A method of iris recognition based on wavelet packet decomposition

Abstract

The present invention discloses an iris recognition method based on wavelet packet decomposition. First, human eye images are collected and iris images are preprocessed to obtain processed iris images. Second, diagonal high-frequency information of iris image wavelet packet decomposition is modulated into iris feature codes by using thresholds, and iris features are extracted. Finally, a Hamming distance classifier is used to classify and match iris features. The method determines whether the two feature codes are from the same iris by calculating the similarity between the current iris feature code and the iris code in the iris template library, thereby classifying the iris image. The iris features extracted by the recognition method of the present invention have good classification performance, thereby improving the accuracy of iris recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of biometric identification, and in particular relates to an iris recognition method based on wavelet packet decomposition. Background Art

[0002] Iris recognition is a type of biometric recognition technology. It has a higher recognition rate and security than biometric recognition technologies such as fingerprints and faces. Feature extraction is a key step in iris recognition. The features for identifying identity are extracted from the iris texture, and then the pattern recognition method is used to confirm or identify the identity. Early iris recognition algorithms mainly include: Daugman's two-dimensional Gabor method and Wildes' Gaussian-Laplacian pyramid method. The Daugman method often has a significant drop in performance when recognizing iris images of poor quality, and the algorithm only represents the features of the iris from a qualitative perspective, and is lacking in representing the detailed features of the iris texture. The Wildes method can only work in the authentication mode (Verification) due to the complexity of the extracted features and the high matching complexity, and has limitations in practical applications. With the development of artificial neural networks, some scholars have also applied neural networks to iris recognition. However, due to the complex structure of artificial neural networks in deep learning, the recognition speed of iris recognition needs to be improved, and traditional shallow networks such as BP neural networks are more likely to fall into local extremes and affect the accuracy of iris recognition.

[0003] Boles uses the integral average of wavelet transform to represent iris features. The algorithm converts two-dimensional iris images into one-dimensional signal extraction features, weakening the correlation between iris images as two-dimensional textures. Lim uses Haar wavelet as the mother wavelet to filter and decompose the iris normalized area four times, modulates the high-frequency area of ​​the fourth layer obtained by decomposition, and forms an 87-bit iris feature code with the mean of the high-frequency coefficients of the wavelet decomposition of the first to third layers. The learning vector quantization (LVQ) network is used to classify the extracted features, with a recognition rate of 98.4%. The iris feature space extracted by Lim's algorithm is small, the operation speed is fast and the storage space is saved, but it only uses 87 binary bits to represent the iris texture, and the recognition rate is difficult to guarantee under massive data. Summary of the invention

[0004] In view of the above problems, the present invention aims to provide an iris recognition method based on wavelet packet decomposition, which modulates the diagonal high-frequency information of the iris image wavelet packet decomposition into an iris feature code using a threshold, and identifies the feature using the Hamming distance.

[0005] To achieve the purpose of the present invention, the technical solution of the present invention is:

[0006] An iris recognition method based on wavelet packet decomposition, characterized by comprising the following steps:

[0007] Step 1: Collect human eye images and perform iris image preprocessing to obtain a processed iris image;

[0008] Step 2: Divide the iris effective area according to the iris texture distribution, and select the feature extraction area in the iris effective area;

[0009] Step 3: Perform two-layer decomposition based on wavelet packet transform on the obtained feature extraction area to obtain diagonal high-frequency information, and perform binary encoding on the obtained diagonal high-frequency information to obtain an iris feature code;

[0010] Step 4: Classify and match the iris features of the obtained iris feature code to finally obtain the classification result.

[0011] Furthermore, the iris image preprocessing operation described in step 1 includes: iris positioning, iris interference detection and normalization processing, and iris image normalization processing.

