adaptive thresholding is calculated based on the maximum and minimum value of width and height of license plate. The detection rate and computation time of the approach is measured on some real dataset collected in different situations. The detection rate of the algorithm is more than 99% and the computation time of the algorithm is faster than other algorithms. [11]

In 2017, Li et al. proposed a powerful license plate detection method by using Convolutional Neural Network (CNN). The system firstly generates the convolutional feature maps from the vehicle image using CNN, and then complete license plate sub-window are extracted on the convolutional feature maps by applying a single-scale sliding-window detector. Finally, a regression network is used to locate the license plate. The system is implemented on their own high-quality data set.

The CNN-based license plate detection method is more powerful than others and it can reduce the detection time. But, the image that takes on the illumination status is difficult to detect and the CNN-based detection method is needed to add the image enhancement methods. [12]

Shi et al. introduced the detection algorithm that applied the visual feature and convolutional neural network. The algorithm contains two main parts: candidate box generation, and candidate box classification and regression. For candidate box generation, there are two types: edge detection based method and color model based method. Edge detection based method is operated by using Gaussian blurring, grayscale conversion, sobel edge detection, binarization, and closing operation. In color model based method, HSV color space conversion, denoising using mean value, dilation and closing operation are processing on the color features. In the candidate box classification and regression step, two cascaded convolutional neural networks, namely C-Net and R-Net, are applied. C-Net is used to determine the place that exist the license plate and R-Net is used to judgment the position and size on the results of C-Net. The experiment is worked on the UCSD dataset and their own collected Chinese license plate images dataset. The algorithm that used visual features and convolutional neural network gets good performance on the given dataset, improved accuracy rate, and faster performance speed. However, the algorithm still has issues such as miss detection in actual scenes that contain many license plates and not satisfaction in the candidate box regression. [13]

The comparison of the previous research works is presented in this section with two tables. This section is aimed to know about the methods and datasets that used in the previous study and the advantage and disadvantages of the previous studies. Table 1 shows the comparison of the license plate detection methods and the datasets that used in the reference papers. Table 2 presents the advantages and disadvantages of the license plate detection methods

Table 1. Detection methods and Lecense Plate Dataset