Advanced
Real-Time License Plate Detection in High-Resolution Videos Using Fastest Available Cascade Classifier and Core Patterns
Real-Time License Plate Detection in High-Resolution Videos Using Fastest Available Cascade Classifier and Core Patterns
ETRI Journal. 2015. Apr, 37(2): 251-261
Copyright © 2015, Electronics and Telecommunications Research Institute(ETRI)
  • Received : August 07, 2014
  • Accepted : November 12, 2014
  • Published : April 01, 2015
Download
PDF
e-PUB
PubReader
PPT
Export by style
Share
Article
Author
Metrics
Cited by
TagCloud
About the Authors
Byung-Gil Han
Jong Taek Lee
Kil-Taek Lim
Yunsu Chung

Abstract
We present a novel method for real-time automatic license plate detection in high-resolution videos. Although there have been extensive studies of license plate detection since the 1970s, the suggested approaches resulting from such studies have difficulties in processing high-resolution imagery in real-time. Herein, we propose a novel cascade structure, the fastest classifier available, by rejecting false positives most efficiently. Furthermore, we train the classifier using the core patterns of various types of license plates, improving both the computation load and the accuracy of license plate detection. To show its superiority, our approach is compared with other state-of-the-art approaches. In addition, we collected 20,000 images including license plates from real traffic scenes for comprehensive experiments. The results show that our proposed approach significantly reduces the computational load in comparison to the other state-of-the-art approaches, with comparable performance accuracy.
Keywords
I. Introduction
Automatic license plate recognition (LPR) systems have a long history owing to their usefulness. Since their first development in the 1970s, many researchers have studied automatic LPR systems using various methods under different conditions [1] [2] . The advancement of camera devices, computer hardware, and computer vision algorithms has enabled automatic LPR systems to be more applicable in difficult environments. Furthermore, the necessity for automatic LPR systems has increased with the recent developments of the automatic illegal parking detection systems [3] .
An LPR system is composed of four stages. First, the system reads images from a camera (image acquisition). The accuracy of automatic LPR systems often depends on the quality of the camera being used and method of its installation. Second, the system detects license plates (LPs) in acquired images (LP detection). Third, the system segments characters in the detected license plate images (LP segmentation). Finally, the system recognizes the characters and then attempts to match them to license numbers in a database (character recognition). Among these four stages, in this paper, we focus on LP detection because it significantly affects the accuracy and is the most time-consuming stage of LPR systems.
Extensive studies on LP detection have been conducted. In [1] , they categorize LP detection methods based on the following different features: boundary/edge, global image, texture, color, character, and their combination. Texture features are more robust than boundary/edge and color, for example, but are also more time consuming in terms of their extraction. A cascade classifier has been suggested to reduce the computation time of feature extraction. Because extracting all features in candidate sub-windows requires a long processing time, several approaches using a partial set of weak classifiers to reject non-LP candidates have been suggested [4] [7] .
Recently, high-resolution videos (720p, 1,080p, and so on) have become increasingly available owing to the advancement of camera devices. As the resolution of an image increases, the search area becomes larger and the time for LP detection increases. For example, one high-definition camera with a 1,920 × 1,200 resolution is easily able to detect two or more lanes for traffic law enforcement (speed violations or crossing the center line), as shown in Fig. 1 . To detect license plates in the high-resolution image in Fig. 1 , most of the detection algorithms run 7.5-times longer than when detecting license plates in a VGA image. Previous algorithms often have a difficulty to be implemented for real-time LP detection in such high-resolution videos.
PPT Slide
Lager Image
(a) Multiple vehicles and their license plates are present in a high-resolution image. Resized image is shown. The original image has 1,920 × 1,200 resolution and (b) cropped license plate images in original resolution. High-resolution cameras can capture multiple license plates in more than two lanes.
Therefore, we propose two approaches for efficient and robust LP detection in real time. First, we propose the fastest cascade structure available, which is able to reject non-LP candidates with a minimal number of weak classifiers. Second, we propose an efficient training method for the different styles and sizes of license plates.
The rest of this paper is organized as follows. Section II introduces other works related to LPR, such as LP detection, AdaBoost-based approaches, and relevant cascade structures. In Section III, we describe the details of our training method, feature extraction, and proposed cascade structure. Section IV presents extensive experimental results. Finally, some concluding remarks are provided in Section V.
