BR102014012954B1 - Iris segmentation method, biometric authentication method and biometric authentication system - Google Patents

Abstract

IRIS SEGMENTATION METHOD, BIOMETRIC AUTHENTICATION METHOD AND BIOMETRIC AUTHENTICATION SYSTEM. The present invention is part of the electronics industry, specifically the biometrics industry, as it refers to a method of extraction, modeling and identification of iris for a biometric authentication system, in which the present solution aims to maintain the quality of the segmentation result, however, with acceptable execution times for real-time iris recognition, not previously foreseen in previous techniques.

Description

Field of Invention

[1] The present invention is part of the electronics industry, specifically the biometrics industry, as it refers to a method of extraction, modeling and identification of iris for a biometric authentication system.

State of the Technique

[2] Currently, Information Technology is a fact present in our daily life, involving practically all activities carried out. However, the extensive use of IT has brought risks associated with information security, mainly in relation to confidentiality, integrity, availability, authentication and non-repudiation of information.

[3] With this, it became necessary to carry out an advanced investigation into new resources in terms of information security in order to eliminate or reduce the risks associated with the aforementioned factors.

[4] With regard to authentication, there is the authentication of information, systems and people. Regarding people authentication, there are conventional methods based on something an individual has (e.g. cards and keys) or information that the individual knows (e.g. password). Such methods are not completely reliable and robust, as these objects and information can be stolen, forgotten or borrowed.

[5] Due to the aforementioned factors, some alternative methods have been proposed. Among these, a method known as biometric recognition gained prominence, since it does not have the aforementioned weaknesses.

[6] In biometric recognition, a characteristic of the individual known in the literature as biometric characteristic is used, and with that comes an area of study known as biometrics.

[7] The term biometrics refers to the identification of an individual based on their physical or behavioral characteristics. The success of biometric technology is due to the mandatory physical presence of the individual at the place of identification and the elimination of the need to memorize passwords or carry identifiers that could be easily forgotten or stolen.

[8] In this way, systems are developed in which some human characteristic is collected and used for authentication. Such systems are called biometric systems.

[9] These biometric systems can be used in any application where there is a need to perform personal authentication. Some examples where such systems are already in use are banking transactions (ATMs and Netbanking services), employee attendance control, access control to restricted areas, airport identification, login to personal computers and computer systems, electronic voting machines, among others. others.

[10] Biometric characteristics are typically divided into two groups: behavioral and physiological. Among the main behavioral biometrics, the following stand out: signature, typing dynamics and gait. Regarding the physiological aspects, the following can be mentioned: face, iris, fingerprint, hand geometry and voice.

[11] Among the main biometric modalities, fingerprint, face, iris and hand geometry are the most studied and implemented biometric characteristics.

[12] All these systems have some disadvantages compared to other systems. For example, systems that use fingerprints have an analogy with criminal investigations and the need for physical contact with the collector, which can result in rejection by individuals, systems that use facial reading have high intra-personnel variability (for example, in the cases of images of the non-frontal faces), systems that use voice recognition, signatures, or hand geometry need the use of specialized equipment, in addition to understanding variability errors, and systems that understand iris authentication have an expensive and high-precision technology , in addition to highly noisy data.

[13] Among the main biometric characteristics presented, iris biometrics has intrinsic qualities that allow its use as an authenticator. It should be noted that the iris is the area of the eyeball that separates the pupil (center of the eye) from the sclera (part of the white texture).

[14] Iris biometrics comprise qualifications in relation to other forms of biometrics, such as low variability within the same individual (intra-class), high variability between different individuals (inter-class), low genetic influence, high randomness, and high stability throughout life.

[15] Thus, iris biometrics must present distinguishability, allowing differentiation between people, immutability, as it is maintained during life, ease of use, good performance in identification accuracy in an acceptable time, acceptability, presenting little or no resistance to authentication process and good fraud resistance. Furthermore, iris biometrics is an emerging technology that presents new areas to be explored.

[16] These properties have made iris identification, as a means of personal identification, gain great attention in recent times and may come to be the replacement for fingerprint biometrics.

[17] Below are the most relevant documents of the State of the Art that refer to an iris detection and recognition method.

