CN101055618A - Palm grain identification method based on direction character - Google Patents

Abstract

The invention discloses a method for identifying the palmprint based on directional characteristic, including the steps of: collecting the palmmprint image, preprocessing palmmprint image, constructing the palmmprint trainning ROI image collection, extracting-establishing the palmmprint direction feature coding RPOC by the palmmprint feature, matching the palmmprint in the dot pairs area. First new palmmprint trainning ROI image is established by rotating the palmmprint trainning ROI image and is added into the palmmprint trainning ROI image collection for compensating the rotation error; secondly, a modified limited Radon conversion MFRAT is designed, and the palmmprint direction characteristic mode is obtained by comparing the energy sizes of linear areas in six direction at the local area of palmmprint image; Finally, the palmmprint direction characteristic mode is matched through the match of dot pair area. The palmmprint direction feature coding RPOC in the method, not only can reflect the structural feature of palmprint, also can carry major distinguishing information, at the same time, having good fault-tolerant ability for light change, displacement revolving between palmmprint images or the like.

Description

Palmprint Recognition Method Based on Orientation Feature

Field of the Invention The present invention relates to a method for identity authentication using human biological features, in particular to a palmprint recognition method based on direction features.

Background technology In the network information society, there is an urgent need to be able to effectively identify people's real identities. There are two main traditional identification methods: one is password-based security mechanism; the other is certificate-based security mechanism. However, these traditional confirmation mechanisms have their inherent drawbacks. For example, multiple passwords are difficult to remember and easy to forget; certificates are easy to be forged, stolen and lost. Therefore, people look forward to using a safer and more convenient identity authentication method.

In the past ten years, the identity authentication technology (Biometrics) based on human body biometrics has been paid more and more attention by people. The so-called biometric identification technology refers to the technology that uses the inherent physical characteristics or behavioral characteristics of the human body to identify personal identities through image processing, pattern recognition and other methods. Compared with the traditional authentication methods based on passwords or ID cards, it can be carried around, difficult to forge and does not need to be remembered, so it has better security, reliability and effectiveness. At present, the physical features studied in biometric identification technology mainly include fingerprints, faces, irises, palm prints, hand shapes, ear prints, and veins; the behavioral features studied mainly include handwriting, voice, gait, and keystrokes. Among the above studies, fingerprint-based identification is the earliest and most mature method, but the wearability and destructibility of fingerprints limit the further promotion of this method to a certain extent. Although the identification method based on iris and cornea has the advantages of high recognition rate, it also has defects such as expensive equipment and poor user acceptance. Face and voice are also research hotspots in biometric recognition technology, but due to the influence of lighting conditions, face posture, noise and other factors, the accuracy rate is also unsatisfactory.

Palmprints are unique and basically invariable throughout life. Compared with fingerprints, palmprints have a much larger area, richer texture information, and easier access to palmprint images. In recent years, research on biometric identification technology based on palmprint recognition has also received extensive attention.

Early palmprint recognition was processed off-line. Collecting palmprints was to use ink to press the palm on white paper, and then use digital cameras, scanners, etc. to obtain digital pictures. However, there are many shortcomings in the application of offline palmprint recognition to identity authentication. First of all, it is not real-time, and most of the practical applications require online real-time recognition; secondly, the quality of the ink palmprint image is not high, and the features cannot be represented with high quality; thirdly, in the offline image, the palmprint It is difficult to effectively locate the ROI (Region of Interest) area. After 2002, international research on palmprint recognition gradually shifted to online palmprint recognition. The biggest difference between online palmprint recognition and offline palmprint recognition is the use of digital equipment to directly obtain high-quality palmprint images, which can be processed in real time. Among them, in related papers for research purposes, the commonly used equipment for obtaining palmprint images is CCD camera, digital scanner, etc. It should be pointed out here that the Biometric Identification Center of Hong Kong Polytechnic University has designed a special palmprint acquisition device, which won the gold medal of the 14th China Invention Exhibition.

In international and domestic paper databases, more than 100 papers on palmprint recognition can be retrieved. Overall, the proposed palmprint recognition methods can be divided into the following categories:

(1) Method based on texture features. This kind of method regards the palmprint image as a kind of texture structure, and uses related methods to extract the texture features of palmprint for recognition. PalmCode proposed by D.Zhang and A.Kong is a classic texture-based palmprint recognition method, which uses Gabor filter to filter the palmprint image, and then applies the zero-crossing criterion to encode the palmprint image [References : D. Zhang, W.K. Kong, J. You, and M. Wong, "Online palmprint identification," IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(9)(2003), pp.1041-1050.]. Then A.Kong et al. used information fusion technology to improve PalmCode and proposed FusionCode, and the recognition rate was further improved [References: A.Kong, D.Zhang, and M, Kamel, "Palmprint identification using feature-level fusion," Pattern Recognition. 39(2006) 478-487].

(2) Method based on line features. L. Zhang et al. first performed wavelet decomposition on the palmprint image, and then used the directional modeling method to extract the important coefficients of the wavelet sub-bands as the main line and important fold features [References: L. Zhang, and D. Zhang, "Characterization of palmprints by wavelet signatures via directional context modeling,” IEEE Transaction on Systems, Man and Cybernetics, Part B.34(3)(2004), pp.1335-1347.]. Wu Xiangqian regards the palm line as a kind of roof line, and determines the palm line according to the zero-crossing point of the first derivative of the image and the amplitude of the second derivative. On the basis of this algorithm, Wu Xiangqian et al. discuss the palm line based on line features. Pattern classification and recognition algorithm [References: X.Q.Wu, D.Zhang, and K.Q.Wang, "Palm line extraction and matching for personal authentication," IEEETransaction on Systems, Man and Cybernetics, Part A, 36(5)(2006) , pp. 978-987.].