[0012] Furthermore, the specific operation steps of step 2 include:

[0013] S21: Divide the iris region into R1 and R2 regions, corresponding to the iris pupil and iris ciliary body, respectively;

[0014] S22: preliminarily selecting the R1 region as the effective iris region, and further dividing the R1 region into R3 and R4 regions, and selecting the R4 region with less interference as the effective iris region;

[0015] S23: Select a feature extraction area R5 of a fixed size from top to bottom and from right to left in the effective iris area.

[0016] Furthermore, the formula of the wavelet packet transform described in step 3 is:

[0017]

[0018] Where i = 1, 2, ..., 2j, j = 1, 2, ..., J, J = log2N, N is the number of wavelet packet decomposition layers, f(t) is the signal to be decomposed, It represents the i-th wavelet packet on the j-th layer, that is, the wavelet packet coefficient, and G and H are wavelet decomposition filters.

[0019] Furthermore, the specific operation steps of step 3 include:

[0020] S31: Perform the first-layer decomposition on the selected feature area to obtain a node subgraph;

[0021] S32: performing two-layer decomposition on each node subgraph to obtain image low-frequency information, horizontal high-frequency information, vertical high-frequency information and diagonal high-frequency information of each subnode;

[0022] S32: Select the combination of diagonal high-frequency information of the second layer as the iris feature extraction object, set the threshold to 0 to perform binary encoding on the extracted high-frequency coefficients, and the encoding rule of the feature space element C(i) is:

[0023]

[0024] S33: After encoding, an iris feature code is obtained.

[0025] Furthermore, it is characterized in that the specific steps of classifying and matching the iris features in step 4 include:

[0026] Step 41: Select an iris feature code X from the obtained iris feature code and compare it with the iris template feature code in the iris database. Compare and take the ratio of the inconsistent coding bits generated by XOR operation to the total template bits as the Hamming distance. The calculation formula of Hamming distance is:

[0027]

[0028] Where N is the total number of template bits;

[0029] Step 42: Determine the obtained Hamming distance HD, the smaller the distance is, the more similar the two types of iris patterns are;

[0030] Step 43: Repeat steps 41 and 42 until all iris feature codes are compared to obtain the final classification result.

[0031] Compared with the prior art, this method has the following beneficial effects:

[0032] First, the present invention uses wavelet packet analysis for iris feature extraction. Compared with other iris feature extraction methods such as wavelet analysis, the present invention can extract richer texture feature information to obtain a higher recognition rate, thereby improving the accuracy of iris recognition;

[0033] Second, the present invention uses the Hamming distance classifier to identify the extracted iris feature code, which can obtain a higher correct recognition rate;

[0034] In summary, the present invention proposes an iris recognition method based on wavelet packet multi-scale decomposition, which decomposes the segmented effective iris area using wavelet packets, extracts iris features, and uses Hamming distance to identify the features, thereby ultimately improving the classification performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a structural diagram of the iris recognition principle;

[0036] Figure 2 This is the result of locating the iris edge using the Daugman algorithm;

[0037] Figure 3 This is the result of detecting eyelids using the parabola detection method;

[0038] Figure 4 This is the result of testing eyelashes using the Masek method;

[0039] Figure 5 This is the result image after normalizing the iris image;

[0040] Figure 6 It is the wavelet packet decomposition structure diagram;

[0041] Figure 7 is the wavelet decomposition structure diagram;

[0042] Figure 8 It is a diagram of the iris texture structure;

[0043] Fig. 9 It is the result of dividing the iris sub-region on the original image;

[0044] Fig.10 It is the result of dividing the iris sub-region on the normalized image;

[0045] Fig.11 is the normalized image disturbed by eyelids and eyelashes;

[0046] Fig.12 It is the division result diagram of the iris effective area R4;

[0047] Fig.13 (a)-(b) are the division results of the feature extraction area. Fig.13 (a) is the effective area of ​​iris, Fig.13 (b) Selecting a feature extraction area in the iris effective area;

[0048] Fig.14 It is the tree structure diagram of wavelet packet decomposition;

[0049] Fig.15 This is a schematic diagram of wavelet packet decomposition of the iris feature extraction area;