II. Related Work
In this section, we introduce previous state-of-the-art approaches for LP detection. The overall descriptions of these state-of-the-art methods for automatic LPR are presented in [1] and [2] . In addition, several binarization methods for LP character segmentation are compared in [8] .
- 1. LP Detection
Several approaches based on edge detections and morphological operations have been proposed for LP detection because license plates have clear edges and character shapes [9] [10] . These approaches are intuitive and powerful when the scenery images of the license plates are not noisy. However, the character edges in a license plate can be unobtainable owing to environmental challenges. On the other hand, text feature–based license plate detection approaches do not require an accurate extraction of such edges [6] , [11] . These approaches perform better than edge detection–based or morphological operation–based approaches when the boundary of a license plate is unclear. Our approach is also that of a text feature–based approach. These approaches are computationally expensive, especially when the scene is complex and many edges are present. However, our algorithm is very fast and can detect license plates in a high-resolution image within a significantly shorter period of time in comparison with state-of-the art text feature–based LP detection approaches.
- 2. AdaBoost-Based Approaches
Adaptive boosting (AdaBoost) was combined with the modified census transform (MCT) and Haar-like features for various purposes. The MCT was proposed by Fröba and Ernst [5] and improves the local binary pattern (LBP) [12] and census transform (CT) [13] for face detection. The authors in [14] proposed adaptive AdaBoost learning for face detection by regression techniques. AdaBoost learning and MCT features were also used for a cascade classifier of LP detection systems [7] and sign detection at subway stations [15] . Viola and Jones [4] developed a face detection framework by rapidly calculating the Haar-like features with boosting. Several other approaches have used Haar-like features for an LP detection system [16] [17] .
- 3. Cascade Structure
A cascade structure is useful for the detection problem. Although we use a binary classifier regardless of whether a detection window contains a wanted object, the number of positive windows (containing a wanted object) is much less than the number of negative windows. To reduce the computational time, it is better to reject a higher number of negative windows with a smaller number of features. For this reason, most previous approaches use a cascade structure [4] [7] , [14] [19] . These approaches have a fixed number of stages in a cascade structure and train the thresholds for the classifier in each stage. Such a cascade structure largely reduces the processing time, but not maximally. In addition, training the thresholds for each stage is difficult when the number of stages is large. Therefore, we propose the fastest available cascade structure that rejects negative samples in the earliest stages. In addition, our algorithm does not need to train all thresholds for every stage.
III. Approach
- 1. Creating Training Set for LP/Non-LP Classifiers Using Core Patterns.
A. Various LP Types
We categorized the license plates into nine different types depending on the background or character colors, the number of rows, and the locations of the characters. Figure 2 shows the shapes of all license plate types. The variety of license plate types makes the LP detection more complex.
PPT Slide
Lager Image
Example of each type of license plate. (A1) to (A4) have characters that are darker than their respective plate backgrounds, and (B1) to (B5) have characters that are brighter than their respective plate backgrounds.
B. Creating Positive Training Set with Core Patterns
The height/width ratio ( H / W ratio) of the license plates is not uniform across plate types. Because we prefer to use a detection window of a fixed size for fast searching, we need further processing on the license plate sample images with a high variation in H / W ratio. We suggest the following four ideas to unify the H / W ratio of the license plate sample images:
  • ∎ Training a classifier of a specific type of license plate versus the others.
  • ∎ Changing the aspect ratio of the license plate sample images to keep all plate sample images in a uniform aspect ratio.
  • ∎ Cropping the core sub-image (core patterns), which has the same aspect ratio, from the license plate images.
  • ∎ Including the neighbor region of the license plate images so as to not lose any regions of the plates.
For the first idea, multiple classifiers need to be created, which increases the entire processing time based on the number of classifiers. When we apply the second idea, the detection performance decreases, unless we resize the source images with various ratios. These ideas are not appropriate for real-time implementation. On the other hand, the third (core patterns) and fourth ideas (entire plates) solve this problem with the same aspect ratio. The training samples from these two ideas, core patterns and entire plates, are presented in Fig. 3 . To demonstrate the effectiveness of the proposed idea using core patterns, we compared the LP-detection performance from training using the core patterns with training using images of entire plates, as shown in Fig. 4 . The LP-detection training using the images of entire plates performs better than that from using the core patterns for a true positive rate of under 0.94. However, training using core patterns becomes superior to training using images of entire plates for high true positive rates. Since most LP-detection related industries require a 99% or higher detection rate, our algorithm using core patterns is more appropriate for LP detection. Therefore, we use core patterns for our other experiments.