(3) The method based on the characteristic feature or called the subspace method. Mainly use techniques such as eigenvalue decomposition and Gram-Schmidt orthogonalization to reduce the dimensionality of the palmprint image and obtain the corresponding eigenvectors. G.M.Lu and X.Q.Wu respectively proposed palmprint recognition methods based on PCA (Principal Component Analysis) and LDA (Linear Discriminant Analysis) [References: G.M.Lu, D.Zhang, and K.Q.Wang, "Palmprint recognition using eigenpalms features," PatternRecognition Letter, 24(9-10)(2003), pp.1463-1467.], [References: X.Q.Wu, D.Zhang, and K.Q Wang, "Fisherpalms based palmprint recognition," PatternRecognition Letter, 24(1 5) (2003), pp.2829-2838.].

(4) Method based on direction information. This method is similar to the calculation and estimation of the direction field of a fingerprint image. For the palmprint image, it is to obtain the direction of each pixel of the palmprint image, so that the palmprint image is mapped from the gray space to the direction feature space, and then matched. At present, in the field of palmprint recognition, algorithms based on directional features have achieved a higher recognition rate, because the directional information of palmprint lines can carry more discrimination information, and it is not sensitive to variations such as illumination changes. A.Kong and David.Zhang (Zhang Dapeng) proposed CompetitiveCode on the basis of FusionCode. In this method, the Gabor filter in 6 directions is used to filter the image, and then the Winner-take-all rule is used to extract the strongest response direction as recognition Features [References: A. Kong, and D. Zhang, "Competitive coding scheme for palmprint verification," Proc.Of the ^17th ICPR, vol(1)(2004), pp.520-523.].

Among the above recognition methods, the first type of texture-based method was studied relatively early, but this method is easily affected by factors such as illumination changes, and the recognition rate is difficult to improve. The second type of line-based methods suffer from many limitations, for example, many palm lines are blurred and difficult to extract. The third type of recognition method based on subspace technology is currently a research hotspot, but this method only considers the correlation between image pixels and does not use the structural information of palmprint images, and the recognition results on large-scale databases still need further research. verify.

In contrast, the fourth type of palmprint image recognition method based on directional features can obtain better recognition results. At present, in the comparison of the recognition results of various methods, the method based on directional features has the best recognition rate. The classic palmprint recognition method based on directional features is the CompetitiveCode we mentioned above [References: A.Kong, and D.Zhang, "Competitivecoding scheme for palmprint verification," Proc.Of the ^17th ICPR, vol(1 ) (2004), pp.520-523.].

Through international patent database retrieval, Zhang Dapeng (David. Zhang) and others of Hong Kong Polytechnic University's biometric identification center applied for a US invention patent in June 2004. The patent announcement number is WO/2005/124662, and the name is "based on palm line Characteristic palmprint recognition method". The core content of the palmprint recognition method disclosed in this invention is CompetitiveCode.

However, the palmprint recognition method based on CompetitiveCode still has some deficiencies. The specific performance is as follows: (1) the Gabor filter used in the palmprint recognition method of CompetitiveCode is not the best tool for extracting the directional features of the palmprint image; (2) the feature extraction speed is relatively slow, because the Gabor filter with a larger template is used It is time-consuming to filter the palmprint image; (3) no solution to the rotation problem is proposed, and the recognition rate is difficult to further improve; (4) CompetitiveCode's palmprint recognition method uses Hamming distance for matching and lacks good fault tolerance.

SUMMARY OF THE INVENTION The object of the present invention is to overcome the deficiencies in the palmprint recognition method of CompetitiveCode in the prior art, and propose a new palmprint recognition method based on direction features. The palmprint recognition method based on direction features not only has a very fast feature extraction speed, but also has a greatly improved recognition rate. This method has stronger robustness and better practicability than CompetitiveCode's palmprint recognition method.

The object of the present invention is achieved like this: the palmprint recognition method based on direction feature, comprises

Step (1) palmprint image collection

The palmprint image is collected by the palmprint image collection device to obtain a palmprint image grayscale matrix that can be used for further processing. The palmprint images collected in the registration stage are called palmprint training images, and the palmprint images collected in the recognition stage are called palmprint test images.

Step (2) palmprint image preprocessing

Before extracting the features of the palmprint image, the palmprint image needs to be preprocessed. When collecting palmprint images, the palmprint image of the entire palm is generally collected, but it is not suitable to use this kind of palmprint image for matching. On the one hand, because the palmprint image of the entire palm is relatively large and the processing speed is slow, it is not applicable. On the other hand, because the palmprint image of the entire palm has not been processed by positioning, there are large rotation and displacement errors, which makes the matching result unstable. Therefore, in the palmprint image recognition method, firstly, the palmprint image is rotated and corrected by locating the position of the palm and fingers, and then a square area of 128×128 pixels is cut in the center of the palmprint image as the palmprint training ROI (Region of Interest) ) image, and finally perform feature extraction and matching on the palmprint training ROI image in the square area.

