`Neubauer et al.
`
`USOO6553131B1
`(10) Patent No.:
`US 6,553,131 B1
`(45) Date of Patent:
`Apr. 22, 2003
`
`(54) LICENSE PLATE RECOGNITION WITH AN
`INTELLIGENT CAMERA
`
`6,141,620 A * 10/2000 Zyburt et al. ............... 701/117
`6,339,651 B1
`1/2002 Tian et al. .................. 382/105
`
`(75) Inventors: Claus Neubauer, Monmouth Junction,
`N S. Jenn-Kwei Tyan, Princeton,
`
`(73) Assignee: Siemens Corporate Research, Inc.,
`Princeton, NJ (US)
`Subject to any disclaimer, the term of this
`patent is extended or adjusted under 35
`U.S.C. 154(b) by 0 days.
`
`(*) Notice:
`
`* cited by examiner
`Primary Examiner Bhavesh M. Mehta
`Assistant Examiner Seyed Azarian
`(57)
`ABSTRACT
`An intelligent camera System and method for recognizing
`license plates, in accordance with the invention, includes a
`camera adapted to independently capture a license plate
`image and recognize the license plate image. The camera
`includes a processor for managing image data and eXecuting
`a license plate recognition program device. The license plate
`(21) Appl. No.: 09/396,950
`recognition program device includes a program for detecting
`orientation, position, illumination conditions and blurring of
`(22) Filed:
`Sep. 15, 1999
`the image and accounting for the orientations, position,
`(51) Int. Cl." .................................................. G06K 9/00
`(52) U.S. Cl. ........................................ 382,105.382170 E. A.C. R s W.A.RA
`(58) Field of Search/17727s. 159,173.
`S. E. for Segmenting characters depicted in the baseline image by
`218 104 170 169 17 173. 340,907.
`employing a projection along a horizontal axis of the base
`s
`s
`s
`s 93s 9 3. 701/1 17
`line image to identify positions of the characters. A Statistical
`s 1 - s
`classifier is adapted for classifying the characters. The
`classifier recognizes the characters and returns a confidence
`References Cited
`Score based on the probability of properly identifying each
`U.S. PATENT DOCUMENTS
`character. A memory is included for Storing the license plate
`4567,609 A * 1/1986 Metcalf
`recognition program and the license plate images taken by
`382/105
`5,651,075 A
`7/1997 Frazier et al... E. an image capture device of the camera.
`ClCalT . . . . . . . . . . . . . . . . . . . . . .
`5,809,161 A * 9/1998 Auty et al. ................. 382/104
`5,847,661. A 12/1998 Ricci .......................... 340/902
`19 Claims, 5 Drawing Sheets
`
`ITULO COCOS COTUS O
`
`C
`
`C, O Ol
`
`(56)
`
`image capture
`10
`
`Course localization 12
`
`Sub-sample Image
`14
`
`Fine localization
`24
`
`Segmentation 26
`
`Projection profile
`28
`
`Compute vertical edges
`f6
`
`Filter projection profile
`30
`
`Saliency map
`18
`
`Determine segments
`32
`
`Coarse localization
`result/region of interest
`determined
`20
`
`| TH detacho
`Tilt detection and
`refinement of borders
`22
`
`Character normalization
`and classification
`34
`
`Compare with
`rules and codes
`36
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`BOSCH EXHIBIT 1007
`
`Page 1 of 13
`
`
`
`U.S. Patent
`
`Apr. 22, 2003
`
`Sheet 1 of 5
`
`US 6,553,131 B1
`
`
`
`Image Capture
`10
`
`-
`
`COOrse localization 12
`
`Sub-sample image
`14
`
`Fine localization
`24
`
`Segmentation 26
`
`Projection profile
`28
`
`Compute vertical edges
`16
`
`Filter projection profile
`30
`
`Saliency map
`18
`
`Determine segments
`32
`
`Coarse localization
`result/region of interest
`determined
`20
`
`Tilt detection and
`refinement of borders
`22
`
`Character normalization
`and classification
`34
`
`Compare with
`rules and codes
`36
`
`FIG.
