CN104200207A - License plate recognition method based on Hidden Markov models - Google Patents

Abstract

A license plate recognition method based on Hidden Markov models includes 1, license plate positioning; 2, character segmentation, namely performing binaryzation processing on images, and determining the character positions according to the height-width ratio of characters of Chinese license plates; 3, character recognition, namely applying decomposition rules on each Chinese character in a set in the process of Chinese character processing to generate progressive split graph including radicals as nodes, finding out the optimization problem of an optimal radical set representing the Chinese character set in a formulation manner through the maximum likelihood and minimum description length, solving the optimization problem to acquire the optimal radical set, and using the optimal radical set in the character recognition algorithm based on the Hidden Markov models; when phabetic and digital characters are processed, extracting character skeletal accumulation characteristics, including stroke slope accumulation characteristics, inflection point amplitude accumulation characteristics and contour depth accumulation characteristics, by scanning four sides of characters. The method has high automation level, and the Chinese character recognition accuracy rate is improved greatly.

Description

A kind of licence plate recognition method based on hidden Markov model

Technical field

The present invention relates to license plate recognition technology, especially a kind of licence plate recognition method.

Background technology

License plate recognition technology (Vehicle License Plate Recognition, VLPR) is a kind of application of Video Image recognition technology in vehicle license identification.

License plate recognition technology requires the license plate in motion to be extracted and to identify from complex background, by car plate extract, the technology such as image pre-service, feature extraction, Recognition of License Plate Characters, identify vehicle identification number.Can realize the functions such as parking lot fee collection management, magnitude of traffic flow control index measurement, vehicle location, automobile burglar, high way super speed robotization supervision, electronic eye used for catching red light runner, toll station.For safeguarding traffic safety and urban public security, prevent traffic jam, realizing intelligent traffic administration system has real meaning.Existing license plate recognition technology, can identify comparatively accurately to English character and numeral, still, for the identification of Chinese character, has the poor technological deficiency of accuracy rate.

Summary of the invention

In order to overcome the deficiency that intelligent degree is lower, Chinese Character Recognition accuracy rate is lower of existing car plate recognition method, the invention provides the licence plate recognition method based on hidden Markov model that a kind of intelligent degree is higher, effectively promote Chinese Character Recognition accuracy rate.

The technical solution adopted for the present invention to solve the technical problems is:

A licence plate recognition method based on hidden Markov model, described licence plate recognition method comprises the steps:

Step 1, car plate location: the video data to watch-dog is decoded, and mask data frame, forms the view data of every frame video;

First image is carried out to gray processing processing, according to luminance equation, image is converted to gray-scale map, by pretreated image projection is processed, statistical picture is at the number of level, vertical direction up contour point respectively; In vertical and level two directions, scan, record meets the position of predetermined threshold value, obtains the region of car plate by comparison;

Step 2, Character segmentation: picture is carried out to binary conversion treatment, according to the depth-width ratio of the character of Chinese car plate, determine the position of character;

Step 3, character recognition, comprise following process:

3.1) while processing Chinese character, each Chinese character application decomposition rule in this set is usingd to generate and comprise the progressive broken away view as the radical of node, with maximum likelihood and minimum description length, carry out the formulistic optimization problem of finding out the optimum radical collection that represents Chinese character set, separate this optimization problem to obtain optimum radical collection, and in the character recognition algorithm based on hidden Markov model, use this optimum radical collection;

3.2), while processing English and numerical character, by four side scanning extraction character bone Accumulations of character, comprise stroke slope value accumulation feature, flex point amplitude Accumulation, profile depth Accumulation.

Further, described step 3.1) in, the process of the character recognition algorithm based on hidden Markov model is as follows: utilize HMM model to carry out modeling to each character, the sequence of observations using character feature as HMM model, finds out status switch with Viterbi algorithm; The training of HMM can adopt BW algorithm, obtains HMM model λ=(A, B, π) represent, A is that state transition probability distributes, and B is observed reading probability distribution, and π is initial state distribution, the parameter finally obtaining is that the maximum likelihood of this hidden Markov model is estimated, call HMM model λ=(A, B, π), adopt forward direction algorithm, calculate conditional probability P (the O| λ of character to be detected under each HMM _i), conditional probability P (O| λ _i) value maximum be optimum matching, assert that character to be detected is the corresponding character of this HMM model.