Especially also comprise: the structure of step (3) palmprint training image set

In the palmprint recognition system, first one or several palmprint images are collected and stored in the system as palmprint training images. However, due to imperfect preprocessing operations, there is often a certain rotation error between the palmprint image to be recognized and the palmprint training image, which easily leads to wrong recognition. At present, the field of palmprint recognition has not proposed an effective solution to this problem. The present invention solves this problem by constructing new palmprint training images. By observation, the maximum rotation error between the palmprint image to be recognized and the palmprint training image is about 10°. Assume that there is a palmprint training ROI image A in the system for a person's palmprint, and the rotation angle α of the palmprint training ROI image A is 3°, 6°, 9°, -3°, -6°, - 9°, get the rotated palmprint training ROI image, that is, form a new palmprint training ROI image set A1, A2, A3, A4, A5, A6, suppose the system also includes multiple palmprint training ROIs of B and C image, then the final palmprint training ROI image set is A, A1, A2, A3, A4, A5, A6, B, C, and the structure of the palmprint training ROI image set can effectively compensate for the rotation error;

Among them, the rotation angle α and the number of new palmprint training images generated can be adjusted according to the actual situation.

Step (4) Palmprint Feature Extraction-Establish Palmprint Orientation Feature Coding RPOC

The main feature of the palmprint is the line feature, and these lines are directional, so the directional feature can also effectively express the essential structure of the palmprint. In the local area of the palmprint image, the palmline can be approximated as a short straight line, so the finite Radon transform (the finite Radon transform, FRAT for short) can be used to calculate the direction of the palmline. However, due to the "surround effect" of FRAT, it is not accurate when calculating the direction feature. The present invention has designed a novel improved Radon transform (Modified Finite Radon Transform, is referred to as MFRAT to accurately calculate the palmprint feature. The calculation palmprint feature of MFRAT is as follows:

Define Z _p = {0, 1, ..., p-1}, where p is a positive integer, for the real-valued equation image f[x, y] on the finite two-dimensional grid Z _p ² , MFRAT is defined as:

$\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; MFRAT</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; \&Element;</span>}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, L _k is a straight line formed by some points of f[x, y] in the two-dimensional grid Z ² _p :

L _k = {(i, j): j=k(ii ₀ )+j ₀ , i∈Z _p } (2)

In the above formula, L _k is the straight line equation, (i ₀ , j ₀ ) is the center point of Z ₂ ^p . k is expressed as the slope of L _k . Then L _k is expressed as a straight line passing through the center point (i ₀ , j ₀ ) of Z ^{2 p} with different slopes (directions). L _k has another representation method L(θ _k ), where θ _K is the angle value corresponding to k.

r(L _k ) in formula (1) represents the integration (summation) of L _k in different directions, that is, r(L _k ) represents the energy of L _k in different directions. The direction information of the palmprint is calculated by comparing the size of r(L _k ). Since the pixel value of the palm line is generally small in the palmprint image, the direction of the minimum value in r(L _k ) is selected as the final direction information. See the formula below:

${\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="0"\; label="i"\; filenames="A20071011128900292.png"\; state="\{\{state\}\}">i</figure-callout></span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; j</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; min</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

In the entire palmprint image, the direction information of the entire image is calculated by moving Z ² _p pixel by pixel or multiple pixels. The formula of the orientation map of the palmprint image is:

Among them, k(i, j) is the value of k in formula (3) θ _k (i, j).

In MFRAT, there are three parameters that can be adjusted in the application, namely p, N and W. One is p, which determines the size of the two-dimensional grid Z ² _p , which is equivalent to determining the length of L _k ; the other is the number N of k, which indicates the energy of how many lines are calculated. If N is large, the calculation amount will be large , if N is small, there are too few directional features. Generally speaking, the size of N is between 6 and 12; the third is the width W of L _k , and the size of W can be adjusted according to application requirements. Generally speaking, the size of W Between 1 and 4. In the present invention, p is set to 16; N is set to 6; W is set to 4, and the final palmprint feature image, that is, the palmprint training ROI image is 32×32 pixels.

Step (5) based on point-to-area palmprint matching

In other palmprint recognition methods, the Normalized Hamming Distance or Angular Distance is often used for feature matching. However, matching results using Hamming distance or angular distance are often not robust because they are based on pixel-to-pixel matching. Generally speaking, due to displacement and rotation errors between the palmprint image to be recognized and the palmprint training image, the pixels of the two images cannot be precisely overlapped. The palmprint matching is performed based on the point-to-area distance function designed by the present invention, which can effectively improve the accuracy of the palmprint matching.

Suppose A is a palmprint training image and B is a palmprint test image, they are from the same palm, but collected at different time periods. Both A and B have a size of m×n pixels. Further assume that A(i, j) and B(x, y) are two corresponding points at the same position. If there is no displacement or rotation error between A and B, then we know that A(i, j) and B(x, y) are coincident, that is, "i=x" and "j=y". However, as mentioned in the previous paragraph, due to displacement and rotation errors, A(i, j) often does not coincide with B(x, y). On the other hand, the probability of A(i, j) appearing near B(x, y) is relatively high. According to the above analysis, the designed point-to-area based matching can be expressed as:

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CircleTimes;</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

In formula (5), s(A, B) represents the matching distance from A to B. "" indicates the logical "etc" operation, that is, A(i, j) and The value of any pixel in B(i, j) is equal, then A(i, j) The value of B(i, j) is 1, otherwise it is 0. B(i,j) is a local region centered on B(i,j), which can be defined as different shapes. Similarly, the matching distance from B to A is:

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CircleTimes;</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

The final matching distance is:

S(A,B)=S(B,A)=Max(s(A,B),s(B,A)) (7)

Use point-to-region matching. In the present invention, B(i, j) is set as a cross-shaped area with an area of 5 pixels (B(i-1, j), B(i+1, j), B(i, j), B(i, j -1), B(i, j+1)), or a square area of 9 pixels in size ((B(i-1, j-1), B(i-1, j+1), B(i-1 , j), B(i+1, j), B(i, j), B(i, j-1), B(i, j+1), B(i+1, j+1), B (i+1, j-1)).