`
`Page 2 of 13
`
`
`
`U.S. Patent
`
`Apr. 22, 2003
`
`Sheet 2 of 5
`
`US 6,553,131 B1
`
`Projection on x-axis
`
`12 24 36 48 60 72 84 96 108 120
`
`TP-filter from x-projection
`
`12 24 36 48 60 72 84 96 108 120
`
`Segmentation
`
`12 24 36 48 60 72 84 96 108 120
`
`FG. 2A
`
`255.00
`204.00
`FIG. 2B 53.66
`102.00
`51.00
`
`162.00
`129.60
`FIG. 2C '375
`64.80
`5240
`
`FIG. 2D 666
`
`
`
`FIG. 3
`
`Page 3 of 13
`
`
`
`U.S. Patent
`
`Apr. 22, 2003
`
`Sheet 3 of 5
`
`US 6,553,131 B1
`
`16 x 20 pixels
`Input image
`
`4 x convulution
`5 x 5 kernels
`
`16 x 20 pixels
`
`Subsampling
`
`
`
`8 x 10 pe
`
`fully connected
`204
`100 ...As to Oooooooooooooo of -1
`7
`\ 205
`(t DOOOOOOOOOOOOODDDDDDD-1
`
`
`
`40 Output neurons
`classes; 0-9, A-Z,...,
`
`FG. L.
`
`Page 4 of 13
`
`
`
`U.S. Patent
`
`Apr. 22, 2003
`
`Sheet 4 of 5
`
`US 6,553,131 B1
`
`
`
`Processor
`303
`
`
`
`
`
`
`
`External system
`304
`
`Comera 300
`
`License plate recognition
`device program 312
`
`Statistical classifier
`314
`
`FIG. 5
`
`Page 5 of 13
`
`
`
`U.S. Patent
`
`Apr. 22, 2003
`
`Sheet 5 of 5
`
`US 6,553,131 B1
`
`
`
`(ERH
`
`Confidence:
`0.9 08 0.9 0.90.9 0.85 0.9 0.90.9
`
`FIG. 6
`
`Page 6 of 13
`
`
`
`1
`LCENSE PLATE RECOGNITION WITH AN
`INTELLIGENT CAMERA
`
`US 6,553,131 B1
`
`BACKGROUND
`
`1. Technical Field
`This disclosure relates to optical recognition methods and
`Systems and more particularly, to an intelligent camera for
`character recognition.
`2. Description of the Related Art
`Most image processing Systems for industrial applications
`and Video Surveillance are still based on a personal computer
`(PC), a frame grabber and a separate CCD camera. PC based
`vision systems for traffic Surveillance have been described in
`the prior art. In contrast to PC based systems intelligent or
`Smart cameras are recently becoming more popular, Since
`they offer Several advantages. They require minimal Space,
`because processor and Sensor are integrated in one package.
`They are more robust and reliable in outdoor environments,
`they require leSS maintenance and they are well Suited for
`cost Sensitive applications. However, intelligent cameras
`offer less computational power, Since they are typically one
`or two generations behind the current processor generation
`and they are limited with respect to main memory (e.g., 8
`MB) and hard disk space (e.g., a 5 MB flash disk).
`In principle, license plate recognition is may be similar to
`optical character recognition (OCR) for document process
`ing. However, existing OCR engines cannot be Successfully
`used as license plate readers because they cannot tolerate an
`extreme range of illumination variations, Such as inhomo
`geneous illumination and shadows, blurring due to dirt,
`Screws, particles, etc. In addition, OCR products are limited
`due of their memory and processing Speed requirements.
`Therefore, a need exists for versatile algorithms for appli
`cations that go beyond simple identification or measuring
`tasks. A further need exists for a System that reliably
`recognizes license plates for applications ranging from Sur
`veillance to automated Systems for determination of the
`parking fees.
`
`SUMMARY OF THE INVENTION
`A method for recognizing license plates employing an
`intelligent camera with a processor and a memory, in accor
`dance with the present invention, includes capturing an
`image including a license plate by the intelligent camera,
`and detecting a region in which the license plate is located
`by performing a coarse localization of the image.
`Orientation, position, and illumination conditions of the
`image are detected and accounted for to obtain a baseline
`image of the license plate. A fine localization of the baseline
`image is performed to obtain a more accurate depiction of
`Vertical resolution of the baseline image of the license plate.
