Automatic License Plate Recognition Using Python And OpenCV

Automatic License Plate Recognition using Pythonand OpenCVK.M. SajjadDepartment of Computer Science and EngineeringM.E.S. College of Engineering, Kuttippuram, Keralame@sajjad.inAbstract—Automatic License Plate Recognition system is a realtime embedded system which automatically recognizes the licenseplate of vehicles. There are many applications ranging fromcomplex security systems to common areas and from parkingadmission to urban traffic control. Automatic license plate recognition (ALPR) has complex characteristics due to diverse effectssuch as of light and speed. Most of the ALPR systems are builtusing proprietary tools like Matlab. This paper presents an alternative method of implementing ALPR systems using Free Software including Python and the Open Computer Vision Library.Keywords—License plate, Computer Vision, Pattern Recognition, Python, OCR.Fig. 1.Proposed SystemI. I NTRODUCTIONThe scientific world is deploying research in intelligenttransportation systems which have a significant impact onpeoples lives. Automatic License Plate Recognition (ALPR)is a computer vision technology to extract the license numberof vehicles from images. It is an embedded system which hasnumerous applications and challenges. Typical ALPR systemsare implemented using proprietary technologies and hence arecostly. This closed approach also prevents further researchand development of the system. With the rise of free andopen source technologies the computing world is lifted tonew heights. People from different communities interact ina multi-cultural environment to develop solutions for mansnever ending problems. One of the notable contribution ofthe open source community to the scientific world is Python.Intel’s researches in Computer Vision bore the fruit calledOpen Computer Vision (OpenCV) library, which can supportcomputer vision development.II. P ROPOSED S YSTEMIn India, basically, there are two kinds of license-plates,black characters in white plate and black characters in yellowplate. The former for private vehicles and latter for commercial, public service vehicles. The system tries to address thesetwo categories of plates. The high-level block diagram of theproposed system is shown in Fig. 1.A. CaptureThe image of the vehicle is captured using a high resolutionphotographic camera. A better choice is an Infrared (IR)camera. The camera may be rolled and pitched with respectto the license plates.Fig. 2.Captured Image (Original image)Character recognition is generally very sensitive to theskew. The readable characters can become distorted due tothe obliqueness of the camera. Using a better camera withmore definition and resolution will increase the success ratioof the system.B. PreprocessPreprocessing is the set algorithms applied on the imageto enhance the quality. It is an important and commonphase in any computer vision system. For the present systempreprocessing involves two processes: Resize – The imagesize from the camera might be large and can drive thesystem slow. It is to be resized to a feasible aspect ratio.Convert Colour Space – Images captured using IR or photographic cameras will be either in raw format or encoded intosome multimedia standards. Normally, these images will bein RGB mode, with three channels (viz. red, green and blue).

through the image and find the connected pixels. Each of theconnected components (blobs) are labelled and extracted. Fig.4 shows the filtered blobs.E. SegmentationFig. 3.Localization (Thresholded image)Segmentation is the process of cropping out the labelledblobs. These blobs are expected to be the required portionof the license number. A special algorithm called ImageScissoring[1] is introduced here. In this algorithm, the licenseplate is vertically scanned and scissored at the row on whichthere is no white pixel and the scissored area is copied into anew matrix, as in Fig. 5.Number of channels defines the amount colour informationavailable on the image. The image has to be converted tograyscale.C. LocalizeRear or front part of the vehicle is captured into animage. The image certainly contains other parts of thevehicle and the environment, which are of no requirementto the system. The area in the image that interests usis the license plate and needs to be localized from thenoise. Localization is basically a process of binarizing theimage. As shown in Fig. 3, the image is converted to blackand white. There are two motivations for this operation –1. Highlighting characters and 2. Suppressing background.Fig. 5.Segmented blobsThere are unwanted blobs even after segmentation. Theseare classified using special algorithms.Localization is done by an image processing techniquecalled Thresholding. The pixels of the image are truncated totwo values depending upon the value of threshold. Thresholdrequires pre-image analysis for identifying the suitablethreshold value. Adaptive thresholding technique determinesa local optimal threshold value for each image pixel so as toavoid the problem originating from non-uniform illumination.Fig. 6.D. Connected Component AnalysisClassified BlobsF. Character RecognitionFinally, the selected blobs are send to a Optical CharacterRecognition (OCR) Engine, which returns the ASCII of thelicense number.III. T OOLSA. PythonFig. 4.Connected Components (Blobs)In order to eliminate undesired image areas, a connectedcomponent algorithm is first applied to the binarized platecandidate. Connected component analysis is performed toidentify the characters in the image. Basic idea is to traversePython is a remarkably powerful dynamic, object-orientedprogramming language that is used in a wide variety ofapplication domains. It offers strong support for integrationwith other languages and tools, and comes with extensivestandard libraries. To be precise, the following are somedistinguishing features of Python: Very clear, readable syntax. Strong introspection capabilities. Full modularity. Exception-based error handling.

