Baixe o app para aproveitar ainda mais
Prévia do material em texto
Accuracy Improvement of A Province Name Recognition on Thai License Plate Somkheart Kraisin Information and Communication Technology for Embedded Systems Sirindhorn International Institute of Technology Email: m592204040646@studentmail.siit.tu.ac.th Natsuda Kaothanthong School of Management Technology Sirindhorn International Institute of Technology Email: natsuda@siit.tu.ac.th Abstract—License plate recognition is mechanism for extract- ing information on a license plate and converting to an encoded text. The information on license plate is letters, number, and province’s name. Although, the license plate recognition for Thai has been introduced for more than ten years, it has a low accuracy rate in the province part. This is due to a low resolution of the province, which has approximately 15-18 pixels of the height. This research aims to improve the accuracy of the province name recognition on Thai license plate. The experiments are conducted on images that were captured from the South of Thailand. A two step classification is proposed to improve accuracy. Given an image of a province’s part, it is classified according to the length of the name. Then, the HOG feature is extracted for predicting the province’s name using a classifier. The proposed method achieve 90% accuracy, which is higher than classifying using only Extreme Learning Machine and Decision Tree. Index Terms—OCR, Thai License Plate Recognition, HOG, Extreme Learning Machine I. INTRODUCTION Thai License Plate Recognition is a technique that is used to convert information on image to encoded text [1] [8] [10] [7] [2] [14] [15]. Information on license plate can be used to identify car ’s owner. It has been utilized in many applications such as car park system, vehicle detection system, car monitoring etc. In the South of Thailand, police have been used license plate recognition software to preventing the car bomb. Many researches were proposed for recognizing the license plate of Japan [8] [10], Taiwan [7], Brazil [2], China [14], United Kingdom [15]. However, such methods cannot be implemented in Thailand because of the different layout and the language. Accuracy rate of license plate recognition in Thailand is limited due to the similarity of Thai characters. Although many researches were proposed to recognize Thai characters [1] [12], it cannot uniquely recognized characters as shown in Fig 1. Moreover, the characters’ height of the province is approximately 15-18 pixels, which is very small as compared to the size of the license plate. Therefore, the accuracy of the province’s name recognition of the current method [1] is limited. Fig. 1: Pair of Similar Characters in Thai language To recognize each characters on province portion in Thai license plate is practical due to the low resolution of the province’s part. R. Juntanasub et.al. [11] proposed a method for recognizing the province’s name using Hausdorff Distance to compare the similarity of the two names. Although, it achieved an accuracy of 92%, the recognition time is high. S. Subhadhira et.al. [1] applied a HOG feature and an Extreme Learning Machine for a recognition. It has been reported that the accuracy is 95.05%. However, the accuracy drops to 60.83% when utilizing with the images of a real situation. In this research, a method to improve an accuracy for recognizing the province’s name of images captured form the real situation is proposed. The overview of the method is depicted in 2. Given a set of images of the province’s part, it is classified using the length of the image into N groups. See, S1, S2, and S3 as in Fig.2. For each group, a classifier is built for recognizing the province’s name. The paper is organized as follow: the previous work and the preliminary are presented in Section II and III, respectively. The proposed method can be found in Section IV. The exper- imental result and the conclusion are discussed in Section V and VI, respectively. II. PREVIOUS WORK Methods were proposed for recognizing license plates of different countries: Japan [8] [10], Taiwan [7], Brazil [2], China [14], and United Kingdom [15]. C. Lin et.al. [7] proposed a technique that applied a SVM and HOG to detect the license plate in an image. A Convolution Neural Networks was used to recognize English characters and Arabic numbers Images of Province Part Length Classification S1 S2 S3 Image Classification Image Classification Image Classification Output Fig. 2: Overview of the proposed method. on