Compared with the prior art, the beneficial effects of the present invention are:

First, use the method of constructing the palmprint training ROI image to compensate the rotation error.

As mentioned in the background technology of the present invention, due to the imperfect preprocessing operation, the palmprint test image to be recognized often has a large rotation error with the palmprint training image collected in the registration stage, which is easy to cause wrong recognition. The invention constructs a new palmprint training ROI image for compensating the rotation error. For example, there is a palmprint training ROI image A in the system, and the palmprint training ROI image A is rotated, and the angles α are 3°, 6°, 9°, -3°, -6°, -9° respectively, and the obtained New palmprint training ROI image sets A ₁ , A ₂ , A ₃ , A ₄ , A ₅ , A ₆ . If there are multiple palmprint training ROI images in the system, in addition to the palmprint training ROI image A, there are also B, C, etc., then the final palmprint training ROI image set is A, A ₁ , A ₂ , A ₃ , A ₄ , A ₅ , A ₆ , B, C. Among them, the rotation angle α and the number of generated new palmprint training ROI images are adjusted according to the actual situation. Compensating the rotation error by constructing the palmprint training ROI image set can effectively reduce the negative impact of the rotation error on palmprint image recognition.

Second, the palmprint orientation feature code RPOC is proposed, that is, Robust Palmprint OrientationCode, which improves the finite Radon transform MFRAT (Modified Finite RadonTransform), which can quickly and accurately extract the palmprint orientation feature with high-precision recognition rate. In the prior art, Competitive Code, a classic palmprint recognition method based on direction information, uses Gabor filters in six directions to filter the image, and determines the direction value of the pixel by comparing the response values of the filters. However, the convolution of the Gabor filter and the image is time-consuming, because a large number of multiplication and addition operations are required, and the Gabor filter cannot simulate linear features well.

The present invention proposes an improved finite Radon transform MFRAT. In the local range of the palmprint image, in addition to the high-precision recognition rate, a major advantage of the palmprint direction feature coding RPOC is that it has a very fast processing speed, because when using MFRAT as a feature When extracting, the addition operation is mainly used, and the addition operation is performed in the linear area in 6 directions, which can effectively reduce the time of palmprint feature extraction, and can fit the linear feature very well, and has a better performance than the Gabor filter. Ability to extract directional features. In Figure 5MFRAT of the specific embodiment, the width W of the line is 1, and the number of directions is 6, and the (a)(b)(c)(d)(e)(f) images represent the integral of lines in different directions (summed), these directions are 0°, 30°, 60°, 90°, 120°, 150°.

In addition, Fig. 7 is a comparative figure of the present invention, (a) image is the original palmprint ROI image, (b) is the palmprint recognition method based on CompetitiveCode, (c) is the palmprint direction feature code RPOC, It can be seen that the palmprint direction feature code RPOC of the present invention can better reflect the palmprint structural features. In Fig. 11, the ROC curves of palmprint orientation feature code RPOC and several classic palmprint recognition methods (Palmcode, Fuioncode, Competitivecode) are shown. In the ROC curve, given a FAR value, the larger the value of GAR, the better the recognition rate. It can be seen that the recognition performance of the palmprint orientation feature code RPOC is significantly better than other classical palmprint recognition methods in the prior art.

Third, a novel point-to-area palmprint feature matching is proposed, which has better fault tolerance. In the prior art, Normalized Hamming distance or Angular distance is often used for feature matching. However, matching results using Hamming distance or angular distance are often not robust because they are based on pixel-to-pixel matching. Generally speaking, due to the displacement and rotation errors between the palmprint test image to be recognized and the palmprint training image collected in the registration stage, the pixels of the palmprint test image to be recognized and the palmprint training image collected in the registration stage cannot be accurately overlapped. .

The point-to-area matching based on the design in the present invention can be expressed as:

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CircleTimes;</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; B</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

In formula (8), s(A, B) represents the matching distance from A to B. "" indicates the logical "etc" operation, that is, A(i, j) and The value of any pixel in B(i, j) is equal, then A(i, j) The value of B(i, j) is 1, otherwise it is 0. B(i,j) is a local region centered on B(i,j), which can be defined as different shapes. Similarly, the matching distance from B to A is:

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CircleTimes;</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

The final matching distance is:

S(A,B)=S(B,A)=Max(s(A,B),s(B,A)) (10)

As a further improvement to the prior art, the present invention B(i, j) is set as a cross-shaped area with an area of 5 pixels (B(i-1, j), B(i+1, j), B(i, j), B(i, j -1), B(i, j+1), B(i, j), B(i, j-1), B(i, j+1)), or a 9-pixel square area ((B( i-1, j-1), B(i-1, j+1), B(i-1, j), B(i+1, j), B(i, j), B(i, j -1), B(i, j+1), B(i+1, j+1), B(i+1, j-1).Therefore the present invention designs based on the point-to-area distance function to control Pattern matching has strong fault tolerance and can effectively improve the matching accuracy.

Description of drawings

Fig. 1 is a flow chart of a palmprint recognition method based on direction features.

Fig. 2 is the palmprint image that the present invention gathers.

Fig. 3 is the original palmprint training ROI image after the present invention cuts.

Fig. 4 is the schematic diagram that the present invention constructs ROI image collection of palmprint training.

Fig. 5 is a schematic diagram of the 9×9 size MFRAT of the present invention.

Fig. 6 is a schematic diagram of a 16×16 size MFRAT of the present invention.

Fig. 7 is a comparison chart of RPOC and Competitivecode features of the present invention.