`Characters depicted in the baseline image are Segmented by
`employing a projection along a horizontal axis of the base
`line image to identify positions of the characters. The
`characters are classified based on a Statistical classifier to
`obtain a confidence Score for the probability of properly
`identifying each character. The above Steps are recursively
`performed until each confidence Score exceeds a threshold
`value to recognize the characters.
`In alternate methods, the Step of detecting orientation,
`position, and illumination conditions of the image and
`accounting for the orientations, position, and illumination
`conditions of the image to obtain a baseline image of the
`license plate may include the Step of comparing each char
`
`15
`
`25
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`2
`acter in the image of the license plate with examples of
`images with different illuminations to account for illumina
`tion effects on the image. The Step of detecting a region in
`which the license plate is located by performing a coarse
`localization of the image may include the Steps of Sub
`Sampling the image to reduce a number of pixels, extracting
`Vertical edges in the image, generating a Saliency map based
`upon the vertical edges to identify regions in the image with
`a probability of including the license plate and extracting a
`localization result which includes the image of the license
`plate. The Step of Segmenting characters depicted in the
`baseline image by employing a projection along a horizontal
`axis of the baseline image to identify positions of the
`characters may include the Steps of providing a projection
`profile of pixel intensities acroSS Vertical lines of pixels in
`the baseline image, filtering the projection profile and iden
`tifying locations of characters in the baseline image depicted
`by area below a threshold value in the filtered projection
`profile.
`The Statistical classifier may employ a convolutional
`network, and the Step of classifying the characters based on
`a Statistical classifier to obtain a confidence Score for the
`probability of properly identifying each character may
`include the Step of training the classifier by employing
`Virtual Samples of characters. The method may include the
`Step of comparing character blocks and characters to pre
`determined license plate codes and conventions to check
`accuracy of recognition. The Step of recursively performing
`the method StepS until each confidence Score exceeds a
`threshold value may include the Step of considering adjacent
`characters together to attempt to improve the confidence
`Score. The above method steps may be implemented by a
`program Storage device readable by machine, tangibly
`embodying a program of instructions executable by the
`machine to perform the method steps.
`An intelligent camera System for recognizing license
`plates, in accordance with the invention, includes a camera
`adapted to independently capture a license plate image and
`recognize the license plate image. The camera includes a
`processor for managing image data and executing a license
`plate recognition program device. The license plate recog
`nition program device includes means for detecting
`orientation, position, illumination conditions and blurring of
`the image and accounting for the orientations, position,
`illumination conditions and blurring of the image to obtain
`a baseline image of the license plate. The camera includes
`means for Segmenting characters depicted in the baseline
`image by employing a projection along a horizontal axis of
`the baseline image to identify positions of the characters. A
`Statistical classifier is adapted for recognizing and classify
`ing the characters based on a confidence Score, the confi
`dence Score being based on a probability of properly iden
`tifying each character. A memory is included for Storing the
`license plate recognition program and the license plate
`image taken by an image capture device of the camera.
`In alternate embodiments, a trigger device may be adapted
`to cause an image to be captured responsive to an event. The
`event may include an approach of a vehicle. The means for
`detecting may include examples of images with different
`illuminations to account for illumination effects on the
`image for each character within the image. The means for
`Segmenting may include means for providing a projection
`profile of pixel intensities acroSS Vertical lines of pixels in
`the baseline image, a filter profile for filtering the projection
`profile and means for identifying locations of characters in
`the baseline image depicted by area below a threshold value
`in the filtered projection profile. The Statistical classifier may
`
`Page 7 of 13
`
`
`
`3
`includes one of a convolutional network and a fully con
`nected multilayer perceptron. The memory may include a
`database for predetermined license plate codes and conven
`tions for checking accuracy of recognition. The intelligent
`camera System may include a parking lot control System
`coupled to the intelligent camera System for determining
`parking fees based on character recognition of license plates.
`These and other objects, features and advantages of the
`present invention will become apparent from the following
`detailed description of illustrative embodiments thereof,
`which is to be read in connection with the accompanying
`drawings.