license plate in Taiwan. There was a misclassification of the similar characters such as “B” and “8”, “U” and “0”. K. Miyamoto et.al. proposed a Japan license plate recognition using an image analysis [10]. Template matching is applied to recognize the segmented characters on the license plate. S. Subhadhira et.al. [1] proposed a license plate recognition for Thai using an Extreme Learning Machine. Given an input image of a Thai license plate, it is segmented into lower and upper part. The upper part is divided into two sub- parts: a series of letters and numbers. The lower part contains a province name. Each part is binarized using an image processing tools. Then, the feature of each part of the license plate is extracted using Histogram of Gradient (HOG) [4] [13]. The classifier called Extreme Learning Machine that is applied to recognize each part, i.e. a series of letters, numbers, and a province name. The accuracy rate of recognition on province name in research is 95.05%. P. Sa-ngamuang [10] proposed technique to recognize Thai characters on a license plate by using an Essential-Elements-Based technique. R. Juntanasub and N. Sureerattanan [11] divided an image into 5 × 5 blocks. The intensity of black color in each block is used as a representation. Hausdorff Distance Technique to measure similarity of the two images. III. PRELIMINARY STUDY A. Thai License Plate Recognition Structure of Thai license plate is composed of three parts: letter series, number, and province name as shown in Fig 3. The combination of the three parts is used to represent a unique car in Thailand. In other words, two cars can have the same combination of the letter series and the number. However, the two cars can be differentiate using the province’s name. In this way, the accuracy of the province’s name recognition on the license plate plays an important role. B. License Plate Dataset The dataset of the license plate images was from the car check point at Songkla province. Only the images that were taken during the day time were used in this work. Each license plate is manually cropped from image as shown in Fig 4. The segmented image is binarized and rotated before segmenting province part by using the following method. Fig. 3: Thai License Plate Structure Fig. 4: Dataset Preparation To segment upper and lower part from a binarized image as shown in Fig 5, a vertical and a horizontal projection are applied as shown in Fig 6 and Fig 7, respectively. 1) Vertical Projection: It is a technique that is used to separate border frame of license plate in vertical by using intensity that can be calculated as follow equation. Pv(j) = j=H∑ j=0 f(xj , yj) (1) where f(xj , yj) represents intensity, H represents height of image. 2) Horizontal Projection: To separate border in horizontal direction, horizontal projection is applied by using equation below. Ph(j) = j=w∑ j=0 f(xj , yj) (2) where f(xj , yj) is intensity, w represents width of image. 3) Border position: Let PV be intensity in vertical projec- tion, border position of image can be calculated by follow equation. dv[i] = Pv(i+ 1)− Pv(i) (3) Maximum and minimum of dv represent border position of license plate image. dh[i] = Ph(i+ 1)− Ph(i) (4) Let ma represents maximum and mi represents minimum. dv[ma] and dv[mi] are position border in image. Fig. 5: Binary License Plate Image Fig. 6: Vertical and Horizontal Projection Fig. 7: Province Segmentation C. Histogram of Gradients Histogram of Gradients isa feature that was proposed for scene recognition. It is used to represent importance information of images and was used in character recognition [12] [16] [17].To compute HOG. It can be divided into 5 steps as follows. 1) Each image is resized to 160 × 32 pixels and converted to gray-scale. 2) Images are computed gradients in X and Y direction. 3) Sliding window size 2 × 2 from left to right and top to bottom of image, to move window it is moved 1 cell/step, it is overlap with last cell 50%, each cell has 8 × 8 blocks.Each cell is computed Histogram of gradients into 9 bins. 4) Divide each value by k k = √√√√i=N∑ i=1 Hv(i) + eps (5) where eps=1.5× 10−10, Hv represents HOG vector, N represents length of vector. 