[0050] Fig.16 This is a schematic diagram of the second-layer diagonal high-frequency information of wavelet packet decomposition;

[0051] Fig.17 This is a graph showing the statistical distribution of diagonal high-frequency information;

[0052] Fig.18 are the sym2 scaling function and wavelet function;

[0053] Fig.19 It is the distribution curve of iris feature Hamming distance;

[0054] Fig. 20 This is a schematic diagram comparing the ROC curves of different algorithms. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0056] Reference Figure 1 It can be known that the basic principle of iris recognition is: iris recognition mainly includes steps such as human eye image acquisition, iris image preprocessing, iris feature extraction, and iris feature matching. Through the above steps, high-resolution and high-quality human eye images containing irises are acquired, and then the iris images are located and normalized. Feature engineering-related algorithms are used to extract easily distinguishable feature codes from iris textures, and the extracted feature codes are matched with sample feature templates in the database to complete identity authentication or recognition.

[0057] Reference Figure 1-20 It can be seen that an iris recognition method based on wavelet packet decomposition includes the following steps:

[0058] Step 1: Collect human eye images and perform iris image preprocessing to obtain a processed iris image;

[0059] The iris image preprocessing operation includes: iris positioning, iris interference detection and normalization processing, and iris image normalization processing, specifically:

[0060] 1. Iris positioning:

[0061] Iris positioning refers to the positioning and detection of the boundaries between the iris and sclera, as well as between the iris and pupil. Since the collected iris image also contains the human eye area, in order to facilitate subsequent feature extraction and recognition, it is necessary to segment the iris area from the human eye image. Iris positioning plays a fundamental role in whether the iris can correctly identify the identity. Iris positioning algorithms are mainly divided into two categories according to the positioning principle: one is the calculus method based on the iris edge gradient represented by Daugman, and the other is the Hough transform method represented by Wildes. The present invention uses Daugman's calculus method to locate the iris. The core of this method is to use an effective calculus operator to calculate the iris edge parameters. The calculus operator is:

[0062]

[0063] Where: * is the convolution operation, G σ (r) is the Gaussian function used to smooth the image, r is the radius of the circle detection operator, and I(x,y) is the grayscale value of the iris image at point (x,y). The Gaussian function in formula (1) is:

[0064]

[0065] Where r0 is the center of the Gaussian function and σ is the standard deviation of the Gaussian function.

[0066] Traverse all points on the iris image, and calculate the gradient of the average pixel value of the entire circle at each step of the radius r from small to large. The value corresponding to the maximum gradient is used as the positioning parameter of the inner and outer boundaries of the iris;

[0067] The convolution operation is used to smooth the iris image and remove the influence of noise. The size of the filter template will affect the positioning accuracy. The Daugman method is used to detect the iris edge and the positioning results are shown in the attached figure. Figure 2 shown.

[0068] 2. Iris interference detection and normalization processing:

[0069] The iris area extracted by iris positioning contains interference from eyelids, eyelashes and light spots. If these interference noises are not removed, they will be mixed with the real iris texture features in the subsequent iris feature extraction and encoding process, thereby affecting the accuracy of iris recognition. Since the three-dimensional space search is reduced to a one-dimensional space search and the speed is faster, the present invention uses the least squares method to fit the parabola eyelid method to detect the eyelid. Here, this method detects the eyelid, and the results are shown in the attached figure. Figure 3 As shown;

[0070] It is difficult to detect eyelashes in human eye images. The present invention adopts the Masek method to detect eyelashes. This method uses the large difference between the grayscale value of eyelashes and the grayscale value of nearby irises to set a threshold to clear the eyelash area. The processing results are shown in the attached figure. Figure 4 shown.