PPT Slide
Lager Image
Positive training images (gray scale) of different license plate types. Because of the H/W ratio difference of license plates, training images in left column (entire plate) should include a large portion of non-LP regions.
PPT Slide
Lager Image
ROC curves comparing the performance of LP detection using training images of core patterns (red star) and entire plates (blue lozenge).
C. Creating Negative Training Set
We collected ten negative samples from each training image using the following procedure:
  • Choose a sub-window at a random position in the training image.
  • If the chosen sub-window is overlapped by more than 25% of the window by another selected sub-window or positive training sample (already annotated), then the sub-window is discarded. Otherwise, the sub-window is selected.
  • Repeat this process until ten samples are collected.
Examples of some of the collected negative training samples are shown in Fig. 5 . In the next subsection, the training samples are resized to the size of the fixed detection window in preparation for the next step; that is, feature extraction.
PPT Slide
Lager Image
Example of negative training images.
- 2. Local Structure Features by MCT
The MCT was proposed to capture local structures not captured by the LBP [12] . The LBP at pixel p is defined as
(1) LBP(p)= ⊗ q∈N(p) C( I(p),   I(q) ),
where I ( p ) is the intensity value (gray scale) of pixel p , C ( a , b ) is a comparison function that returns a value of one if a b and returns zero otherwise, N ( p ) is a set of eight surrounding pixels of p , and ⊗ is a bit concatenation operation.
The MCT at pixel p is defined as
(2) Γ(p)= ⊗ q∈ N ′ (p) C( I ¯ (p),  I(q) ),
where N ′( p ) is a set of nine pixels including the eight surrounding pixels of p and p itself, and Ī ( p ) is the average intensity of the set of pixels N ′( p ). The MCT requires one more bit to represent a transform than the LBP, but the MCT can capture more local structures than the LBP. Figure 6 shows an example of the LBP and MCT, where I ( p ) is the highest among the intensities of the surrounding pixels of p . When I ( p ) is higher than any I ( q ) such that q N ( p ) then LBP( p ) is 00000000 (2) and only assigns the same kernel. However, MCT can capture such a local structure.
PPT Slide
Lager Image
(a) Sample intensity values of 3 × 3 pixels, (b) LBP feature vector becomes 00000000(2) because intensity of all surrounding pixels of the center is lower than the intensity of the center pixel, and (c) MCT feature vector becomes 011110001(2).
Figure 7 shows a sample license plate image and its MCT visualizations. The characters in an 18 × 6 sized MCT image are not recognizable at all by the human eye, and the characters in a 24 × 8 sized MCT image are more recognizable but confusing. We can clearly recognize the characters in the other larger-sized MCT images. Because the MCT visualizations of a 36 × 12 sized MCT image and a 42 × 14 sized MCT image are very similar, we expect that the 36 × 12 dimension size is sufficiently large for MCT feature extraction.
PPT Slide
Lager Image
All training images are resized to a fixed pixel dimension, such as 18 × 6, 24 × 8, 30 × 10, 36 × 12, and 42 × 14. This figure shows resized training images, their MCT visualizations, and red marked dots for positions of the selected stronger (5%, 10%, and 100%) weak classifiers in each column.
- 3. Proposed Cascade Structure with AdaBoost Learning
To detect license plates in a given image, we classify millions of sliding windows using our classifier. Our classifier consists of look-up-table (LUT) weak classifiers using the AdaBoost learning procedure in a novel cascade manner. More details of each component are presented in the following subsections.
A. LUT Weak Classifier
An LUT weak classifier is proposed by Wu and others [20] . An LUT classifier is able to efficiently classify multi-Gaussian distributed samples, uses fixed bins, and is sensitive to the number of bins. Because the MCT features have a fixed number of bins (511), it is more appropriate to build LUT weak classifiers using MCT features than other features such as Haar-like features.