`
`BRIEF DESCRIPTION OF DRAWINGS
`This disclosure will present in detail the following
`description of preferred embodiments with reference to the
`following figures wherein:
`FIG. 1 is a block/flow diagram showing a system/method
`for employing an intelligent camera for license plate recog
`nition in accordance with the present invention;
`FIG. 2A is an image of a license plate;
`FIGS. 2B-2D show segmentation steps for segmenting
`the characters of the license plate image of FIG. 2A, in
`accordance with the present invention;
`FIG. 3 depicts illustrative examples of virtual samples
`which may be employed in accordance with the present
`invention;
`FIG. 4 is a Schematic diagram of a convolutional network
`for employing the present invention;
`FIG. 5 is a block diagram of an intelligent camera in
`accordance with the present invention; and
`FIG. 6 depicts a license plate at various Stages of pro
`cessing in accordance with the present invention.
`
`DETAILED DESCRIPTION OF PREFERRED
`EMBODIMENTS
`The present invention is directed to a robust intelligent
`character recognition System and method for determining
`characters under various conditions, for example, in differ
`ent illumination conditions, at different angles, etc. In one
`particularly useful embodiment, the present invention is
`employed in license plate recognition.
`A robust car identification System for Surveillance of
`parking lot entrances that runs completely on a Stand alone
`low cost intelligent camera is provided. To meet accuracy
`and Speed requirements, hierarchical classifiers and coarse
`to find Search techniques are applied at each recognition
`Stage for localization, Segmentation and classification.
`The present invention provides an efficient hierarchical
`decomposition of a recognition task, where coarse Segmen
`tation and classification methods are applied. Ambiguous
`patterns may be forwarded to more Sophisticated methods to
`achieve a desired recognition time (for example, 2-3 Sec
`onds based on an IntelTM 486/100 MHz processor). Given
`multiple Segmentation hypotheses, the reliable recognition
`or rejection of Segmented characters is one important aspect
`for the performance. The character recognition task is pref
`erably accomplished by employing a convolutional network.
`It should be understood that the elements shown in FIGS.
`1, 5 and 6 may be implemented in various forms of
`hardware, software or combinations thereof. Preferably,
`these elements are implemented in Software on one or more
`appropriately programmed general purpose digital process
`ing unit having a processor and memory and input/output
`
`15
`
`25
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`US 6,553,131 B1
`
`4
`interfaces. Referring now in Specific detail to the drawings
`in which like reference numerals identify similar or identical
`elements throughout the several views, and initially to FIG.
`1, a flow diagram of an illustrative license plate recognition
`method is shown. In block 10, character identification is
`based, in one example, on frontal or rear views of the car to
`obtain an image of a license plate. For example, the car may
`be stopped in front of a gate. An inductive Sensor may be
`employed to externally trigger image capture. In a preferred
`embodiment, the camera is fixed So that the orientation of
`the license plates is approximately constant. However, the
`Size and position of the license plates can vary slightly due
`to the position and the type of car. The approaching car may
`be detected by the camera itself and the external inductive
`Sensor trigger may be eliminated.
`In block 12, a course localization is performed preferably
`based on multi-resolution analysis. Different resolutions of
`the image are checked to assist in character recognition of
`Symbols or characters on the license plate. Coarse localiza
`tion hones in on the license plate image within the image
`captured. This is performed by Sub-Sampling the image, in
`block 14, to reduce the number of pixels therein (i.e., to
`reduce computational complexity). In block 16, Vertical
`edges are computed (since vertical edges dominate in license
`plates, however, horizontal edges may also be computed).
`Vertical edge computation determines vertically disposed
`pixel groups. In block 18, a Saliency map is generated. A
`Saliency map blends vertical edges to achieve intensity
`regions in the image. Highest peak values in the intensity
`regions have the highest probability of being the license
`plate and are Selected for further processing. In block 20, the
`invention hones in on the image of the license plate to
`provide the course localization.
`In block 22, tilt detection and refinement of an area of
`interest is determined. This includes addressing illumination
`effects, positions (distances away from the license plate, left
`and right borders), rotations, blurring and other effects.
`Comparisons to normalized correlation models or templates
`are performed to find Similar illuminations, distances, etc. to
`attempt to reduce the effects on the image. Advantageously,
`this provides a robustness to the System which is not
`achieved by OCR algorithms.