Further, in described step 2, the depth-width ratio of the character of Chinese car plate is defined as follows:

L≈4.5H

H≈1.9W _c

W _n≈0.3W _c

D _n≈0.5W _c

Wherein, the string length that L is car plate, the height that H is character string, W _cfor the width of ordinary symbol, ordinary symbol does not comprise ' 1 ', W _nfor the width of character ' 1 ', D _ninter-character space for the both sides of character ' 1 '.

Beneficial effect of the present invention is mainly manifested in: take the technology such as Digital Image Processing, pattern-recognition, computer vision as basis, the vehicle image of shot by camera or video sequence are analyzed, the number-plate number of identification automobile, has greatly improved traffic intelligence management.Utilize hidden Markov model to detect car plate, the especially Chinese character of cognitron motor-car simultaneously, there is higher accuracy rate.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of the motor vehicle photo of car plate to be detected.

Fig. 2 is the schematic diagram of the car plate picture after binary conversion treatment.

Fig. 3 is the schematic diagram of the car plate picture after separating character.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1～Fig. 3, a kind of licence plate recognition method based on hidden Markov model, by the vehicle on monitored road surface being detected and automatically extracting vehicle license information, and processes, and described licence plate recognition method comprises the steps:

Step 1, car plate location.Video data to watch-dog is decoded, and mask data frame, forms the view data of every frame video, thereby is convenient to the follow-up identification of single image being carried out to information of vehicles.

The colors such as the car plate photographing is all generally coloured image, and license plate has the surplus of the yellow end, wrongly written or mispronounced character of the blue end.For these license plate images are handled together, will first car plate be carried out to gray processing processing, according to luminance equation, image is converted to gray-scale map.Its conversion formula is:

Y＝0.299×R+0.587×G+0.144×B

Wherein R, G, B are respectively three colouring components of the red, green, blue of input color image, and Y is the gray-scale value of this pixel in gray-scale map.

By pretreated image projection is processed, statistical picture is at the number of level, vertical direction up contour point respectively.Owing to there being the character informations such as Chinese letters on car plate, therefore abundant than other edges of regions.Suitable threshold value can be set, in vertical and level two directions, scan, record meets the position of threshold value, by comparison, can obtain the region of car plate.

Step 2, Character segmentation.Picture is carried out to binaryzation (black and white) to be processed.Grey level range in one width gray level image is 0～255, in image the gray-scale value of the pixel of each point be f (x, y) ∈ 0,1 ..., 255}, being provided with a threshold value is T (0≤T≤255):

$g(x,y)=\left\{\begin{array}{cc}255,& f(x,y)\&GreaterEqual;T,\\ 0,& f(x,y)<T.\end{array}\right.$

Wherein, g (x, y) represents after binaryzation the value of each pixel in image.If g (x, y)=255, illustrate that this point is target, otherwise, be background.As can be seen here, from complicated image, target is intactly extracted from background, threshold value choose the key that just becomes binaryzation.If threshold value is chosen too high, too much impact point is just by the background that is included into of mistake; Threshold value is chosen too lowly, there will be contrary situation.For object can correctly be split, key is to select suitable threshold value.According to test, the threshold value of setting is 125.

General licence plate character has 7, and length all meets certain standard.The overall length of character zone is 409mm, the wide 45mm of each character, and high 90mm, the 2nd and the 3rd character pitch is 34mm, all the other character pitches are 12mm.After although license plate image is processed, can extract the character above it and the certain noise of timing meeting substitution, in character string, between each character of original expression can there is not too large variation in the mutual relationship between the parameters of relation.The depth-width ratio that is below the character of Chinese car plate is defined as follows:

L≈4.5H

H≈1.9W _c

W _n≈0.3W _c

D _n≈0.5W _c

Wherein, the string length that L is car plate, the height that H is character string, W _cwidth for ordinary symbol.The ordinary symbol here does not comprise ' 1 ', W _nfor the width of character ' 1 ', D _ninter-character space for the both sides of character ' 1 '.4 formula are above the major parameter of relation between character on Chinese car plate, can determine according to the depth-width ratio of character the position of character.