Figure 8 illustrates the LUT training and classification algorithm. Once we determine the size of the detection window (36 × 12), all of the training images are resized to this size. The MCT is applied at positions p = ( x , y ) except for the edge pixels, and the output has 511 possible feature indices. A histogram with 511 bins is created from Γ( p ) of all training samples. The LUT classifiers are assigned the values “+1” if the number of positive samples is larger than the number of negative samples; otherwise, they are assigned the value “−1,” as shown in (3) below.
h(p,k)={ +1    if    L pos (p,k)> L neg (p,k), 1     otherwise,
where
L p pos (k)
and
L p neg (k)
are the numbers of positive and negative samples whose feature index is k at position p , respectively. Note that we use the values “+1” and “−1,” while previous approaches [4] [6] , [19] used “+1” and “0.”
PPT Slide
Lager Image
Example LUT classifier for license plate detection.
While the LUT classifier in Fig. 8 considers all samples having the same weight, it can be more effectively learned using the AdaBoost algorithm, which is described in the next subsection.
B. AdaBoost Learning Procedure
AdaBoost is a method for training a boosted classifier, which is a linear combination of weak classifiers, to achieve a stronger classifier. This algorithm selects a weak classifier with the smallest weighted error in every round. The weighted error can be calculated by the updated weight of the training samples, and the weight of the training samples is updated by increasing the weight if the corresponding sample is misclassified, and by decreasing the value otherwise. The details of the AdaBoost learning procedure are as follows:
AdaBoost Round: t = 1, … , T .
  • Choose a weak classifierht(p, Γ(p)) at pixelp. The pixel is chosen using p=argminp  ϵt(p),whereϵt(p)=∑i|ht(p,Γ(p))≠CiDt(i),Ciis the class of training samplei(+1 if positive, −1 otherwise), andDt(i) is the weight of training sampleiat roundtsuch thatDt(i) ∑iDt(i)=1.
  • Compute the weight of the weak classifier from the errorαt=12ln(1−ϵt(p)ϵt(p)).
  • Update the weight of the training samplesDt+1(i)=Dt(i)e−αtCiht(p,Γi(p))∑iDt+1(i).
The tables of the local indices are regenerated using
L pos (p,k)= i S pos D t (i)I( Γ(p)=k ) .
Finally, the strong classifier Γ( p ) is calculated by the linear combination of weak classifier ht ( p , Γ( p )), as in
H(Γ)= t=1 T α t h t ( p,Γ(p) ).
C. Proposed Fastest Available Cascade Structure
To achieve a high rate of accuracy, a low false-positive rate, and a reduction in computational load, we use a cascade structure. When searching for license plates using a sliding window, the number of non-plates is much larger than the number of plates. Therefore, the early stage of the cascade framework rejects a large portion of negative samples, whereas most of the positive samples are accepted. Most of the previous detection algorithms using AdaBoost use a three- or four-staged cascade structure.
Here, we propose a cascade structure that we believe to be the fastest in existence. Unlike state-of-the-art cascade structures collecting 5% to 10% of the weak classifiers in the first stage, our cascade structure only collects two weak classifiers. Whenever every n th additional weak classifier is combined to a strong classifier for n > 1, the strong classifier rejects negative samples in the n th stage. To the best of our knowledge, the suggested structure is the fastest cascade structure for detecting license plates. In addition, our approach is general enough to apply to other object detection problems. Flow charts of a baseline cascade structure with three stages and the proposed cascade structure are presented in Figs. 9(a) and 9(b) , respectively. The number of valid weak classifiers is represented by N .
PPT Slide
Lager Image
(a) Baseline cascade structure and (b) proposed cascade structure. Note that the proposed structure has a constant threshold Th for all decisions, while the baseline cascade structure has a different threshold Thi for the ith strong classifier.
Our cascade structure has an advantage in rejecting negative samples in the early stages, but we are concerned that the structure may reject more positive candidates than the baseline structures. Here, a positive candidate indicates a testing detection window that has a 75% overlapping region between a positive candidate and license plate. To check that our cascade classifier can correctly detect a license plate, we have to ensure that it never rejects all positive candidates. We examined the changes of the strong classifier value, H (Γ), in all stages for the strongest positive candidates in 10,000 license plates and for the 10,000 randomly selected negative candidates, as shown Fig. 10 . The strong classifier value H (Γ) for the strongest positive candidates does not drop below the single threshold Th (= 0) with any number of weak classifiers, as shown in Fig. 10(a) . On the other hand, some H (Γ) for the negative candidates drop below the threshold with a sufficient number of weak classifiers. These distributions enable our algorithm to reject more false positives without the chance of increasing the number of false negatives.