Fig. 8 is a schematic diagram of a point-to-region matching operator in the present invention.

Fig. 9 is a distribution diagram of matching values of true matching and false matching in the verification test of the present invention.

Fig. 10 is a distribution diagram of FAR and FRR of the verification test of the present invention.

Fig. 11 is the ROC curve diagram of the RPOC result comparison of the palmprint direction feature encoding of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

Fig. 1 is a flow chart of a palmprint recognition method based on direction features. In Fig. 1, the palmprint recognition method based on direction features includes registration process and recognition process.

The registration process is: the user uses the collection device to collect palmprint images, and the user's personal information such as name, ID number, etc. is stored in the system. Generally, 1 to 3 collections are required for each user, and 1 to 3 palmprint training images are stored in the template database. Using palmprint image preprocessing, cut the central area of 128×128 size in the palmprint training image as the palmprint training ROI image. To construct the palmprint training ROI image set is to rotate one of the palmprint training ROI images in small angles to form several rotated palmprint training ROI images, that is, to form a new palmprint training ROI image set. Use MFRAT to extract the directional features of all palmprint training ROI images, and store the palmprint directional feature templates that form the palmprint training image set into the template database. Wherein, the palmprint direction feature template of each palmprint training ROI image set corresponds to the previously stored personal identity information such as name and ID number.

The identification process is: the user uses the collection device to collect palmprint images. To preprocess the palmprint test image, cut the center area of 128×128 size in the palmprint test image as the palmprint test ROI image. Use MFRAT to extract the directional features of the palmprint test ROI image to form a palmprint directional feature template to be recognized. For identity verification operations, the user to be identified also needs to input identity information such as an ID number into the system. The palmprint test direction feature template is matched with several palmprint training direction feature templates with the same ID number in the training template database. Palmprint matching using point-to-region. The value of the maximum similarity between the test direction feature template and the training direction feature template in the template database is taken as the final matching value. If the matching value is greater than the preset threshold T, the verification operation is successful. Otherwise verification fails. For the identity recognition operation, the user to be identified does not need to input identity information such as an ID number into the system. The palmprint test direction feature template is matched with all training feature templates in the training template database, and the system returns the training pattern with the largest matching value with the test direction feature template. The ID number of the direction feature template. If the matching value is greater than the preset threshold T, the identification is successful. Otherwise, the identification fails.

Fig. 2 is the palmprint image that the present invention gathers. In the collected palmprint image, the background is black, which is convenient for segmenting the palmprint area image. This image resolution is about 75dpi. Although the resolution is low, features such as the main line and wrinkles of the palm print are still quite clear and can be used for identity authentication. The first advantage of using low-resolution palmprint images is that the image acquisition equipment is cheap, which is beneficial to reduce costs, and the second advantage is that the size of the image is small and the processing speed is fast.

Fig. 3 is the original palmprint training ROI image after the present invention cuts. In the cropped palmprint training ROI image, it still contains the main features of the palmprint, such as main lines and wrinkles. The main purpose of using the palmprint training ROI image is to locate the palmprint image, so that the palmprint images collected at different times from the same palm have relatively small displacement and rotation errors, which is convenient for palmprint matching.

Fig. 4 is the schematic diagram that the present invention constructs ROI image collection of palmprint training. Through observation, the maximum rotation error between the palmprint image to be recognized, that is, the test image and the palmprint training image collected in the registration stage is about 10°. Assume that there is a palmprint training ROI image A in the system for a person's palmprint, and the palmprint training ROI image A is rotated 3°, 6°, 9°, -3°, -6°, -9° respectively, Get new training palmprint ROI images A ₁ , A ₂ , A ₃ , A ₄ , A ₅ , A ₆ . If there are multiple palmprint training ROI images in the system, in addition to the palmprint training ROI image A, there are also B, C, etc., then the final palmprint training image set is A, A ₁ , A ₂ , A ₃ , A ₄ , _A5 , _A6 , B, C. Through the construction of palmprint training ROI image set, the rotation error can be effectively compensated.

Fig. 5 is a schematic diagram of the 9×9 size MFRAT of the present invention. In Figure 5MFRAT, the width W of the line is 1, and the number of directions is 6, and the (a)(b)(c)(d)(e)(f) images represent the integral (summation) of lines in different directions , these directions are 0°, 30°, 60°, 90°, 120°, 150°, respectively. Using this MFRAT, direction features can be calculated one pixel (central pixel) at a time.

Fig. 6 is a schematic diagram of a 16×16 size MFRAT of the present invention. In the MFRAT of Figure 6, the width W of the line is 4, and the number of directions is 6. The (a)(b)(c)(d)(e)(f) images represent the integral of lines in different directions. These The directions are 0°, 30°, 60°, 90°, 120°, 150°. Using this MFRAT, the orientation features of 4×4 pixels (dark pixels in the central region) can be calculated at a time. The 4×4 pixels have the same directional feature value, and in the feature image, the 4×4 pixels can be regarded as 1 pixel, so the feature image of the original image of 128×128 has a size of 32×32 pixels.

Fig. 7 is a comparison chart of RPOC and Competitivecode features of the present invention. Among them, (a) image is the original palmprint ROI image, (b) is the palmprint recognition method based on CompetitiveCode, and (c) is the palmprint orientation feature code RPOC. It can be seen that the palmprint orientation feature code RPOC can better reflect the palmprint structural features. It should be pointed out here that the size of Figure 7(b) and (c) is 32×32 pixels, and they are enlarged for the convenience of display. Each small square represents a pixel. In addition, in Figure 7(b) and (c), different gray values represent different orientation values.