Step 3, character recognition.Because the first character of the car plate of most of China is Chinese character, second to the 7th character is letter or number, this just can be divided into license plate image identifying two parts and process, first is the process of identification Chinese character, second portion is identification letter and digital process, because Chinese-character stroke is more, different with the processing procedure of letter or number.

While processing Chinese character, each the Chinese character application decomposition rule in this set is usingd generate the progressive broken away view that comprises as the radical of node, by the formulistic optimization problem of finding out the optimum radical collection that represents Chinese character set of maximum likelihood and minimum description length, separate this optimization problem to obtain optimum radical collection and use this optimum radical collection in the character recognition algorithm based on hidden Markov model.

While processing English and numerical character, by four side scanning extraction character bone Accumulations of character, comprise stroke slope value accumulation feature, flex point amplitude Accumulation, profile depth Accumulation.

Utilize HMM model to carry out modeling to each character, the sequence of observations using character feature as HMM model, finds out status switch with Viterbi algorithm; The training of HMM can adopt BW algorithm, obtains HMM model λ=(A, B, π) represent, A is that state transition probability distributes, and B is observed reading probability distribution, π is initial state distribution, and the parameter finally obtaining is that the maximum likelihood of this hidden Markov model is estimated.Call HMM model λ=(A, B, π), adopt forward direction algorithm, calculate conditional probability P (the O| λ of character to be detected under each HMM _i).Conditional probability P (O| λ _i) value maximum be optimum matching, assert that character to be detected is the corresponding character of this HMM model.

Claims (3)

1. the licence plate recognition method based on hidden Markov model, is characterized in that: described licence plate recognition method comprises the steps:

Step 1, car plate location: the video data to watch-dog is decoded, and mask data frame, forms the view data of every frame video;

First image is carried out to gray processing processing, according to luminance equation, image is converted to gray-scale map, by pretreated image projection is processed, statistical picture is at the number of level, vertical direction up contour point respectively; In vertical and level two directions, scan, record meets the position of predetermined threshold value, obtains the region of car plate by comparison;

Step 2, Character segmentation: picture is carried out to binary conversion treatment, according to the depth-width ratio of the character of Chinese car plate, determine the position of character;

Step 3, character recognition, comprise following process:

3.1) while processing Chinese character, each Chinese character application decomposition rule in this set is usingd to generate and comprise the progressive broken away view as the radical of node, with maximum likelihood and minimum description length, carry out the formulistic optimization problem of finding out the optimum radical collection that represents Chinese character set, separate this optimization problem to obtain optimum radical collection, and in the character recognition algorithm based on hidden Markov model, use this optimum radical collection;

3.2), while processing English and numerical character, by four side scanning extraction character bone Accumulations of character, comprise stroke slope value accumulation feature, flex point amplitude Accumulation, profile depth Accumulation.

2. a kind of licence plate recognition method based on hidden Markov model as claimed in claim 1, it is characterized in that: described step 3.1), the process of the character recognition algorithm based on hidden Markov model is as follows: utilize HMM model to carry out modeling to each character, the sequence of observations using character feature as HMM model, finds out status switch with Viterbi algorithm; The training of HMM can adopt BW algorithm, obtains HMM model λ=(A, B, π) represent, A is that state transition probability distributes, and B is observed reading probability distribution, and π is initial state distribution, the parameter finally obtaining is that the maximum likelihood of this hidden Markov model is estimated, call HMM model λ=(A, B, π), adopt forward direction algorithm, calculate conditional probability P (the O| λ of character to be detected under each HMM _i), conditional probability P (O| λ _i) value maximum be optimum matching, assert that character to be detected is the corresponding character of this HMM model.

3. a kind of licence plate recognition method based on hidden Markov model as claimed in claim 1 or 2, is characterized in that: in described step 2, the depth-width ratio of the character of Chinese car plate is defined as follows:

L≈4.5H

H≈1.9W _c

W _n≈0.3W _c

D _n≈0.5W _c

Wherein, the string length that L is car plate, the height that H is character string, W _cfor the width of ordinary symbol, ordinary symbol does not comprise ' 1 ', W _nfor the width of character ' 1 ', D _ninter-character space for the both sides of character ' 1 '.