PPT Slide
Lager Image
(a) Strong classifier value changes by number of weak classifiers for selected positive samples and (b) strong classifier value changes by number of weak classifiers for random negative samples.
IV. Experiments
- 1. Dataset Description
To evaluate the proposed algorithms, we used 20,000 vehicle images in our experiments. These images were collected by high-definition cameras with a 1,624 × 1,224 resolution. The cameras were located on more than ten camera poles and were used for 24 hours to reflect the real traffic scene variability. Because the vehicle images were collected from a real traffic scene, some of the license plate types appeared more frequently than the other types. The numbers of the collected samples for the nine types are presented in Table 1 . The dataset is challenging because of variations in illumination, background, and plate, such as the location, size, type, color, and font. We implemented an LP detection system on a PC with 3.4 GHz i7 quad core CPU using C++.
Number of images in database for each license plate type.
LP type A1 A2 A3 A4
# of images 5,090 3,420 1,278 3,326
LP type B1 B2 B3 B4 B5
# of images 3,242 40 2,682 50 872
- 2. Detection Window Size
The size of the detection window has a high impact on the efficiency and accuracy of the detection system. We tested five detection window sizes (18 × 6, 24 × 8, 30 × 10, 36 × 12, and 42 × 14) to find the optimal solution. Using an 18 × 6 sized detection window, we only achieve a 0.8 true-positive rate with a high false-negative rate (5.0×10 −5 ), as shown in Fig. 11(a) . We expected this result from the MCT visualizations in Fig. 7 because the character texture is not recognizable in 18 × 6 sized MCT images. The performance of LP detection is highly improved by choosing a 24 × 8 or larger sized detection window. Figure 11(a) also shows that 30 × 10, 36 × 12, and 42 × 14 sized detection windows are comparable to achieve high true-positive rates. To achieve a 0.9998 true-positive rate, we can choose both 36 × 12 and 42 × 14 sized detection windows due to their low false-positive rates.
The processing time per frame for different window sizes is measured to find the optimal size of the detection window, as shown in Fig. 11(b) . We choose the 36 × 12 sized detection window between 36 × 12 and 42 × 14, as the processing time using 42 × 14 is about 50% longer than when using 36 × 12.
PPT Slide
Lager Image
(a) ROC curves comparing the performance of LP detection in five different sizes of detection window, and magnified ROC curves comparing the performance of LP detection in the three largest sizes of detection window and (b) average processing time per frame in five detection window sizes.
- 3. Proposed Cascade Structure vs. Baseline Cascade Structures vs. Haar-like-Based Detector
To compare the accuracy and efficiency of our proposed cascade structure with the baseline cascade structures, we implemented the LP detection system using both the proposed cascade structure and the two baseline cascade structures. In addition, to compare the performance of MCT-based LP detectors with Haar-like-based LP detectors, we implemented Haar-like-based LP detectors similar to [13] . Baseline cascade structure 1 has four stages with 17 (5%), 34 (10%), 68 (20%), and 194 (all) weak classifiers [6] . Baseline cascade structure 2 has N /3 stages; the first stage having three weak classifiers and each stage thereafter having an additional three.
We tested our approach and the baseline approaches using 10,000 images from nine LP types. The performance of the proposed cascade structure is comparable to that of the baseline cascade structures, shown in Fig. 12(a) . Haar-like-based LP detectors have also been tested using 2,045 images from the A1 LP type. Note that we also tested Haar-like-based LP detectors using images from all LP types, but their detection rates were significantly low due to the variation in LP types. Therefore, the MCT-based LP detector is more suitable to achieve a high true-positive rate (> 0.99).
The strength of our algorithm lies in its speed. Our algorithm is about three-times faster than baseline cascade structure 1, 1.5-times faster than baseline cascade structure 2, four-times faster than the proposed cascade structure with Haar-like features, and 25-times faster than a 4-stage cascade structure with Haar-like features, as shown in Fig. 12(b) .
PPT Slide
Lager Image
(a) ROC curves comparing the performance of LP detection in baseline cascade classifiers and proposed cascade classifier and (b) processing time (ms/frame) measured according to false-positive rate.