5) Create final vector by changing 2-D vector to 1-D vector. D. Classifier Extreme Learning Machine (ELM) is a feed forward net- works. ELM consists of an input layer, a hidden layer, and an output layers. Decision tree is a tree-like structure algorithm that is widely used in Machine Learning. Each node of decision tree is used to represent attribute, each branch represents outcome of the nodes and each leaf node is used to represent class label. IV. PROPOSED METHOD From our preliminary experiment using HOG feature and ELM classifer, we found a misclassification of Bangkok and Pattalung. From our observation, Bangkok and Pattalung province have different length. In this way, a length is used to filter the province name before classifying the province name. For an illustration see Fig. 2. A. Length Classification 1) Computing Image’s Length: Given an binarized im- age of the province part on a license plate, which was segmented using a horizontal and a vertical projection. Let L = {L1, . . . , Ln} be the length of the n images in the dataset where Li for i = {1, . . . , n}. The length Li of an image i can be computed as follow: Li = x=end−1∑ x=start Gi(x, y) (6) where Gi(x, y) = { 1 if x < end, (y = 0) 0 otherwise start represents start position of province’s characters, end represents last position of province’s characters on province portion. Gi represents province image in training data. 2) Filtering image’s length: Given the length L of the n images, they are grouped into b groups using a histogram. In this work, b = {3, 4} are used to group the images of the province part. Let I be a length interval, which can be computed as follow: I = Lmax − Lmin b , (7) where Lmax is the maximum length and Lmin is the minimum length of the images. Given an interval I and b, the lower value of an interval of each group is computed using the following equation: Ijmin = Lmin × j where { j = 1..b } (8) The upper value of an interval of each group is computed using the following equation: Ijmax = I j min + I where { j = 1..b } (9) where Ijmin is the lower value and I j max is the upper value of the group j for j = {1..b } . To assign each image into a group Kj for j = {1..b } , each image Gi in the dataset is assigned to the group Kj if the length is between [Ijmin, I j max). B. Province Name Recognition To train model, each image in a group Kj for j = {1..b } are resize to 160 × 32 pixels. A HOG feature is extracted to represent an image. Let Hi represents HOG of an image Gi, Hi is used to train a recognition model. In this work, an Extreme Learning Machine (ELM) and Decision Tree are used. Images of Province Part Length Classification S1 S2 S3 Image Classification Image Classification Image Classification Output Fig. 8: Overview of the province’s name recognition. C. Recognize Province Name An overview of the proposed province name recognition is shown in Fig. 8. Given an testing image of the province part T , it is classified by its length. If the length of T falls between the interval [Ijmin, I j max), the image T is in the group j. From the overview in Fig. 8, the length of the image T is in group 2. Then, T is resized to 160 × 32 pixels before extracting a HOG feature, denoted by HT . Given a feature HT and the group j, the province name recognition model of the group j is used to recognize the province’s name. V. EXPERIMENT In this experiment, the image dataset is collected from real situation, which were taken in the South of Thailand. The dataset composes of 5 provinces (classes). For each class, there are 69 images where 45 are used for training and the rest is for testing. Examples of the images in each class is shown in TableI. From the examples, the length of ‘Bangkok’ class is the longest while ‘Pattalung’ is the shortest. TABLE I. Example of images in each class. Bangkok Narathivat Pattalung Pattani Songkla A. Experimental Setting To obtain the number of group for classifying the images’ length, an experiment is conducted in order to show the class distribution. The distribution of the province images in each group for b = 3 and b = 4 are shown in Table III and Table V, respectively. Examples of images in each bin using b = 3 and b = 4 are shown in Table IV and Table VI, respectively. It can be seen from the result in Table V that b = 4 gives better distribution since the majority of each group falls in TABLE II. Example of testing images in each class. Bangkok Narathivat Pattalung Pattani Songkla different group. On the other hand, for b = 3 the majority of the provinces which have similar length are in the first bin. TABLE III. The distribution of the length of the province images using 3 groups. Group Bangkok Narathiwat Pattalung Pattani Songkla 1st 3 36 45 44 45 2nd 41 7 0 1 0 3rd 1 2 0 0 0 TABLE IV. Example of images when using b = 3. Group Example of Images 1st Bin 2nd Bin 3rd Bin TABLE V. The distribution of