[18] Document CN102629319 describes a robust method of segmenting the iris region based on a specific threshold detector, which can be widely used in many application systems that use iris recognition for identity recognition and security prevention. However, said document uses a variation of the Hough Transform technique, which, unlike the present invention, is avoided due to the high computational cost.

[19] Document CN101866420 is used in the pre-processing process of optical volume iris holographic recognition, and performs the application of optical volume holographic image recognition technology to the iris recognition field. Optical volume holographic iris recognition technology is an application of optical volume holography image recognition in iris recognition, which has the function of parallel iris recognition. By adopting the processing method of the present invention, fast iris positioning is performed, the calculation amount is reduced, the calculation accuracy is improved, and the iris recognition rate is improved. Also, the aforementioned document uses the Hough Transform technique, unlike the cited document, the present application has the advantage of avoiding this transform due to its high computational cost.

[20] Document KR100794361 describes a method for detecting eyelids and eyelashes to improve iris recognition performance is used to detect eyelids from an iris image accurately, detect eyelashes in an iris image accurately, and remove eyelashes detected from the iris image through image interpolation without decreasing the number of iris codes, thus improving iris recognition performance. An area of the iris, eyelids and reflected light are detected in the iris image (S20). The detected eyelids and the reflected light are moved through the interpolation (S30). Both eyelid endpoints included in the iris area are detected using a gray level difference at the edge of the iris area and an eyelid survey area is detected at both eyelid endpoints (S40). Candidate eyelid spots are detected in the detected eyelid survey area (S50). The eyelid is approximately detected using the so-called Hough transform for parabolas, considering the rotation of an eye from the candidate eyelid points (S60). Unlike said document, the present patent application detects the pupil, iris and eyelids, efficiently isolating the iris region.

[21] Document KR20080025610 describes an iris detection method and a device for that purpose that after detecting the pupil boundary and center point, searching the navigation area from the candidate center to the pupil center using the profile information of the row and column of the boundary filtered binary image by inserting the image information that captured the image of the eye part and masking each point at the candidate pupil center point in the candidate navigation center in the binary image and then performing the operation of masking with eight pupil threshold mask templates for each concentric vision center that changed the ray value to the candidate pupil center point origin, performing the masking operation with six iris threshold masking operations corresponding to six positions for concentric circles that have changed the radius value to the detected pupil center point origin, it provides a means of detecting iris to detect the iris boundary area, allowing the eye to detect the iris more efficiently and accurately compared to conventional iris detection methods, even if the eye is obscured by eyebrows or eyelids in the image information that captured the area From the eyes. Although said document also uses the histogram of regions of interest in part of the technique, in the method proposed in the present invention information from an additional step is used: the application of Sobel operators. The pixels are divided into 4 vectors, represented by quadrants and based on opposite quadrants we estimate the region. To locate the pupil, the technique of the authors of that document is based on pixels whose difference is 45°, not to mention the process of locating the angle of each pixel. In the present invention, the differential is to detect two points on the edges of opposite quadrants that are exactly apart by the diameter value, in addition to mentioning the Sobel operators. Also in the cited document, the image histogram may vary according to ambient light, a factor not mentioned in the described technique. Unlike the aforementioned document, in the present invention the method is efficient even in environments that influence image capture, providing reflections in the iris and pupil regions.

[22] Document KR20040026905 is similar to the previous document cited, as it deals with an image quality assessment device and method for iris recognition. Despite also using the histogram of regions of interest in part of the technique, and mentioning the Sobel operators, it is not clear how it is used. However, it does not mention how the information on gradient, angle and division of pixels into quadrants is used, as is done in the method proposed in the present invention, its main objective being the "projection of pixels".

[23] In terms of iris biometry, several approaches have been developed in recent years, among which the methods reported in the works of J. Daugman, “How Iris Recognition Works”, IEEE Trans. On Circuits and Systems for Video Technology, vol. 14, no. 1, pp. 21-30, 2004, and Y. Du, “Partial Iris Recognition Using a 1-D mApproach: Statistics and Analysis,” presented at the 2005 IEEE International Conference on Acoustics, Speech, and Signal Processing in Philadelphia.