[0071] 3. Iris image normalization:

[0072] In order to eliminate the differences in size, position and angle of the extracted iris images, the iris images of different sizes and resolutions must be normalized and converted into regions of the same size, thereby eliminating the effects of translation, scaling and rotation on iris recognition and facilitating subsequent matching and recognition. Among the existing normalization models, the most widely used Rubber-Sheet model maps each point in the iris image to the polar coordinate system through formula (3), so that the extracted iris regions of different sizes are transformed into rectangular regions of the same size, such as Figure 5shown.

[0073] Step 2: Divide the iris effective area according to the iris texture distribution, and select the feature extraction area in the iris effective area;

[0074] The distribution of iris texture is shown in the following figure. Figure 8 As shown in the figure, it can be seen that there is a ring-shaped structure called the curling wheel about 1.5mm away from the edge of the pupil in the iris area. This structure roughly divides the iris into two parts: the part close to the pupil is called the iris pupil part; the part close to the outer edge of the iris is called the iris ciliary body part. Because the pupil part has more radial stripes and wrinkles, it appears to have richer texture details in appearance. More iris features can be extracted in this area for recognition;

[0075] According to the distribution of the pupil and ciliary body, the iris area is divided into R1 and R2 areas, corresponding to the approximate iris pupil and iris ciliary body, respectively. Fig. 9 As shown in the figure, the normalized iris image is divided into two parallel non-intersecting sub-areas in the vertical (radial) direction. Fig.10 Considering that the R1 region is the iris pupil part, which contains rich texture features, and the R2 region contains fewer iris texture features and is easily blocked by the upper and lower eyelids, the R1 region is initially selected as the effective iris region.

[0076] However, for an iris normalized image analysis with eyelash interference, as shown in the attached Fig.11 As shown in the figure, the white area is the detected eyelid and eyelash area, which is part of the R1 area. The R1 area is further divided into two parts, R3 and R4. It can be seen from the figure that the R3 area is seriously blocked by the upper eyelid and disturbed by the eyelashes. And the analysis of a large number of original iris images found that the upper eyelid covers the iris more seriously than the lower eyelid, and the upper eyelashes interfere more seriously with the iris area, so the R4 area is selected as the effective iris area;

[0077] As attached Fig.12 As shown, the R4 region is determined based on the detected eyelid and eyelash regions, and the eyelid and eyelash regions are assigned a uniform, high grayscale value (255 for 8-bit grayscale images). Let the upper eyelid and eyelash region be L1, and all points in the L1 region are detected. The point that is the "rightmost" point with the largest ordinate value is P(x1, y1); let the lower eyelid region be L2, and all points in the L2 region are detected. The point that is the "topmost" point with the largest abscissa value is Q(x2, y2).

[0078] Step 3: Perform two-layer decomposition based on wavelet packet transform on the obtained feature extraction area to obtain diagonal high-frequency information, and perform binary encoding on the obtained diagonal high-frequency information to obtain an iris feature code;

[0079] Wavelet transform only decomposes the low-frequency part of the signal layer by layer, and does not further decompose the high-frequency part obtained by decomposition. Therefore, wavelet transform is insufficient in characterizing signals containing a large amount of detail information (fine edges or textures). However, wavelet packet transform can further decompose the high-frequency part of the signal after decomposition, and can perform better feature extraction on iris images containing rich texture detail information. Therefore, the present invention performs feature extraction based on wavelet packet transform.

[0080] Let f(t) be the signal to be decomposed, represents the i-th wavelet packet on the j-th layer, called the wavelet packet coefficient, G and H are wavelet decomposition filters, H is related to the scale function, and G is related to the wavelet function. Then the wavelet packet transform can be defined as:

[0081]

[0082] Where i = 1, 2, ..., 2 j , j = 1, 2, ..., J, J = log2N, N is the number of wavelet packet decomposition layers;

[0083] From the attached Figure 6 and attached Figure 7 It can be seen that in Figure 6 The left node of each subtree represents the low-frequency information after the signal wavelet packet decomposition, and the right node represents the high-frequency information after decomposition. Figure 7 The middle nodes (1,0), (2,0), and (3,0) are the low-frequency parts after wavelet decomposition, and (1,1), (2,1), and (3,1) are the high-frequency parts after wavelet decomposition. Compared with wavelet packet decomposition, the high-frequency parts after wavelet decomposition are not further decomposed;