the length of the province images using 4 groups. Group Bangkok Narathiwat Pattalung Pattani Songkla 1st 0 0 35 17 36 2nd 0 13 9 22 9 3rd 7 27 1 5 0 4th 38 5 0 1 0 TABLE VI. Example of images in each bin when using b=4. Group Example of Images 1stBin 2nd Bin 3rd Bin 4th Bin B. Experimental Result The experiments are conducted using two classification: 1) without applying the length classification and 2) the one with the length classification. For the experiments using the length classification, they are conducted using two different size of b, i.e., b = {3, 4}. The summary of the experimental result is shown in Table VII. It can be seen from the result that the length classification can improve the classification accuracy. Comparing the size of TABLE VII. Accuracy Comparison of Recognition Method Classification Method Accuracy (%) Extreme Learning Machine without length classification 60.83 Decision Tree without length classification 60.83 Extreme Learning Machine and b = 3 histogram 83.33 Decision Tree and b = 3 histogram 66.67 Extreme Learning Machine and b = 4 histogram 90.00 Decision Tree and b = 4 histogram 70.00 TABLE VIII. Confusion Matrix of Decision Tree without length classification. Class Bangkok Narathiwat Pattalung Pattani Songkla Accuracy Bangkok 13 0 1 9 1 54.17% Narathivat 1 14 2 3 4 58.33% Pattalung 1 1 18 3 1 75.00% Pattani 0 5 4 13 2 54.17% Songkla 1 3 4 2 14 58.33% Average accuracy 60.83% the bins, using b = 4 achieves a better result. The detail of the experimental result can be found in the following subsection. 1) Without Length Classification: In this experiment, the province’s images are classified using the two methods, i.e., Decision Tree and Extreme Learning Machine. Results can be found in Table VIII and IX, respectively. From the recognition result, there are misclassified im- ages. For example, Songkla is recognized as Narathivat and Bangkok is recognized as Pattani. Regarding to the examples in Table I, the length of Bangkok is much longer than Pattani. 2) With Length Classification: To improve the classification as shown in Table VIII and Table IX, the images are classified according to their length using a histogram. The classification result using b = 3 histogram with the Decision Tree and the Extreme Learning Machine are shown in Table X and TableXI, respectively. It can be seen that accuracy of the classification using the Extreme Learning Machine is improved from 60.83% to 83.33%. On the other hand, there are a number of misclassified of Songkla when using the Decision Tree. Table XII and Table XIII shows the classification result using b = 4 histogram. The accuracy of the Decision Tree classifier improves from 60.83% to 70% when using b = 4 bins. The misclassification classes from ‘Bangkok’ to ‘Pattani’ is improved from 9 to 1. See, Table X and Table XII for comparison. The accuracy using Extreme Learning Machine is 90%, which is better than the Decision Tree. The classification result TABLE IX. Confusion Matrix of Extreme Learning Machine without length classification. Class Bangkok Narathiwat Pattalung Pattani Songkla Accuracy Bangkok 12 3 4 4 1 50.00% Narathiwat 2 18 1 2 1 75.00% Pattalung 0 2 20 1 1 83.33% Pattani 2 3 4 12 3 50.00% Songkla 3 6 3 2 11 45.83% Average accuracy 60.83% TABLE X. Confusion Matrix of Decision Tree using b = 3 bins. Class Bangkok Narathiwat Pattalung Pattani Songkla Accuracy Bangkok 20 3 0 1 0 83.33% Narathivat 0 14 3 3 4 58.33% Pattalung 0 0 19 3 2 79.17% Pattani 0 0 4 11 9 45.83% Songkla 0 3 2 3 16 66.67% Average accuracy 66.67% TABLE XI. Confusion Matrix of Extreme Learning Machine using b = 3 bins. Class Bangkok Narathiwat Pattalung Pattani Songkla Accuracy Bangkok 24 0 0 0 0 100.00% Narathivat 0 21 0 1 2 87.50% Pattalung 0 2 20 2 0 83.33% Pattani 0 3 4 17 0 70.83% Songkla 0 4 2 0 18 75.00% Average accuracy 83.33% for each class is better that using b = 3 histogram. The result of ‘Songkla’ province is similar to the one from b = 3 histogram. This is because its length was classified in the same bins as ‘Pattalung’ and ‘Pattani’. Comparing the result between b = 3 and b = 4, the one using b = 4 gives a better result. VI. CONCLUSION Thai License Plate recognition still has low accuracy to recognize province name, the main reason is province portion has low resolution. This research aims to improve accuracy rate by using two step classification. First step, the image is classified by filtering image ’s length, second step, image is classified by using classifier. From the experiment, classifying image into N