In addition, we analyze how the proposed cascade structure operates faster than the baseline structures. The number of rejected sub-windows (mostly negative samples) is measured in each stage of the cascade classifier. The number of rejected sub-windows decreases more drastically in the proposed cascade structure compared with baseline cascade structure 1, as shown in Fig. 13(b) . Because the classification time for sub-windows is proportional to the number of weak classifiers used, it is more ideal for the cascade classifier to have a smaller number of rejected sub-windows when it has a larger number of weak classifiers. While baseline cascade structure 1 rejects 726 and 118 sub-windows at stage 3 (68 weak classifiers) and stage 4 (194 weak classifiers), respectively, as shown in Fig. 13(a) , the proposed cascade structure rejects only 3.99 or fewer sub-windows with 68 or more weak classifiers. Therefore, the proposed cascade structure is able to significantly reduce the processing time for the LP detection. Note that the values in Fig. 13 are obtained by taking the average of all testing samples. If the value is 0.1, then rejections occur once in every ten testing samples.
PPT Slide
Lager Image
Number of rejected sub-windows are represented in bar graph: (a) baseline cascade classifier 1 rejects sub-windows when it collects 17, 34, 68, and all weak classifiers and (b) proposed cascade classifier rejects sub-windows whenever it collects one additional weak classifier.
Figure 14 shows successful examples of the proposed cascade structure reducing more false positives than the baseline cascade structure 1. When H (Γ) of the negative samples is higher than the threshold at all four stages, baseline cascade structure 1 fails to reject the negative samples. Note that our test images were obtained under various environmental conditions.
PPT Slide
Lager Image
LP detection results comparing proposed algorithm using fastest cascade structure and baseline algorithm using 4-stage cascade structure. Even in highly challenging environments, our algorithm is able to reduce false-positives more effectively.
V. Conclusion
We are able to detect LPs from high-resolution videos in real time by efficiently rejecting false positives using the proposed fastest available cascade structure and by training using the core patterns. Our novel cascade structure enables our system to reject false positives during the early stages. In addition, we show that training with the core patterns is more effective than training with images of entire plates. Our LP detection system is fast enough to be implemented in real time for embedded systems and performs accurately under the most difficult of environmental conditions.
BIO
kilyhan@etri.re.kr
Byung-Gil Han received his BS degree in electronics and electrical engineering and his MS degree in electronics engineering from the Department of Electronics Engineering, Kyungpook National University, Daegu, Rep. of Korea, in 2005 and 2007, respectively. Since 2010, he has been with the Electronics and Telecommunications Research Institute, Daejeon, Rep. of Korea. His major research interests include computer vision, pattern recognition, and video surveillance.
Corresponding Author  jtlee@ utexas.edu
Jong Taek Lee received his BS degree in electrical engineering from the Korea Advanced Institute of Science and Technology, Daejeon, Rep. of Korea, in 2005 and his MS and PhD degrees in electrical and computer engineering from the University of Texas, Austin, USA, in 2007 and 2012, respectively. Since 2012, he has been with the Electronics and Telecommunications Research Institute, Daejeon, Rep. of Korea. His major research interests include computer vision, activity recognition, and video surveillance.
ktl@etri.re.kr
Kil-Taek Lim received his BS, MS, and PhD degrees in electronics engineering from the Department of Electronics Engineering, Kyungpook National University, Daegu, Rep. of Korea, in 1993, 1995, and 1999, respectively. From 2004 to 2011, he was an assistant professor of the Computer and Information Engineering Department of Gyeongju University, Rep. of Korea. Since 2012, he has been with the Electronics and Telecommunications Research Institute, Daejeon, Rep. of Korea. His major research interests include computer vision, pattern recognition, and video surveillance.
yoonsu@etri.re.kr
Yunsu Chung received his MS and PhD degrees in electronics engineering from the Department of Electronics Engineering, Kyungpook National University, Daegu, Rep. of Korea, in 1995 and 1998, respectively. Since 1999, he has been with the Electronics and Telecommunications Research Institute, Daejeon, Rep. of Korea. He is currently working as the leader of the Regional Industry IT Convergence Research Section. His major research interests include biometrics, video surveillance, human–robot interface, and human–computer interface.