`In block 24, a fine localization based on local edge and
`regional features is performed, i.e., a vertical fine localiza
`tion for license plates. After Successful localization, multiple
`Segmentation hypotheses are created by an algorithm, in
`block 26, based on nonlinear projections onto a baseline
`image. The nonlinear projections are employed in a profile,
`in block 28 (See FIG. 2B) and filtered in block 30 (See FIG.
`2C). As a result, the left and right boundary of each character
`is determined, in block 32 (See FIG. 2D). In block 34, the
`recognition System preferably employs a convolutional neu
`ral network classifier or other Statistical classifier, which
`identifies the characters and returns a confidence value for
`each character. Based on the confidence measure, the Seg
`mentation hypothesis with the highest overall confidence is
`accepted. After image analysis, a table of rules, for example,
`valid city codes and country Specific rules are applied to
`Verify, correct or reject Strings, in block 36. For example,
`German license plates include two blocks with letters fol
`lowed by one block with digits only and the overall number
`of characters may not exceed eight. These rules can signifi
`cantly narrow the Search and improve performance.
`Localization of License Plates
`In block 12, license plate localization is preferably based
`on multi-resolution analysis. The coarse localization relies
`on the fact that a license plate (text) includes a high amount
`
`Page 8 of 13
`
`
`
`S
`of Vertical edges compared to the remaining parts of a car
`and its Surrounding in an image. Vertical edges are computed
`and a Saliency map is generated on a reduced resolution.
`Pixels with high intensity in the Saliency map correspond to
`positions that are likely to include the license plate. The
`highest peal value in the Saliency map corresponds to the
`license plate. In rare instances other peaks may occur. In this
`case a candidate list of possible locations is ordered by the
`intensity of the vertical edge feature which is computed and
`evaluated until a reasonable result is found. This method
`provides a robust technique for localizing the license plate
`with images having Size and orientation variations due to
`varying distance, position and type of license plate or car.
`Horizontal edges or other local features may be employed
`for coarse localization as well. Position and Size variations
`as well as Small rotations are accounted for by the System in
`block 22.
`In block 24, after coarse localization, fine localization is
`performed. The orientation of the license plate is precisely
`determined and corrected. The left and right border of the
`license plate are detected. The height and position of the
`license plate within the image are determined to more
`accurately retrieve a region of interest. This refinement is,
`again, based on Vertical edge features of the original image
`(not a Sub-Sampled image) to gain higher accuracy. After the
`position, orientation and size of the license plate are
`determined, the region of interest is resampled to a prede
`termined vertical resolution (e.g., 20 pixels). This resolution
`is maintained for Subsequent Steps to reduce computational
`overhead. Accurate positioning and Size normalization along
`the Vertical access improve the accuracy of Segmentation
`and classification.
`Segmentation of the Character String
`Based on the position and Size of the detected license plate
`the region of interest is extracted and re-sampled So that the
`Vertical resolution is about 20 pixels for the character height,
`this may be adjusted depending on the application. The 20
`pixel resolution for the character height was found to be
`Sufficient for the Subsequent Segmentation and recognition
`as will be described.
`In contrast to U.S. license plates the characters of German
`license plates are not equally Spaced and the character width
`is Strongly varying. Furthermore, two different fonts and
`layouts are used. While Sophisticated Segmentation algo
`rithms like Hidden Markov Model, which are also Success
`fully used for handwriting recognition or Speech
`recognition, may perform better, here a simple Segmentation
`based on projections to the baseline is advantageously
`employed to minimize execution time.
`Referring to FIGS. 2A-D, segmentation steps for block
`26 of FIG. 1 are illustratively depicted for a license plate 100
`shown in FIG. 2A. Based on a central part of the region
`including the characters of FIG. 2A, a profile or projection
`is generated in FIG. 3B from right to left along the image,
`based on a darkest pixel in each column (in the y direction).
`FIG. 2B is a plot of pixel intensity (X-axis) versus column
`number (y-axis). A low pass filtered profile, for example, the
`profile depicted in FIG. 2C is subtracted from the profile in
`FIG. 2B resulting in a profile of FIG. 3D. In FIG. 2D, the
`areas or regions below a threshold (e.g., 1.00) each include
`a character.