[24] In all these works, the first step of an iris recognition process consists of capturing the eye image and improving the acquired image.

[25] Next, the enhanced image is segmented so that the iris region can be outlined. A review of the main methods of segmentation of the iris region used in the prior art can be found in L. Masek's bachelor's thesis, “Recognition of Human Iris PAtterns for Biometric Identification”, from the Dept. The University of Western, Australia, 2003.

[26] These methods, although highly efficient in locating and extracting the iris region, are in general highly computationally expensive, compromising their real-time applications.

[27] The greatest difficulty in using the iris stems from the various noises present in the images collected, such as reflections caused by ambient light, obstructions of the eyelids over the iris region, presence of eyelashes, among others, problems not solved in the aforementioned techniques. .

[28] Due to the aforementioned adverse factors and others that may make it difficult to locate the iris, there are many studies in the literature with the objective of reducing or eliminating such effects.

[29] An example in this sense that can be cited is the state of the art described by Chirayuth Sreecholpech and Somying Thainimit in the work “A Robust Model-based Iris Segmentation”, from the Kasetsart Signal and Image Processing Laboratory, Department of Electrical Engineering from Kasetsart University, Bangkok, Thailand.

[30] The technique described in this document uses the statistical values of mean and standard deviation. The local mean and standard deviation are calculated using the gradient magnitude values of the edge pixels. This calculation is performed after applying an edge enhancement filter (Sobel Operators), presented in item II-A of that document. The mean and standard deviation values are used (in this case, in Formula 2 of the document in question) in search of local maxima and, in this way, model the iris by a circle.

[31] Thus, the state of the art uses an integral formula to detect the boundary, based on the values described in the previous paragraph, that is, there is a huge computational time spent searching for a local maximum, as well as performing integral calculations.

[32] Furthermore, it is important to note that the method described in the prior art is used only to model the iris, and no additional treatment or system is proposed.

[33] Additionally, the document in question also mentions the use of an 11x11 pixel window used over the magnitude gradient, to calculate the local mean and standard deviation.

[34] Specifically with regard to the pupil, an analysis in the literature demonstrates that there is pupil dilation/erosion, a process that occurs in practice due to the presence of light. In this case, a fixed region, pre-defined in absolute pixel values, may no longer be representative if there is any process of pupil dilation/erosion. Thus, if the pupil undergoes processes that change its shape, the dimension of the search region will not be altered, and image treatment errors may occur.

[35] It should also be noted that this technique described in the aforementioned document is similar to the technique developed by John Daugman, cited above.

[36] Another document that can be cited is that of authors J. Raja Sekar, S. Arivazhagan and R. Anandha Murugan, “Methodology For Iris Segmentation and Recognition Using Multi-Resolution Transform”, published at the Third International Conference on Advanced Computing (ICoAC), in 2011.

[37] To locate the pupil, the technique described in this 2011 document uses three lines to find the pupil boundaries and models the iris boundaries by two parabolas. In both cases, the techniques decrease precision, quality, and definition, which can result in parsing errors.

[38] Additionally, as in the prior art, the search region is defined in terms of absolute pixels. The same problem may be present: a fixed window may no longer be representative if there is some process of pupil dilation/erosion.

[39] Finally, a last prior technique to be cited is the one described by M. Vatsa, R. Singh and A. Noore in “Improving Iris Recognition Performance Using Segmentation, Quality Enhancement, Match Score Fusion, and Indexing”, published in IEEE Transactions, “Systems, Man and Cybernetics, Part B: Cybernetics” volume 38, issue 4.

[40] The authors of the article divide the location of the iris into two steps: first, they estimate its location by modeling the iris by an elliptic curve, and later, the authors use equations to more accurately determine the boundaries of the iris. Through these equations, the boundary is drawn point by point.

[41] The authors use the same method to detect the pupil and the iris, not using any edge enhancement filter, in addition to not using the fact of having located the pupil to reduce the iris search region, as performed by the technique proposed in the present invention.

[42] Thus, the technique described in the article entails a great need for computational resources, as well as a high processing time, limiting real-time applications.

