publication ravi

Upload: ravi-theja

Post on 04-Jun-2018

218 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/13/2019 Publication Ravi

    1/4

    License Plate Extraction Using Adaptive Threshold and Line Grouping

    A. Ravi Theja, Shashank Jain, Abhishek Aggarwal and V. Krishna Rao Kandanvli,Member,IEEE

    Department of Electronics and Communication Engineering

    Motilal Nehru National Institute of TechnologyAllahabad, India

    [email protected], [email protected], [email protected], [email protected]

    AbstractLicense plate recognition system is an essential toolfor the Traffic Police Department and so, is widely being used

    for road traffic surveillance in many countries. This paper

    aims at locating the license plate in any given traffic

    surveillance image, by detecting the rectangle like shapes in

    that image. Line grouping algorithm extracts these shapes in

    the input image to locate the license plate. But its accuracy falls

    sharply for images where the license plate boundary suffers

    from occlusions leading to broken boundaries. The use of

    adaptive threshold and contours, instead of canny edgedetector, to extract the line segments greatly increases the

    accuracy of line grouping algorithm, which is also verified by

    the results.

    Keywords-adaptive threshold; contours; license plate

    extraction; line extraction; line grouping; occlusion

    I. INTRODUCTIONThe importance of automatic license plate recognition

    systems is greatly increasing due to the heavy accretion inthe number of motor vehicles every year and such systemshave already been installed in many countries including UK,

    US and Korea. UK has an extensive automatic number platerecognition (ANPR) CCTV network and the police arecapable of tracking any car in the whole country using thisnetwork.

    The first step in any recognition system is to locate thelicense plate in the image or video, as the success ofrecognition is directly dependent on the success of thelocalization of the license plate.

    The difficulties faced by real time recognition systemsare poor image resolution caused by blur due to the distance

    between the camera and the car, poor lighting, low contrastdue to overexposure, reflection or shadows and license plateobscured because of dirt etc. The robustness of any systemdepends on how well it is able to tackle these issues.

    Several approaches have been suggested to deal with theproblems associated with license plate localization. Theseinclude vertical edge density methods [1]-[5], line grouping[4] and morphological operations [5]. Vertical edge densitymethods are reliable when the blur, due to the distance

    between the camera and plate, is low and consequently, thevertical edges in the license plate can be detectedsuccessfully by the edge detector. The shape of the license

    plate is a more reliable feature than the vertical edges, as itis more resistant to blurring. In [4], line grouping has beenemployed initially to detect the presence of license plate likeshapes in the image and upon failure of line grouping,vertical edge density has been used. In line grouping, linesare fitted on the edge segments which are obtained byconnecting the edge pixels detected from a canny edgedetector [10]. These lines, prior to further processing, arefiltered on the basis of their length such that the dimensionsof the license plate lie between their respective thresholds(MinWidth length of plate MaxWidth and MinHeight Height of plate MaxHeight).

    (a) (b)

    Figure 1. Input Image with l icense plate suffering from occlusions and edge image with broken or discontinuous boundaries

  • 8/13/2019 Publication Ravi

    2/4

    The plate boundaries normally suffer from occlusions.When these boundaries are processed according to theabove process, each of the boundary yields two or more thantwo shorter lines instead of a single boundary line (as shownFigure 1). The shorter lines are further filtered out due totheir length and so the plate is not detected in the subsequent

    process.

    In general, there are two stages in the plate recognitionprocess. In first stage, the color image is converted to abinary image. In second stage, in order to obtain the platelocation, the binary image is processed via varioustechniques. The canny edge detector [4] relies on the pixelgradient at the boundaries of the plate region and its outputis prone to occlusions and noise. Whereas adaptivethresholding [6] is not affected by occlusions in the

    boundary of the plate region as it relies on the contrast inpixel values rather than the pixel gradient. The advantage ofusing adaptive threshold is to eliminate unimportant regionsin the image. This will reduce the number of lines extractedso that the next stage has lesser data to process. In [6], apart

    from adaptive thresholding, horizontal scanning and verticalprojection methods have been used to extract the licenseplate.

    In this paper, we have used adaptive thresholding toextract the line segments. By line grouping algorithm [4], theline segments are grouped to locate the license plate. Theresults obtained are compared with [4] to illustrate theeffectiveness of using adaptive thresholding for lineextraction. This paper is organized as follows. In Section II,the line extraction and line grouping is explained. Section IIIgives the experiments and results and Section IV concludesthis paper.

    II. LINE EXTRACTION AND GROUPINGA. Line Extraction

    The input color image is first converted to grayscaleimage by using the method described in [11]. The grayscaleimage is processed with adaptive thresholding process [7].The mask size is such that it never fits inside the plate andalways has a considerable portion outside the plate region inall positions. The threshold value used in adaptivethresholding is calculated as explained in [7].

    The contours in the binary image are extracted using themethod described in [8] and then they are approximated by

    polygons [9]. While approximating the contours withpolygons, the parameter, specifying the distance of the mostdistant point from the accumulated approximation, is kept afraction of the total length of the contour. The sides of the

    polygon are extracted and further processed in stage B.

    B. Line Grouping [4]This stage extracts line segments and preserves only the

    ones satisfying the pre-specified conditions [4] (see [4,

    conditions X1 to X11]). The thresholds have been set foreach camera configuration.

