GENERAL DESCRIPTION

VEDA LPR represents our current software solution in the domain of vehicle license plate recognition (LPR), being intended to automatically identify and recognize / "read" such license plates in digital images.

The VEDA LPR engine not only represents an (Windows / 32-bit) extension, but almost a completely new design, development and implementation, started from an old (DOS / 16-bit) LPR / ALPR / ANPR approach of ours, namely VEDA CAR (1993-94).

The software may use, as input, both color and grayscale image files (grabbed either by video cameras or by digital photo ones), and raw 256 gray levels images already "prepared" in (a buffer of) the computer memory. It doesn't depend on any specific image acquisition board (frame grabber). Supported image file formats include: JPEG, PNG, PCX, BMP, TIFF and IMG (Imaging Technology 256 gray levels uncompressed format).

VEDA LPR isn't directly meant as a universal and ready-to-run solution, but merely as a foundation on which we/one can build customized ANPR / ALPR specific applications.

Thus, once interfaced with some specific image acquisition chain (camera, frame grabber, triggering devices), possibly a controller for an access device (gate / barrier), and a database management software, the VEDA LPR engine can be easily used to monitorize, control and/or validate the access in restricted areas, public and/or private parking lots or garages, to monitorize highway (or other roads) traffic, red-light or speed limit infringements, to control the input / output at customs, toll collection points, etc.

A VEDA LPR engine SDK / API is available for developers / integrators.

USED METHOD

Same as for VEDA CAR, we designed the new LPR engine as having two main parts: an image segmentation subsystem and a recognition / training (OCR) subsystem. The first one is meant to globally analyze a 256 gray levels input image and to find the area(s) of interest potentially containing alphanumeric characters. Once such an area found, it is converted to a monochrome bitmap, using a smart, adaptive, binarization technique. Some "noise" filters and geometrical distortions (skew, slant, rotation) compensation algorithms are then applied on the obtained black and white bitmap. The next target is to split this bitmap into rows of text (if needed), and these ones into their component characters. Vectors of appropriate parameters are generated starting from these character bitmaps, as internal representations for each one. These vectors are then passed further as input to the second subsystem. At this level, if the purpose is recognition, a powerful neural-like pattern classification algorithm is used to "identify" them, based on (one or more) existing OCR knowledge bases. If the purpose is training a (selected) knowledge base, specific representations (patterns) derived for each character sample, explicitly "tagged" interactively by the user, are "learned". A dedicated user interface is provided for this purpose. Thus, on further recognition sessions, the system's experience is enhanced, and, after a number of training sessions, the knowledge becomes good enough to ensure average recognition rates above 99%.

In order to be as fast as possible, the global image segmentation system (interest area finder) doesn't use traditional image analysis techniques such as Fourier analysis, Sobel or Laplace edge detection methods and the likes. Instead, we developed a "contrast vaults" analysis method and algorithm for providing the most effective results. Briefly, a fast, image-to-"contrast vaults" measure map, transformation is performed on a per-block basis, i.e. a numeric value derived from an average "contrast vaults" measurement is attached to each block. Next, a conventional (histogram-based) method is applied on that map for finding a proper segmentation threshold, meant to isolate "blobs" from the background. Some filtering and connectivity analysis are applied on the initially retained "runs" of blocks in order to obtain the candidate blobs. These latter ones are subject of a blobs filtering process, and only the appropriate ones are kept. Usually, area(s) containing the car license plate proves to be (among) the remained "blob(s)". Once the interest area(s) thus segmentated, a smart, adaptive threshold based binarization (using some statistical methods and several test points) is performed, and black/white resulted bitmap(s) is (are) filtered of "noise", de-skewed/de-slanted, and finally passed further for segmentation to the recognition (or, respectively, to the training) subsystem.

The recognition / training (OCR) subsystem is based on one of our proprietary approaches / technologies in the field of neural and not-neural but somehow alike classifiers. This technology features a high degree of noise tolerance and power of generalization. It practically allows to learn and to further recognize any graphic symbol. It is also used as the basic recognition / training engine for our VEDA OCR/NeurOCR application.