[43] It is also possible to cite some relevant state-of-the-art articles involving iris analysis, such as J. G. Daugman, “How Iris Recognition Works” IEEE Trans. On Circuits and Systems for Video Technology, vol. 14, no. 1, pp. 21-30, 2004; J. Daugman, “Probing the uniqueness and randomness of IrisCodes: Results from 200 billion iris pair comparisons.” Proceedings of the IEEE, vol. 94, no. 11, pp 1927-1935, 2006; ICE 2006 Report: P.Jonathon Phillips , W. Todd Scruggs, Alice J. O'Toole, Patrick J. Flynn, Kevin W. Bowyer, Cathy L. Schott, and Matthew Sharpe.“FRVT 2006 and ICE 2006 Large-Scale Results.” NISTIR 7408, March 2007; L. Masek, “Recognition of Human Iris Patterns for Biometric Identification”, BA, Dept. The University of Western, Australia, 2003; R. C. Gonzáles and R. E. Woods, “Digital Image Processing”, ed. Edgard Blücher, 2000; Y. Du, B. Bonney, R. W. Ives, D. M. Etter, and R. Schultz, “Partial Iris Recognition Using a 1-D mApproach: Statistics and Analysis,” presented at the 2005 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , Philadelphia, PA, Mar 2005; R. F. L. Chavez, “A Proposal for Improving the Efficiency of a Human Iris Recognition System”, di Master's thesis, Electrical and Computer Engineering, Dept. Communications, State University of Campinas, São Paulo, 2007; R. Wildes, “Iris Recognition: An Emerging Biometric Techology,” Proceedings of the IEEE, vol. 85, pp. 13481363, 1997; CASIA Iris Image DataBbase (ver 3.0), Institute of Automation, Chinese Academy of Sciences; Roger Fredy Larico Chavez, Yuzo Iano & Vicente Idalberto B. Sablon, “Human Iris Recognition Process: rapid iris location”, Revista Cientifica Periódica - Telecomunicações, Vol. 09, No. 01, November 2006. All these documents also include the same problems previously mentioned.

[44] Finally, the work described by Brito, D. F., and Ling, L.L., “Fast Iris Image Segmentation”, ISSNIP Biosignals and Biorobotics Conference 2010, which presents techniques for segmenting an ocular image, locating the pupil, iris and eyelids. It is important to point out that the methods described in the article, although they present similarities with the methods proposed in the present invention, do not satisfactorily resolve the limitations of the state of the art pointed out and discussed so far, among which the location of the iris, in order to overcome existing adverse conditions (reflections, eyelashes, eyelids, image quality captured in different environments, etc.). At the same time, there is also a need for a balance between the computational cost and the location of the iris, in addition to trying to avoid the use of system training, especially in excess.

[45] The main technical differences of the 2010 article in relation to the present invention are described, in detail and divided by topics, below:

As for pupil shaping:

[46] In the 2010 article, the only estimated model is assumed to be correct, while in the present invention there is an extra step, a model verification, which allows verifying if the estimated model is satisfactory and, if not, are performed until two more attempts. This extra step allows the correct location of the pupil in images even under the most adverse conditions (reflexes, eyelashes, glasses). The improvement comprised in the present invention was carried out precisely for the system to work with noisier images.

As for the technique that models the iris:

[47] For iris location, for example, the regions used to determine iris boundaries are different. In the present invention the regions used are larger.

[48] For the determination of the iris boundaries, for example, while in the article it is possible to use several filters and each filter is applied to the image, the modeling is carried out and then a decision is made whether the iris was detected or not. In the present invention, 3 filters are always applied to the image, modeling takes place and the best estimate among the three is chosen.

[49] For eyelid detection, for example, the search regions (in this case, those that determine which pixels will be used) are also different: in the present invention the regions are smaller and a smaller region in the iris was chosen because of the fact that when the iris is covered, when modeling the borders of the iris, the limit of the iris region will be beyond the eyelid to be located, that is, the eyelid points that must be used are inside the iris, whose border has already been determined by the method. In this way, it is not necessary to use all points below the pupil boundary until the end of the image (as is done in the article). In addition, the reduction of the search area eliminates the need to use unnecessary pixels, allowing performance gains and greater precision in the location.