    III. EXPERIMENTS AND RESULTSThe presented adaptive threshold and line grouping

    algorithm is coded in C using Intels OpenCV library and itis tested on a 2.8 GHz Pentium D processor with 512Mb

    RAM. The method, described in [4], is also implemented inC and it will run on the same machine. While implementing[4], the line extraction process is altered by extractingcontours from the binary image returned from canny edgedetector [8]. The sides of the polygon are extracted by usingthe polygon approximation [9] of these contours. Theextracted sides are then processed by applying the conditionsin [4]. For comparing the presented method with [4], a set of40 images are selected where the boundaries of the license

    plate suffered from occlusions or blurring. The results areshown in Table I.

    Table I COMPARISON OF RESULTS

    MethodNumber ofImagesTested

    CorrectRate

    Average numberof Rectanglegroups Detectedin X11

    RejectionRate

    Linegroupingwith Canny

    40 37.5% 11 62.5%

    LinegroupingwithAdaptiveThresholding

    40 87.5% 2 12.5%

    It is observed that the thresholds for line grouping withcanny edge detector are broader as compared to thethresholds with adaptive thresholding. The reason is that thelines extracted by canny, obtained from the broken

    boundaries of the plate, are shorter than the originalboundaries of the license plate. Therefore, the thresholdsMinWidth, MaxWidth, MinHeight and MaxHeight are set tobe broader. Otherwise, the plate is not detected, as the linesmaking the boundary are filtered out in the initial stage dueto condition X2 and X4 in [4]. Due to the broader thresholdsset for the method in [4], the number of rectangle groupsformed by condition X11 is more as compared to the

    presented method.

    IV. CONCLUSIONThe accuracy of line grouping algorithm, reported in [4],

    is 98%, but it is reduced significantly when applied onoccluded plate boundaries or blurred images. The reason forthe loss in performance can be traced to the noncontiguous

    boundaries returned by canny and the broader thresholds setto compensate for the broken boundaries. The use ofadaptive threshold and contour method improves the

    performance of the line grouping method as it gives

  • 8/13/2019 Publication Ravi

    3/4

    contiguous plate boundaries to process. Application ofadaptive threshold on the input image in figure 1 results intoan image shown in figure 5. With the help of presentedalgorithm, the license plate has been detected for imagesshown in Figures 2, 3 and 4.

    REFERENCES

    [1] P. Rattanathammawat and T. H. Chalidabhongse, A Car PlateDetector using Edge Information, Proc. Int. Symp. Communicationsand Information Technologies (ISCIT06), IEEE Press, 2006, pp.1039-1043, doi: 10.1109/ISCIT.2006.339936.

    [2] A. M. Al-Ghaili, S. Mashohor, A. Ismail, and A. R. Ramli, A NewVertical Edge Detection Algorithm and its Application, Proc. Int.Conf. Computer Engineering & Systems (ICCES08), IEEE Press,2008, pp. 204-209, doi: 10.1109/ICCES.2008.4772997

    [3] Jian-Feng Xu, Shao-Fa Li and Mian-Shui Yu, Car License PlateExtraction using Color and Edge, Proc. Int. Conf. Machine Learningand Cybernetics, IEEE Press, August 2004, vol. 6, pp.3904 - 3907.

    [4] Gisu Heo, Minwoo Kim, Insook Jung, Duk-Ryong Lee and Il-SeokOh, Extraction of Car License Plate Regions Using Line Groupingand Edge Density Methods, Proc. Int. Symp. Information

    Technology Convergence (ISITC07), IEEE Press, November 2007,pp. 37-42, doi: 10.1109/ISITC.2007. 79

    [5] F. Faradji, A.H. Rezaie and M. Ziaratban, A Morphological-BasedLicense Plate Location, Proc. Int. Conf. Image Processing (ICIP07),IEEE Press, SeptemberOctober 2007, vol. 1, pp. I-57-I-60,doi:10.1109/ ICIP.2007.4378890

    [6] Liu Ying and Li Nannan, Design of License Plate RecognitionSystem Based on the Adaptive Algorithm, Proc. Int. Conf.Automation and Logistics (ICAL08), IEEE Press, September 2008,

    pp. 2818-2821, doi: 10.1109/ICAL.2008.4636655

    [7] A. Jain, Fundamentals of Digital Image Processing, EnglewoodCliffs, NJ: Prentice-Hall, 1986.

    [8] S. Suzuki and K. Abe, Topological structural analysis of digitalbinary images by border following, ScienceDirect Computer Vision,Graphics and Image Processing, vol. 30, April 1985, pp. 32-46, doi:10.1016/0734-189X(85)90016-7 .

    [9] D. Douglas and T. Peucker, Algorithms for the reduction of thenumber of points required to represent a digitized line or itscaricature, Cartographica, December 1973, vol. 10, pp. 112122,doi:10.3138/ FM57-6770-U75U-7727.

    [10] J. Canny, A computational approach to edge detection, IEEE Trans.on Pattern Analysis and Machine Intelligence, November 1986, vol.8, pp. 679698, doi: 10.1109/TPAMI.1986.4767851

    [11] W. Wharton and D. Howorth, Principles of Television Reception,London: Pitman, 1971.

    .

    (a) (b)

    Figure 2. Input Image and its corresponding output

    (a) (b)

    Figure 3. Input Image and its corresponding output

  • 8/13/2019 Publication Ravi

    4/4

    (a) (b)

    Figure 4. Input Image and its corresponding output

    (a) (b)

    Figure 5. Input Image in figure 1 on applying adaptive threshold and the contours extracted from the corresponding image