`Touching characters caused by dirt particles, image blur
`ring and shadows that partially cover the license plate may
`cause problems for a projection-based Segmentation and the
`peak between characters can become very Small. Thus, the
`Segmentation is preferably parameterized to be sensitive to
`Spaces. The Segmentation also detects many false positive
`
`15
`
`25
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`US 6,553,131 B1
`
`6
`Spaces. If the classification of the Segmented characters does
`not identify characters with high confidence, then a more
`powerful recognition based Segmentation algorithm may be
`applied, as described in commonly assigned U.S. application
`Ser. No. 09/396,952, entitled “CHARACTER SEGMEN
`TATION METHOD FOR VEHICLE LICENSE PLATE
`RECOGNITION", filed concurrently herewith and incorpo
`rated herein by reference.
`Identification of Characters
`Since the Segmentation process itself is inherently
`ambiguous, it has to be tightly coupled to the character
`identification. Segmentation provides locations that are
`likely to include characters. Consequently, the Overall per
`formance depends highly on the performance of the char
`acter classifier and its ability to recognize valid characters
`and to reject invalid characters due to false positive
`segmentation, as performed in block 26 of FIG. 1. Several
`experiments have been conducted by the inventors to choose
`optimal classifiers for this problem. In one experiment, fully
`connected multi-layer perception and convolutional net
`Works were employed. Experiments demonstrate the advan
`tage of convolutional networks compared to the fully con
`nected multilayer perception; however, either method may
`be used as well as equivalent methods.
`For German license plate recognition, 40 classes have to
`be recognized (10 digits, 26 regular characters, 3 special
`German characters: ae, ue, oe, and the registration label
`appearing between first and Second text field). Furthermore,
`negative examples caused by incorrect Segmentation and
`localization were used for training. All neurons in an output
`layer of the convolutional networks should respond with
`Zero activation to a negative input pattern. Negative patterns
`are particularly important to increase the rejection capability
`in case of a wrong Segmentation. By increasing the number
`of negative samples compared to the number of valid
`characters the classifier is biased towards high rejection
`accuracy instead of using all its capacity for interclass
`Separation.
`In addition to negative samples, virtual Samples (VS) or
`templates improve the generalization of the classifier during
`training, if the training Set has limited size. Virtual Samples
`are generated from the original images by Several affine
`transformations. In particular, the size, orientation and hori
`Zontal and vertical position of the characters may be varied
`randomly within a reasonable range (+/-2 pixel shift and
`+/-2%. Scaling, rotation<about 3 degrees) between training
`cycles and noise is added during training. With Virtual
`Samples, the classifier becomes more robust with respect to
`Segmentation and location tolerances. Examples of Virtual
`samples are illustratively shown in FIG. 3.
`Since the Segmentation may include false positive Spaces,
`all combinations of character positions are considered,
`which are possible given a minimum and maximum width
`for the characters. The neural network evaluates the Sub
`images corresponding to these multiple hypotheses. The
`result of the neural network classification is a posterior
`probability of a class given an observation (see C. Bishop,
`Neural Networks for Pattern Recognition, Oxford Univ.
`Press, 1995). This probability may be used to choose the
`most likely combination of characters.
`Before classification, each Sub-image is normalized with
`respect to brightness and contrast. The Sub-image preferably
`has a fixed size, for example, 16x20 pixels, So that wide
`characters like W still fit into the window. Very narrow
`characters like 'I' or 1 may cause problems, since neigh
`boring characters are partly overlapping with their window.
`Therefore, the character width determined by the segmen
`
`Page 9 of 13
`
`
`
`7
`tation is employed to Skip the left and right area beside the
`character and these pixels are Set to a uniform background
`gray Value.
`Fully Connected Multilayer Perceptron
`A three-layered perceptron (MLP)may be used for clas
`sification of characters and training by performing error
`back-propagation. The number of hidden neurons is varied
`manually for optimal generalization and about 100 hidden
`neurons was found to be preferred.
`Convolutional Neural Network Based on Local Connections
`Referring to FIG. 4, a topology of a convolutional net
`work is illustrated. An input image 200 is presented. A first
`layer includes four planes 201 with, for example, 16x20
`pixels. Four convolution kernels 202 of, for example, a 3x3
`pixel window Size are applied So that four edge orientations
`are extracted (diagonals, vertical and horizontal). A Sub
`Sampling is performed to reduce the planes 201 by, Say, a
`factor two to 8x10 pixels and pixel averaging is performed
`to obtain planes 203. After convolutional and Sub-Sampling,
`two fully connected layers, 204 and 205 follow with, for
`example, 100 and 40 neurons, respectively.