Regarding the location of the eyelids:

[50] In the case of the upper eyelid, the search region in the present invention was subdivided into three regions and only two used when the estimate is a parabola (by least squares), with the objective of not using pixels that will only disturb the estimate. . This detail is not present and was not foreseen in the 2010 article, and in this way, the alternative proposed in the present invention considerably improves the location of the upper eyelid. While in the present invention both the least squares approximation by a line and a parabola are used, in the article only the approximation by a line is used.

[51] In view of the above, the present invention appears as a new solution for iris image segmentation in order to maintain the quality of the segmentation result, but with acceptable execution times for real-time iris recognition, something not previously provided for in the prior art.

BRIEF DESCRIPTION OF THE INVENTION

[52] The present invention relates to an iris segmentation method comprising the steps and sub-steps of: (a) Acquiring an ocular image; and (b) Segment the acquired image; wherein the segmentation step comprises the sub-steps of: (b1) Determining the boundaries of the pupil region; (b2) Determine the boundaries of the iris; and (b3) Detect the eyelids.

[53] Since:

[54] The step of determining the boundaries of the pupil region, step (b1), comprises the following steps: (b1.1) Filtering and reducing undesirable effects; (b1.2) Edge enhancement of relevant regions; (b1.3) Modeling of the pupil, and said modeling comprises determining, first, the parameter Xp, and later, the parameters (Yp, Rp); and (b1.4) Model verification.

[55] In the step of filtering and reducing undesirable effects (b1.1), a filter of dimension 15x15 is preferably used.

[56] In pupil modeling, step (b1.3), the determination of the parameter Xp is performed using the sets S3 and S4.

[57] In pupil modeling, step (b1.3), the determination of the parameters (Yp, Rp) is performed from the parameter Xp and the sets S1, S2, S3 and S4.

[58] The model verification, step (b1.4), comprises building a circle using the parameters (Yp, Xp) and Rp; and verify if a minimum percentage of points on the circumference corresponds to edge pixels in the image, and when a minimum percentage of points on the circumference corresponding to edge pixels in the image is not verified, step (b1.2) is returned using New values for the cutoff threshold used for the creation of the binary image are created, in which said cutoff threshold comprises values between 20 and 150. Alternatively, when a minimum percentage of points of the circumference corresponding to the edge pixels of the image, you can return to step (b1.1), using at least one filter selected from a filter with dimensions 9x9 and a filter with dimensions 21x21.

[59] The determination of the iris boundaries, step (b2), comprises the following steps: (b2.1) Determination of the search regions and filtering process; (b2.2) Edge enhancement of relevant regions; (b2.3) Modeling of the iris; and (b2.4) Choice of the best model.

[60] The determination of the relevant regions, step (b2.1), is performed using the parameters (Xp, Yp) and Rp.

[61] Iris modeling comprises estimating the parameters of the center (Xi, Yi) and the radius Ri.

[62] The choice of the best model, step (b2.4), comprises choosing the best estimate for the parameters sought, (Xi, Yi) and Ri using the equations

[63] Detection of eyelids, step (b3), comprises the following steps: (b3.1) Determination of eyelid search regions; (b3.2) Edge enhancement of the regions determined in step (b3.1); and (b3.3) Determination of eyelid boundary curves.

[64] In the edge enhancement step (b3.2), a median filter of dimension 15x15 is preferably used.

[65] Further, the present invention describes a biometric authentication method comprising the steps and sub-steps of: (a) Acquiring an eye image; (b) Segment the acquired image; wherein the segmentation step comprises the sub-steps of: (b1) Determining the boundaries of the pupil region; (b2) Determine the boundaries of the iris; and (b3) Detection of eyelids. (c) Normalize the segmented iris region; (d) Extract relevant features from the normalized iris; (e) (re) Registration; and (f) (au) Authentication;

[66] Obviously, such a biometric authentication method authenticates a user if the comparison in step f is positive, and does not authenticate a user if the comparison in step f is negative.