References
Du S. 2013 “Automatic License Plate Recognition (ALPR): A State-of-the-Art Review” IEEE Trans. Circuits Syst. Video Technol. 23 (2) 311 - 325    DOI : 10.1109/TCSVT.2012.2203741
Anagnostopoulos C.-N.E. 2008 “License Plate Recognition from Still Images and Video Sequences: A Survey” IEEE Trans. Intell. Transp. Syst. 9 (3) 377 - 391    DOI : 10.1109/TITS.2008.922938
Lee J.T. 2009 “Real-Time Illegal Parking Detection in Outdoor Environments Using 1-D Transformation” IEEE Trans. Circuits Syst. Video Technol. 19 (7) 1014 - 1024    DOI : 10.1109/TCSVT.2009.2020249
Viola P. , Jones M.J. 2004 “Robust Real-Time Face Detection” Int. J. Comput. Vis. 57 (2) 137 - 154    DOI : 10.1023/B:VISI.0000013087.49260.fb
Froba B. , Ernst A. “Face Detection with the Modified Census Transform” IEEE Int. Conf. Automat. Face Gesture Recogn. Seoul, Rep. of Korea May 17–19, 2004 91 - 96
Lee Y. “License Plate Detection Using Local Structure Patterns” IEEE Int. Conf. Adv. Video Signal Based Surveillance Boston, MA, USA Aug. 30–Sept. 1, 2010 574 - 579
Sheng H. 2009 “Real-Time Anti-Interference Location of Vehicle License Plates Using High-Definition Video” IEEE Intell. Transp. Syst. Mag. 1 (4) 17 - 23    DOI : 10.1109/MITS.2010.935911
Yoon Y. 2013 “Best Combination of Binarization Methods for License Plate Character Segmentation” ETRI J. 35 (3) 491 - 500    DOI : 10.4218/etrij.13.0112.0545
Zheng D. , Zhao Y. , Wang J. 2005 “An Efficient Method of License Plate Location” Pattern Recogn. Lett. 26 (15) 2431 - 2438    DOI : 10.1016/j.patrec.2005.04.014
Hongliang B. , Changping L. “A Hybrid License Plate Extraction Method Based on Edge Statistics and Morphology” Int. Conf. Pattern Recogn. Cambridge, UK Aug. 23–26, 2004 2 831 - 834
Jia W. , Zhang H. , He X. 2007 “Region-Based License Plate Detection” J. Netw. Comput. Appl. 30 (4) 1324 - 1333    DOI : 10.1016/j.jnca.2006.09.010
Ojala T. , Pietikainen M. , Maenpaa T. 2002 “Multiresolution Gray- Scale and Rotation Invariant Texture Classification with Local Binary Patterns” IEEE Trans. Pattern Anal. Mach. Intell. 24 (7) 971 - 987    DOI : 10.1109/TPAMI.2002.1017623
Zabih R. , Woodfill J. “Non-parametric Local Transforms for Computing Visual Correspondence” European Conf. Comput. Vis. Stockholm, Sweden May 2–6, 1994 151 - 158
Yun W.-H. 2010 “Disguised-Face Discriminator for Embedded Systems” ETRI J. 32 (5) 761 - 765    DOI : 10.4218/etrij.10.1510.0139
Lee D. 2014 “Robust Sign Recognition System at Subway Stations Using Verification Knowledge” ETRI J. 36 (5) 696 - 703    DOI : 10.4218/etrij.14.2214.0007
Zhang H. “Learning-Based License Plate Detection Using Global and Local Features” Int. Conf. Pattern Recogn. Hong Kong, China Aug. 20–24, 2006 2 1102 - 1105
Dlagnekov L. 2004 “License Plate Detection Using AdaBoost” Comput. Sci. University of California San Diego, CA, USA
Jun B. , Kim D. 2012 “Robust Face Detection Using Local Gradient Patterns and Evidence Accumulation” Pattern Recogn. 45 (9) 3304 - 3316    DOI : 10.1016/j.patcog.2012.02.031
Ban K.-D. 2011 “Tiny and Blurred Face Alignment for Long Distance Face Recognition” ETRI J. 33 (2) 251 - 258    DOI : 10.4218/etrij.11.1510.0022
Wu B. “Fast Rotation Invariant Multi-view Face Detection Based on Real AdaBoost” IEEE Int. Conf. Autom. Face Gesture Recogn. Seoul, Rep. of Korea May 17–19, 2004 79 - 84