bins before recognizing can improve accuracy rate. In this research, Recognition province name by using Extreme Learning Machine obtains accuracy at 90%. TABLE XII. Confusion Matrix of Decision Tree using b = 4 bins. Class Bangkok Narathiwat Pattalung Pattani Songkla Accuracy Bangkok 19 4 0 1 0 79.17% Narathivat 2 20 0 2 0 83.33% Pattalung 0 2 14 3 5 58.33% Pattani 0 5 2 14 3 58.33% Songkla 0 2 2 3 17 70.83% Average accuracy 70% TABLE XIII. Confusion Matrix of Extreme Learning Machine using b = 4 bin. Class Bangkok Narathiwat Pattalung Pattani Songkla Accuracy Bangkok 24 0 0 0 0 100.00% Narathivat 0 22 0 2 0 91.67% Pattalung 0 2 22 0 0 91.67% Pattani 0 0 1 22 1 91.67% Songkla 0 0 3 3 18 75.00% Average accuracy 90% ACKNOWLEDGEMENT This work was supported by the SIIT Young Researcher Grant, under contract no. SIIT 2017-YRG-NK04. The authors would like to thank the anonymous reviewers for their valuable comments and suggestions to improve the quality of the paper. REFERENCES [1] S. Subhadhira U. Juithonglang ,“License Plate Recognition Application Using Extreme Learning Machines ”,2014 Third ICT International Student Project Conference (ICT-ISPC),2014,p.103-106. [2] F. Delmar Kurpiel and R. Minetto and B. T. Nassu,”Convolutional neural networks for license plate detection in images”,2017 IEEE International Conference on Image Processing (ICIP),2017,pp.3395-3399. [3] Y. Hou and Z. Ye and W. Xu and L. Ma, “Remote sensing textual image classification based on extreme learning machine and hybrid rice optimization algorithm”,2017 9th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS),2017,p.777-781. [4] S. Routray and A. K. Ray and C. Mishra, “Analysis of various image feature extraction methods against noisy image: SIFT, SURF and HOG”, 2017 Second International Conference on Electrical, Computer and Communication Technologies (ICECCT),2017,p.1-5. [5] W. Kusakunniran, K. Ngamaschariyakul, C. Chantaraviwat, K. Janvit- tayanuchit and K. Thongkanchorn, “A Thai license plate localization using SVM,” 2014 International Computer Science and Engineering Conference (ICSEC), Khon Kaen, 2014, pp.163-167. [6] David C. Lambert, Extreme Learning Machine implementation in Python, (2013), GitHub repository, https://github.com/dclambert/Python- ELM [7] C. H. Lin and Y. S. Lin and W. C. Liu , “An efficient license plate recognition system using convolution neural networks”,018 IEEE Inter- national Conference on Applied System Invention (ICASI),2018,pp.224- 227. [8] K. Miyamoto and K. Nagano and M. Tamagawa and I. Fujita and M. Ya- mamoto, “Vehicle license-plate recognition by image analysis”,Industrial Electronics, Control and Instrumentation, 1991. Proceedings. IECON ’91, 1991 International Conference,1991,pp.1734-1738 vol.3. [9] P. Sa-ngamuang and C. Thamnittasana and T. Kondo,“Thai Car License Plate Recognition Using Essential-Elements-Based Method”,2007 Asia- Pacific Conference on Communications,2007,pp.41-44. [10] K. Taniyama and K. Hayashi, “Robust car License Plate Recognition system verified with 163, 574 images captured in fields”,Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012),2012,pp.1273-1276. [11] R. Juntanasub and N. Sureerattanan, “Car license plate recognition through Hausdorff distance technique”,17th IEEE International Confer- ence on Tools with Artificial Intelligence (ICTAI’05),2005,pp.5-9. [12] T. Siriteerakul,“Mixed Thai-English Character Classification Based on Histogram of Oriented Gradient Feature”,2013 12th International Con- ference on Document Analysis and Recognition,2013,pp.847-851. [13] T. Surasak and I. Takahiro and C. h. Cheng and C. e. Wang and P. y. Sheng,“Histogram of oriented gradients for human detection in video”,2018 5th International Conference on Business and Industrial Research (ICBIR),2018,pp.172-176. [14] Q. Wang, “License plate recognition via convolutional neural net- works”,2017 8th IEEE International Conference on Software Engineer- ing and Service Science (ICSESS),2017,pp.926-929. [15] X. Zhai and F. Bensaali and S. Ramalingam, “Real-time license plate localisation on FPGA”,CVPR 2011 WORKSHOPS,2011,pp.14-19. [16] S. Iamsa-at and P. Horata, “Handwritten Character Recognition Using Histograms of Oriented Gradient Features in Deep Learning of Artificial Neural Network”,2013 International Conference on IT Convergence and Security (ICITCS),2013,pp.1-5. [17] A. Tikader and N. B. Puhan, “Histogram of oriented gradients for English-Bengali script recognition”,International Conference for Con- vergence for Technology-2014,2014,pp.1-5.
Compartilhar