Fig. 8 is a schematic diagram of a point-to-region matching operator in the present invention. Where (a) is a point-to-point matching operator, and (b) is a point-to-cross-shaped matching operator. (c) is a point-to-small square matching operator.

Fig. 9 is a distribution diagram of matching values in the verification test of the present invention.

In the verification experiment, the palmprint matching from the same palm is called true matching, namely Genuine matching, and the palmprint matching from different palms is called false matching, namely Imposter matching. The result of the true match is the verification of the legitimate user's legal identity. A fake match is one that pretends to be a real user. Figure 9 is the statistics of the matching values in all true matches and false matches, the center point of the distribution of matching values of true matches is about 0.7, and the center point of distribution of matching values of false matches is about 0.5. It can be seen from Figure 9 that true matching and false matching values can be separated better, and it is difficult for an imposter to pretend to be successful. For a perfect system, the distributions of true and false matches should have no intersections.

Fig. 10 is a distribution diagram of FAR and FRR of the verification test of the present invention.

Biometric recognition technology generally uses three indicators to measure the recognition effect, namely False Acceptance Rate (FAR), False Rejection Rate (False Reject Rate, FRR) and Equal Error Rate (Equal Error Rate, EER). FRR refers to the probability that the system rejects a real user as a counterfeiter; FAR refers to the probability that the system accepts a counterfeiter as a real user. EER refers to the error rate when FAR is equal to FRR. FRR and FAR are two parameters of the same algorithm system, put them in the same coordinates, as shown in the figure, FAR decreases as the threshold increases, and FRR increases as the threshold increases . Generally speaking, there is an inverse relationship between FAR and FRR, the larger the FAR, the smaller the FRR, and vice versa. There is an intersection point between FAR and FRR in the figure, which is the point where FAR and FRR are equivalent under a certain threshold (the equivalent point corresponding to this threshold is called EER). It is customary to use EER to measure the overall performance of the algorithm. For a better palmprint recognition algorithm, it is hoped that under the same threshold, the smaller the FAR and FRR, the better.

Fig. 11 is the ROC curve diagram of the RPOC result comparison of the palmprint direction feature encoding of the present invention. In Fig. 11, the ROC curves of palmprint orientation feature code RPOC and several classic palmprint recognition methods (Palmcode, Fuioncode, Competitivecode) are shown. In the ROC curve, the abscissa is the false recognition rate (FAR), and the ordinate is the correct acceptance rate (Genuine Acceptance Rate, GAR). In the ROC curve, given a false recognition rate FAR value, the greater the value of the correct acceptance rate GAR, the better the recognition rate. From Figure 11, it can be seen that the recognition performance of palmprint orientation feature code RPOC is significantly better than other classical palmprint recognition methods.

Example

(1) Image database

Experiments are carried out in the palmprint database of the Hong Kong Polytechnic University Biometrics Recognition Research Center (PolyU_BRC) by using the algorithm of the present invention. These images were collected twice for men and women of different ages, the average interval between the two collections was 2 months, and about 10 palmprint images were collected for each palm each time. Therefore, there are nearly 20 palmprint images for each palm in the database. The size of the palmprint image is 384×284 pixels.

(2) Image preprocessing

Apply the palmprint preprocessing algorithm proposed by Zhang Dapeng et al. [References: D. Zhang, W.K. Kong, J. You, and M. Wong, "Online palmprint identification," IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(9)( 2003), pp.1041-1050.], intercepting palmprint image center size is the palmprint ROI image block of 128 * 128 pixels, carries out palmprint feature extraction and matching on this palmprint ROI image block.

(3) Construct palmprint training ROI image set

We select the first image of each palm collected as the palmprint training ROI image set. The remaining 19 palmprint images are used as the palmprint test image set. Assume that the palmprint training ROI image A of a certain person is rotated 3°, 6°, 9°, -3°, -6°, -9° respectively to the palmprint training ROI image A to obtain a new palmprint training ROI image Set A ₁ , A ₂ , A ₃ , A ₄ , A ₅ , A ₆ . Then the final palmprint training ROI image set is A, A ₁ , A ₂ , A ₃ , A ₄ , A ₅ , A ₆ , and there are 7 palmprint training ROI images in total.

(4) Use MFRAT to extract the direction feature of palmprint

Use MFRAT to extract the direction feature of palmprint, in the present invention, p is set to 16; N is set to 6; W is set to 4. The final feature image, that is, the palmprint training ROI image, has a size of 32×32 pixels.

(5) Use point-to-area palmprint matching for matching

Use point-to-area based palmprint matching.

In the present invention, B(i, j) is set as a cross-shaped area with an area of 5 pixels (B(i-1, j), B(i+1, j), B(i, j), B(i, j- 1), B(i, j+1)), or a 9-pixel square area ((B(i-1, j-1), B(i-1, j+1), B(i-1, j ), B(i+1, j), B(i, j), B(i, j-1), B(i, j+1), B(i+1, j+1), B(i +1, j-1)).

(6) Analysis of palmprint recognition test results

The palmprint recognition test can be divided into two categories, namely verification (Verification) and identification (Identification). Verification is the process of confirming the identity through one-to-one matching between the collected palmprint and a registered palmprint. As a prerequisite for verification, his or her palmprint must have been registered in the palmprint database. The palmprint is stored in a certain compressed format and linked with its name or its identification (ID, PIN). Then at the comparison site, first verify its logo, and then use the system's palmprint to compare with the palmprint collected on-site to prove that its logo is legal. Verification actually answers the question: "Is he who he claims to be?".