High level dynamic data types.Supports object oriented, imperative and functional programming styles. Embeddable. Scalable MatureWith so much of freedom, Python helps the user to thinkproblem centric rather than language centric as in other cases.These features makes Python a best option for scientificcomputing. B. Preprocess As seen before, preprocessing involves resizing and changing colour spaces of the sourceimage. Like any other image processing toolkits,OpenCV also provides fast and quick procedures.B. OpenCVOpenCV is a library of programming functions for realtime computer vision originally developed by Intel and nowsupported by Willogarage. It is free for use under the opensource BSD license. The library has more than five hundredoptimized algorithms. It is used around the world, with fortythousand people in the user group. Uses range from interactiveart, to mine inspection, and advanced robotics. The libraryis mainly written in C, which makes it portable to somespecific platforms such as Digital Signal Processor. Wrappersfor languages such as C, Python, Ruby and Java (usingJavaCV) have been developed to encourage adoption by awider audience. The recent releases have interfaces for C .It focuses mainly on real-time image processing. OpenCV is across-platform library, which can run on Linux, Mac OS andWindows. To date, OpenCV is the best open source computervision library that developers and researchers can think of.C. TesseractTesseract is a free software OCR engine that was developedat HP between 1984 and 1994. HP released it to the community in 2005. Tesseract was introduced at the 1995 UNLVAnnual Test OCR Accuracy [2] and is currently developedby Google released under the Apache License. It can nowrecognize 6 languages, and is fully UTF8 capable. Developerscan train Tesseract with their own fonts and character mappingto obtain perfect efficiency.IV. I MPLEMENTATIONOpenCV with Python is a deadly combination for computervision. We will now revisit each phase in the pipeline and seehow they are implemented using the above tools.A. CaptureStarting from point one of capturing the image, OpenCVlibrary has highly optimized algorithms for all imageprocessing operations. OpenCV provides interface fordifferent camera models. The following code snippet explainshow to interface an in-built web camera and capture a frame.from opencv import highgui as hgcapture hot")frame hg.cvQueryFrame(capture)hg.cvShowImage("Snapshot", frame)Resizeoriginal cv.LoadImageM("image.jpg")thumbnail cv.CreateMat(original.rows/ 10, original.cols / 10, original.type)cv.Resize(original, thumbnail)The original image is resized to the dimensions specifiedin the thumbnail object. Colour space conversionCvtColor(original,gray,CV RGB2GRAY)The above line of code converts the original image togray. More image conversion codes are available at theOpenCV Wiki[3].C. LocalizeThreshold operation is performed in this phase. Toretain the image quality, adaptive threshold algorithms areto be used. Previous researches[4] have concluded thatOtsu’s[5] thresholding algorithm is the efficient way ofbinarizing the image. OpenCV provides complex and efficientadaptive thresholding algorithms including Otsu method.cvThreshold(image, binary image,128,255,CV THRESH OTSU)The above line of code returns a binary image whichis adaptively thresholded. The arguments follow the order:1. Source image, 2. Destination image, 3. Threshold value,4. Resultant value, and 5. Type of threshold. The typeCV THRESH OTSU performs Otsu algorithm on the sourceimage.D. Connected Component AnalysiscvBlobsLib is a library to perform binary images connectedcomponent labelling. It also provides functions to manipulate,filter and extract results from the extracted blobs. The libraryprovides two basic functionalities: Extract 8-connected components in binary or grayscaleimages. Filter the obtained blobs to get the interest objects in theimage. This is performed using the Filter method objectsperthreadareused.myblobs CBlobResult(binary image, mask,0, True)myblobs.filter blobs(325,2000)blob count myblobs.GetNumBlobs()The connected components are labelled using the abovecode snippet. filter blobs method is used to filter outthe blobs of required dimensions.