Our training experiments focused on car license plates were performed on several hundreds (even thousands) of color and gray levels images, taken in various conditions and at different resolutions (e.g. 640x480, 768x512, 768x576, 800x600, 1024x768, 1384x1030, and other - both lesser and greater - ones), with quite various framing of the vehicles in the captured scenes also. These vehicles mainly included Romanian (RO), French (F), Italian (I), Dutch (NL), German (D), British (GB), Spanish (E), Slovakian (SK), Czech (CZ), Israeli (IL), Brasilian (BR) and Hong Kong (HK) car license plates, but also American (USA), Mexican (MEX), Belgian (B), Greek (GR), Turkish (TR), Australian (AU), and other ones. This amount of training finally ensured very good average recognition ratios. We emphasize again that the recognition / training system can, in fact, recognize/learn any kind of machine printed text, if properly segmentated within the image. For avoiding possible confusions / misnomers (like '1' with 'I', '0' with 'O' or 'D', '8' with 'B', etc.) in real(-time) applications, a customized list of allowed number formats / syntaxes may be defined and used as a post-recognition automatic correction guide, from case to case, in such specific applications.

HIGHLIGHTS

• Modern, neural-like, trainable OCR enables a great recognition flexibility providing noise tolerance and the ability to cope with various text fonts, positions and dimensions, by means of learning.
• International number plates recognition, with various dimensions and aspect ratios, painted in various colors, both dark/light and light/dark, provided that they contain separable uppercase letters (from a latin-like alphabet) and digits, all of almost equal (comparable) height among each other, placed in one or two overposed rows.
• Strong noise tolerance, required because of various lighting conditions (indoor, outdoor, natural/IR light, uncontrolled reflections, dirt, etc.), and to compensate some of the acquisition process / equipment limitations.
• Great recognition speed (e.g. about 30-500 milliseconds for a 768x512, 8 bits/pixel 256 gray levels image, on an "old" P4 at 1.5GHz PC with 512MB RAM).
• Interactive training performed on a per-case basis from static images, being, practically, a real time process ("train and test").
• Smart (auto-tuneable) contrast "texture" analysis technique for fast and accurate detection/segmentation of the license plate region.
• Automatic compensation of geometrical distortions (license plate skew, slant, and/or rotation, 3D perspective effects).
• Independence on the acquisition equipment (camera and frame grabber), the only restrictions referring to reasonable overall quality of the images and spatial resolution (recommended in the range of 640x480 - 1024x768 pixels), appropriate framing of the vehicle with appropriate camera zoom and focus.

DEMO DOWNLOADING (aprox. 2.96 Mb)

You can download a FREE DEMO application of VEDA LPR (for Windows / 32-bit), from HERE. What you'll get after downloading will be a vedalpr.zip archive file which must be "unzipped" in a (preferably new, empty) directory / folder on your Windows based computer. This "ZIP" file usually contains a readme.txt text file (which is recommended to be read firstly in order to correctly install and run the DEMO), and an automatic installer of all the necessary archived files: install.bat (which must be run in order to get all the required files extracted on your hard disk). Finally, you can use the extracted DEMO application (vedalpr.exe), or the character patterns training tool (vedapatt.exe) if you want to build/update an OCR knowledge base using some images of your own. Please note that, when trying the DEMO, you must first select the image files to test it on, and the OCR knowledge base to be used. The demonstration can then be started to either automatically load and analyze image after image, with an adjustable delay between them (for seeing the segmentation and recognition result for each one), or, manually, step by step, one image (next/previous) at a time.

We shall always appreciate comments and suggestions about our VEDA LPR engine.

Please note that for extended LPR performance testing, some freely downloadable image data sets are available, providing color and grayscale images of various resolutions, containing frontal and rear views of various vehicles with various kinds of international registration plates, taken from various positions and distances and in various real life conditions, e.g.:

• JavaANPR snapshots (data set from O. Martinsky, Brno University of Technology, Slovakia - 2006).
• MediaLab LPR database (data set from I.E. Anagnostopoulos et al., the Multimedia Technology Laboratory of the National Technical University of Athens, Greece - 2015).
• Baza_Slika LPR database (data set from V. Srebric et al., Dept. of Electronics, Microelectronics, Computer and Intelligent Systems of the Faculty of Electrical Engineering and Computing of Zagreb, Croatia - 2003).
• OpenALPR benchmarks (BR and EU LPR images - including the JavaANPR snapshots mentioned above - uploaded on GitHub by M. Hill - 2016).
  

Each and/or all of the above mentioned ones may be used for testing our VEDA LPR Demo.

[ Top of page ]

CONTACT:    Mr. Mihnea VREJOIU