Identification is to compare the collected palmprints with the palmprints in the palmprint database one by one, and find out the fingerprints that match the test palmprints. This is also called "one-to-manymatching". Recognition actually answers the question: "Who is he?".

1. Analysis of verification results

In the verification test, all the palmprint images in the palmprint test image set should be matched with all the images in the palmprint training image set. If the palmprint test image and the palmprint training image come from the same palm, then the matching between them is called Genuine Matching (Genuine Matching), if the test image and the palmprint training image come from different palms, then the matching between them Matching is called Impostor Matching. The result of the matching is a matching value, and the range of the matching value is between [0, 1]. If the matching value exceeds the given threshold, the verification is considered to be passed, otherwise it is rejected. Figure 9 shows the distribution of matching values between Genuine Matching and ImpostorMatching.

Figure 10 and the table below show the FAR and FRR values of the method of the present invention at different thresholds. It can be seen that when the FAR is 4.0×10 ^-5 %, the FRR is only 1.631%. The EER is about 0.16% when the threshold is 0.616. threshold FAR(%) FRR(%) 0.5770.5780.5900.6000.6060.6120.616 8.5307.9512.6370.9790.5020.2450.148 00.0300.0460.0760.0910.1370.167 0.6200.6300.6400.6500.6600.661 0.0880.017 3.6×10 ^-3 4.4×10 ^-4 4.0×10 ^-5 0 0.2130.3200.4420.8691.6311.745

Figure 11 and the table below show the comparison of the recognition results of different palmprint recognition methods. When the FAR is 4.0×10-5%, the FRR of Palmcode is 17.2%, the FRR of Fusioncode is 12.1%, the FRR of Competitivecode is 4.86%, and the FRR of RPOC is only 1.631%. The EER of Palmcode is 0.98%, the EER of Fusioncode is 0.87%, the EER of Competitivecode is 0.47%, and the EER of RPOC is only 0.16%. Much better results than other methods. At present, from available materials, the palmprint direction feature code RPOC of the present invention has obtained a relatively high recognition rate in the field of palmprint recognition. Palmcode Fusioncode Competitive code RPOC FAR(%)FRR(%)EER(%) 4×10 ^-5 17.20.98 4×10 ^-5 12.10.82 4×10 ^-5 4.860.47 4×10 ^-5 1.6310.16

2. Analysis of identification results

The first image of each palm is used as the palmprint training image, and the remaining 19 images are used as the palmprint test image for the recognition test. The recognition accuracy of palmprint direction feature code RPOC is 98.12%, while the recognition accuracy of PalmCode, FusionCode and CompetitiveCode are 95.41%, 96.46% and 97.85% respectively. In the palmprint recognition test, the palmprint direction feature code RPOC also achieved a good recognition rate.

3. Storage capacity

The size of the feature image of palmprint direction feature coding RPOC, that is, the palmprint training image, is 32×32, and the value of each pixel may be one of the six numbers of 1, 2, 3, 4, 5, and 6. Then use 3 bits for each pixel to represent these values, such as using 001 to represent 1, 010 to represent 2, 011 to represent 3, 100 to represent 4, 101 to represent 5, and 110 to represent 6. To store the direction feature of a palmprint training image in this way, the number of bytes that needs to be used is (32*32*3)/8=384bytes. It can be seen that the storage capacity of the palmprint orientation feature code RPOC is very small, which is very suitable for real-time applications.

4. Processing speed

In addition to the high-precision recognition rate, another major advantage of the palmprint direction feature coding RPOC is its very fast processing speed, because when using MFRAT for feature extraction, it mainly uses addition operations, thus reducing the time overhead of the processor. All experiments were completed on a personal computer with a Pentium processor with a main frequency of 2.4GHZ and a 256M memory, and the programming platform used was Visual C++. The following table lists the average time required for palmprint direction feature encoding RPOC algorithm preprocessing, feature extraction and matching. The average response time of an identity verification using the palmprint orientation feature coding RPOC method is less than 0.4 seconds. The feature extraction time of palmprint direction feature coding RPOC is only 50 milliseconds, while that of CompetitiveCode is 200 milliseconds, and the feature extraction time of palmprint direction feature coding RPOC is only a quarter of that of CompetitiveCode. method processing time Preprocessing Time Feature Extraction Time Matching Time RPOCCompetitive CodeRLOC 316ms50ms200ms2.5ms

Claims (4)

1, a kind of based on directional characteristic palm grain identification method, comprise a: palm-print image capture, the user carries out palm-print image capture by harvester, obtain can be used for the further palmprint image gray matrix of processing, palmprint image in the registration phase collection is called the palmmprint training image, is called the palmmprint test pattern at the palmprint image of cognitive phase collection;

B: palmprint image pre-service, at first by location palm, finger position, palmprint image is rotated correction, train ROI (Region of Interest) image in the square region of the centre of palmprint image cutting 128 * 128 pixels as palmmprint then, the palmmprint training ROI image in last square shaped zone carries out feature extraction and coupling, it is characterized in that this method also comprises:

C: structure palmmprint training ROI image set, if palmmprint to the someone, there is width of cloth palmmprint training ROI image A in the system, to palmmprint training ROI image A respectively anglec of rotation α be 3 °, 6 °, 9 ° ,-3 ° ,-6 ° ,-9 °, obtain postrotational palmmprint training ROI image, promptly form new palmmprint training ROI image set A ₁, A ₂, A ₃, A ₄, A ₅, A ₆, also including several palmmprint training of B, C ROI image in the system of setting up departments, then last palmmprint training ROI image set is A, A ₁, A ₂, A ₃, A ₄, A ₅, A ₆, B, C, the structure by palmmprint training ROI image set can effectively compensate rotation error;