`When the convolutional network is employed for classi
`fication of characters, pixels of the input image, i.e., a
`Segmented character, are extracted into four most likely
`orientations (diagonals, vertical and horizontal). Each
`extraction is Sub-Sampled to reduce its size. Comparisons
`occur to decipher the characters and the best character
`(highest confidence character(s)) are output from the output
`neurons (e.g., 40 for German license plates).
`Training for this network (determining weights of hidden
`and output neurons) is performed by error back-propagation
`and virtual Samples, which are employed in a similar fashion
`as with MLP training. Virtual samples advantageously
`improve the systems accuracy. (See FIG. 3).
`Results
`To test the overall recognition performance 900 license
`plate images were processed in accordance with the inven
`tion. An overall recognition rate of 96% percent was
`obtained for the methods of the present invention as shown
`in Table 1.
`
`TABLE 1.
`
`Recognition
`rate %
`
`Execution time
`sec
`Pentium IITM 333
`MHz
`
`Execution
`time sec
`Intelligent
`Calea
`
`Localization
`Segmentation,
`Character
`classification
`Overall
`
`98.0
`98.0
`
`96.O
`
`O.10
`O.15
`
`O.25
`
`1.O
`1.5
`
`2.5
`
`An average processing time of 2.5 Sec on the target System
`was achieved by the intelligent camera of the present
`invention. In this case, the actual time may vary between
`about 2 and about 4 Seconds depending on image quality and
`appearance of the license plate, however further improve
`ments in processing time are contemplated. For bad images,
`more false positive Segmentation hypotheses are created and
`have to be evaluated by the classifier. In one embodiment,
`localization consumes about forty percent of the execution
`time-segmentation and identification consume about Sixty
`percent of the execution time.
`Referring to FIG. 5, an intelligent camera system 301 is
`shown in accordance with the present invention. In one
`embodiment, the methods of present invention may be
`developed and optimized on a PC. The methods may be
`
`US 6,553,131 B1
`
`8
`loaded on an intelligent camera 300, i.e., a camera capable
`of processing and Storing a program. Camera 300 includes
`a memory 302 and one or more processors 303 for executing
`program steps. Memory 302 preferably stores a current
`image of a license plate and Sub-images thereof. Also,
`license plate rules/codes (e.g., city codes) may be stored in
`memory 302 as well as trained weights of the neural network
`described above. Larger memories may include a plurality of
`images of license plates as well as other data. In one
`embodiment, intelligent camera 300, includes a Siemens VS
`710 camera. The Siemens VS 710 camera uses an INTEL
`processor (486/100 MHz, 8 MB RAM, 4 MB flash disk
`serial and Profi-bus interface). An interface 306 may be
`adapted to work with an external system 304. External
`System 304 may include a parking lot control System which
`includes fee Schedules, time in/out data, etc. for a given
`vehicle. Camera 300 is preferably compatible with a PC to
`permit license plate recognition methods to be downloaded
`and employed by camera 300. Therefore, the PC software
`written, for example, in C/C++ can be used on the camera
`without major adaptation. Routines for image capture and
`communication with the control System may have to be
`linked. Camera 300 may include a sensor 308 which triggers
`image capture through a lens 310 when a condition is
`Sensed. For example, if a vehicle approaches a parking lot
`gate, an image is captured for the vehicles license plate. The
`license plate image is recognized by the above methods of
`the present invention (FIG. 1). Memory 302 may store the
`license plate information as well as other information, Such
`as time in/out, etc. Advantageously, camera 300 may be
`placed remotely in a parking lot or other location without the
`need for a large processing System (a PC, etc.). In accor
`dance With the invention, an automated license plate recog
`nition System with high recognition accuracy is provided
`which includes a robust and low cost intelligent camera.
`With an intelligent camera based on a more powerful
`processor, motion analysis may be employed to extend
`applications towards moving vehicles and various outdoor
`Surveillance tasks. Camera 300 includes a license plate
`recognition program device 312 preferably implemented in
`Software. License plate recognition program device 312
`detects orientation, position, illumination conditions and
`blurring of the ima