[0084] Wavelet packet transform is used to perform two-layer decomposition on the selected feature region. The decomposition tree structure is shown in the attached figure. Fig.14 The decomposed wavelet coefficient image is shown in the attached figure. Fig.15 As shown. Fig.14 It can be seen that after the first layer of decomposition, each node subgraph is further decomposed. Taking node (1,1) as an example, it is decomposed into subnode (2,4) image low-frequency information, (2,5) horizontal high-frequency information, (2,6) vertical high-frequency information, and (2,7) diagonal high-frequency information. After wavelet packet decomposition, the image size of each subnode in the second layer is 500;

[0085] Since the iris texture feature is the detail information with grayscale changes, it is mainly expressed by high-frequency information. Therefore, in order to better characterize the iris texture information and eliminate the influence of noise, the present invention selects the combination of the diagonal high-frequency information of the second layer as the iris feature extraction object, and obtains the following Fig.16 The high-frequency information of four nodes (2,3), (2,7), (2,11), and (2,15) is shown in the figure.

[0086] As attached Fig.17 The statistical distribution diagram of the second-layer diagonal high-frequency information value is shown. It can be seen that the extracted diagonal high-frequency information has a total of 2000 high-frequency coefficient values, of which 991 coefficient values ​​are negative and 1009 coefficient values ​​are positive. After decomposition, 1009 coefficient values ​​of the diagonal high-frequency information are positive. After decomposition, the ratio of positive and negative diagonal high-frequency coefficients is about 1:1. The distribution of this information provides a good basis for iris feature coding. The high-frequency coefficients generated after decomposition are used to describe the similarity between the decomposed wavelet and the decomposed signal. The threshold is set to 0 to perform binary coding on the extracted high-frequency coefficients, and the coding rule of the feature space element C(i) is:

[0087]

[0088] Through the above encoding method, a total of 2000-bit iris feature code is formed.

[0089] Step 4: Classify and match the iris features of the obtained iris feature code to finally obtain the classification result.

[0090] The classification and matching of iris features is achieved by calculating the similarity between the current iris feature code and the iris code in the iris template library, and determining whether the two feature codes come from the same iris based on the size of the similarity value, thereby classifying the iris image.

[0091] The classifier used in the present invention is the Hamming distance, which is used to compare the extracted iris feature code X with the iris template feature code in the database. Whether they are consistent, the ratio of the inconsistent coding bits generated by the XOR operation to the total template bit number N is taken as the Hamming distance. The smaller the Hamming distance obtained, the more similar the two types of iris patterns are.

[0092] Example

[0093] The MATLAB R2019a version was used. The iris images were from the CASIA (V1.0) version of the iris database of the Chinese Academy of Sciences. The iris database contains 108 types of human iris images, each type of which includes 6 human iris images. The sym2 wavelet was selected as the decomposition wavelet basis. Its scaling function and wavelet function are shown in the attached figure. Fig.18 shown.

[0094] In the authentication mode, one iris image of each type is selected and its features are extracted as a template, which is compared with other iris features of the same type, generating a total of 108×6=648 Hamming distances of the same type. In the recognition mode, one iris image of each type is selected and its features are extracted as a template, which is compared with other 107 different types of iris features, generating a total of 50×107=5350 different types of Hamming distances. Fig.19As shown in the figure, the spacing between the distribution curves of the same and different Hamming distances is large, with little overlap, and there is a clear distinction pattern. The mean value μ of the distribution of the same Hamming distance same =0.760, standard deviation σ same =0.087. Mean μ of different types of Hamming distance distribution diff =0.560, standard deviation σ diff =0.032.