D: palm print characteristics extracts-sets up palmmprint direction character coding RPOC (RobustPalmprint Orientation Code)

The principal character of palmmprint is the line feature, and these lines have directivity, and promptly direction character can be expressed the essential structure of palmmprint, extracts directional characteristic MFRAT and is described below:

Definition Z _p=0,1 ..., and p-1}, wherein p is a positive integer, for limited two-dimensional grid Z ² _pOn real-valued Equation f [x, y], MFRAT is defined as:

$r\left[L_k\right]={\mathrm{MFRAT}}_f\left(k\right)=\underset{(i,j)\&Element;L_k}{\mathrm{\&Sigma;}}f[i,j]$

Wherein, f[x, y] be the gradation of image matrix, L _kFor at two-dimensional grid Z ² _pIn, f[x, y] the straight line formed of some points:

L _k＝{(i，j)：j＝k(i-i ₀)+j ₀，i∈Z _p}

In the following formula, L _kBe straight-line equation, (i ₀, j ₀) be Z ² _pCentral point, k is expressed as L _kSlope; L so _kJust be expressed as through Z ² _pCentral point (i ₀, j ₀) straight line of different directions, L _kAlso has another method for expressing L (θ _k), wherein, θ _kIt is angle value corresponding to k;

R (L in the formula _k) expression is to the L of different directions _kCarry out integration and promptly sue for peace, r (L _k) represented the L of different directions _kEnergy; By comparing r (L _k) calculate the directional information of palmmprint; Select r (L _k) the little direction of intermediate value is as (i ₀, j ₀) final directional information; Formula as follows:

${\mathrm{\&theta;}}_{k(i_0,j_0)}=\arg \left({\min }_k\left(r\right[L_k\left]\right)\right),k=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;N$

In whole palmprint image, by the mobile Z of pixel or a plurality of pixels ² _p, the directional information of whole palmprint image is just calculated, and the directional diagram formula of palmprint image is:

Wherein (i j) is formula θ to k _{K (i, j)}The k value;

In MFRAT, there are three parameters in application, to adjust, be respectively p, N and W, wherein: p has determined two-dimensional grid Z ² _pSize, promptly determined L _kLength; The quantity N of k represents to calculate the quantity of line, if N greatly then calculated amount is big, direction character is few if N is little, and N is between 6～12; L _kWidth W can adjust according to application demand, W is between 1～4;

Use MFRAT to extract the direction character of all palmmprint training ROI images, the palmmprint direction character masterplate that forms palmmprint training ROI image set is deposited in the masterplate database;

E: based on the palmmprint coupling of point to the zone

If the A in the different time sections collection from same palm is a width of cloth palmmprint training image, B is a width of cloth palmmprint test pattern, the size of A and B all is m * n pixel, and further establishing does not have displacement and rotation error, A (i between A and B, j) with B (x, y) be two corresponding point in same position, (i is j) with B (x for A at this moment, y overlaps, i.e. " i=x " and " j=y ", since displacement and rotation error, A (i, j) often and B (x, y) do not overlap, but A (i j) appears at B (x, y) near probability is big, and the palmmprint matching list based on putting the zone of design is shown:

$s(A,B)=(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}A(i,j)\&CircleTimes;\stackrel{\&OverBar;}{B}(i,j))/m\&times;n$

In the formula, s (A, B) matching distance of expression from A to B, the operation of "  " presentation logic, (i is j) with B (i for A, the value of any one pixel j) is equal, then A (i, j)  B (i, j) value is 1, otherwise then be 0, (i is with B (i j) to B, j) be the regional area at center, can be defined as different shapes;

Similarly, the matching distance from B to A is:

$s(B,A)=(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}B(i,j)\&CircleTimes;\stackrel{\&OverBar;}{A}(i,j))/m\&times;n$

Final matching distance is:

S(A，B)＝S(B，A)＝Max(s(A，B)，s(B，A))。

2, according to claim 1 based on directional characteristic palm grain identification method, it is characterized in that: the described palmmprint of establishing the someone, there is width of cloth palmmprint training ROI image A in the system, to palmmprint training ROI image A respectively anglec of rotation α be 3 °, 6 °, 9 ° ,-3 ° ,-6 ° ,-9 °, obtain postrotational palmmprint training ROI image, promptly form new palmmprint training ROI image set A ₁, A ₂, A ₃, A ₄, A ₅, A ₆, also including several palmmprint training of B, C ROI image in the system of setting up departments, then last palmmprint training ROI image set is A, A ₁, A ₂, A ₃, A ₄, A ₅, A ₆, B, C, anglec of rotation α wherein, the quantity that generates new palmmprint training ROI image is according to the actual conditions adjustment.

3, according to claim 1 based on directional characteristic palm grain identification method, it is characterized in that: described in MFRAT, there are three parameters in application, to adjust, be respectively p, N and W, wherein p is set at 16, N is set at 6, W is set at 4, and final palm print characteristics image is that palmmprint training ROI image is 32 * 32 pixels.

4, according to claim 1 based on directional characteristic palm grain identification method, it is characterized in that: described based on the palmmprint coupling of point to the zone, wherein (i j) is set at the cross area that area is 5 pixels (B (i-1 to B, j), and B (i+1, j), B (i, j), B (i, j-1), B (i, j+1)), perhaps the square region of 9 pixels ((B (and i-1, j-1), B (i-1, j+1), and B (i-1, j), B (i+1, j), B (i, j), and B (i, j-1), B (i, j+1), B (i+1, j+1), B (i+1, j-1)).

Images