[0095] The experiment uses the ROC curve to evaluate the performance of the iris recognition algorithm proposed in this invention. Fig. 20 It can be seen that the ROC curves of the iris recognition algorithm of the present invention are compared with those of the classic iris recognition algorithm. The (FAR, False Accept Rate) in the figure represents the false acceptance rate, which is defined as the proportion of iris patterns that do not belong to the same category being mistakenly matched by the system as the same category. The (FRR, False Reject Rate) is the false rejection rate, which is defined as the proportion of iris patterns that belong to the same category being mistakenly matched by the system as different categories. It can be seen from the figure that the recognition algorithm of the present invention is superior to the Lim and Boles algorithms in overall performance.

[0096] In the recognition mode, the correct recognition rate (CRR) of the iris recognition algorithm of the present invention is compared with that of the classic iris recognition algorithm, and the results are shown in Table 1. From Table 1, it can be seen that the recognition rate of the recognition algorithm of the present invention is significantly higher than that of the algorithms of Boles and Lim, and is close to that of the algorithm of Daugman.

[0097] Table 1 CRR values ​​of different algorithms

[0098]

[0099] The contents not described in detail in this specification belong to the prior art known to the professional and technical personnel in this field. Although the present invention patent is described in detail with reference to the above embodiments, for the technical personnel in this field, it is still possible to modify the technical solutions recorded in the above embodiments, or to replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An iris recognition method based on wavelet packet decomposition, characterized in that: The following steps are involved: Step 1: Collect human eye images and perform iris image preprocessing to obtain a processed iris image; Step 2: Divide the iris effective area according to the iris texture distribution, and select the feature extraction area in the iris effective area; Step 3: Perform two-layer decomposition based on wavelet packet transform on the obtained feature extraction area to obtain diagonal high-frequency information, and perform binary encoding on the obtained diagonal high-frequency information to obtain an iris feature code; Step 4: Classify and match the iris features of the obtained iris feature code to finally obtain the classification result; The specific steps of step 2 include: S21: Divide the iris region into R1 and R2 regions, corresponding to the iris pupil and iris ciliary body, respectively; S22: preliminarily selecting the R1 region as the effective iris region, and further dividing the R1 region into R3 and R4 regions, and selecting the R4 region with less interference as the effective iris region; S23: Selecting a feature extraction region R5 of a fixed size from top to bottom and from right to left in the effective iris area; The specific steps of step 3 include: S31: Perform the first-layer decomposition on the selected feature area to obtain a node subgraph; S32: performing two-layer decomposition on each node subgraph to obtain image low-frequency information, horizontal high-frequency information, vertical high-frequency information and diagonal high-frequency information of each subnode; S32: Select the combination of diagonal high-frequency information of the second layer as the iris feature extraction object, set the threshold to 0 to perform binary encoding on the extracted high-frequency coefficients, and the encoding rule of the feature space element C(i) is: S33: After encoding, an iris feature code is obtained; The specific steps of classifying and matching the iris features described in step 4 include: Step 41: Select an iris feature code X from the obtained iris feature code and compare it with the iris template feature code in the iris database. Compare and take the ratio of the inconsistent coding bits generated by XOR operation to the total template bits as the Hamming distance. The calculation formula of Hamming distance is: Where N is the total number of template bits; Step 42: Determine the obtained Hamming distance HD, the smaller the distance is, the more similar the two types of iris patterns are; Step 43: Repeat steps 41 and 42 until all the iris feature codes are compared to obtain the final classification result; And, the formula of wavelet packet transform described in step 3 is: Where i = 1, 2, ..., 2j, j = 1, 2, ..., J, J = log2N, N is the number of wavelet packet decomposition layers, f(t) is the signal to be decomposed, It represents the i-th wavelet packet on the j-th layer, that is, the wavelet packet coefficient, and G and H are wavelet decomposition filters.

2. The iris recognition method based on wavelet packet decomposition according to claim 1, characterized in that: The iris image preprocessing operation described in step 1 includes: iris positioning, iris interference detection and normalization processing, and iris image normalization processing.

Images