[67] Finally, the present invention also provides a biometric authentication system that comprises a means for acquiring an eye image; a controller configured so as to segment the acquired eye image to locate the pupil, iris and eyelids; extract the iris region, normalize the extracted iris region; sample the relevant features of the normalized iris region; and store the relevant features in a memory, then acquire a new eye image, segment the acquired eye image, normalize the segmented iris image, extract relevant iris features from the normalized iris image, and compare the relevant extracted features with the stored features in the memory.

[68] Again, the biometric authentication system authenticates a user if the match is positive and does not authenticate a user if the match is negative.

[69] The analysis and authentication method described in the present invention comprises steps and sub-steps, and part of these, especially those referring to iris image segmentation, are responsible for enabling a reduction in cost and computational time, with the advantage that the following steps are carried out taking into account the results of the previous step. Thus, it is possible to eliminate much of the image processing of unnecessary regions to segment the image.

[70] Furthermore, the proposed method avoids the need to use techniques that involve machine learning. The fact that there are numerous image capture environments and that total control is not possible would make constant training in the use of such techniques necessary.

[71] In addition, a significant amount of images would be needed for training, which would make the use of a biometric system unfeasible in some cases, such as in a small company where there is not a sufficient number of employees to obtain the minimum number of images for training, as is necessary in several previously mentioned proposed techniques.

[72] As iris biometry comprises some particularities that make the iris have many advantages over other biometric types, such as the fact that each person has a highly constant iris pattern, which remains constant. if relatively stable throughout the life of the human being, unlike other biometric signs, there is no physical contact with the reader, the image collection is carried out quickly and without constraints for the person, the iris, in terms of the quality of biometric attributes, largely satisfies the qualifications required of a very high level biometric system, iris biometrics outperforms most other biometric methods in terms of its high accuracy, robustness and reliability, among others, and in this context, the present invention comprises several advantages over other biometrics techniques.

[73] Thus, the present invention solves the problems of the prior art by describing a new solution for the segmentation of iris image, which maintains the quality of the segmentation result, and also comprises runtimes for iris recognition that allow its use in real time.

BRIEF DESCRIPTION OF THE DRAWINGS

[74] Figure 1A is a representation of an original image obtained by any image acquisition system.

[75] Figure 1B is a representation of an image after edge enhancement without applying the median filter on the original image of Figure 1A.

[76] Figure 1C is a representation of an image after edge enhancement with the application of a 15x15 median filter on the original image of Figure 1A.

[77] Figure 1D is a representation of an image after edge enhancement with the application of a 25x25 median filter on the original image of figure 1A.

[78] Figure 2 is an example of a matrix of the application of a median filter of size 3x3.

[79] Figure 3 is an example of a histogram of an original image obtained by any image acquisition system.

[80] Figure 4A is a representation of an original image obtained by any image acquisition system.

[81] Figure 4B is a representation of an image after applying the 9 x 9 dimensional median filter on the original image of figure 4A.

[82] Figure 4C is a representation of an image after applying the cut-off threshold to the image of Figure 4B.

[83] Figure 4D is a representation of an image after edge enhancement, gradient calculation and new cut threshold in the image of figure 4C.

[84] Figure 4E is a representation of an image after removing the false edges created by applying the filter on the image in Figure 4D.

[85] Figure 4F is a representation of the detected pupil region of the original image after the sub-steps performed in Figures 4A to 4E.

[86] Figure 5 is a table that illustrates the definition of the points that delimit the search regions, in the order (row, column).

[87] Figure 6 is a representation of an image that illustrates the two regions whose pixels will be used to determine the location of the iris.

[88] Figure 7 is a table that illustrates the definition of the points that delimit the eyelid search regions, in the order (row, column).

[89] Figure 8A is an illustration of the upper eyelid search regions.

[90] Figure 8B is an illustration of the lower eyelid scan regions.

[91] Figure 9A is a representation of an original image obtained by any image acquisition system in which the method of the present invention is carried out.

[92] Figure 9B is a representation of an image containing the sub-regions within the upper eyelid search region according to a preferred embodiment of the present invention.

[93] Figure 9C is a representation of a binary image containing the search region with the pixels used in the approximation highlighted according to a preferred embodiment of the present invention.

[94] Figure 9D is a representation of an original image with approximations by a straight line and a parabola for the upper eyelid according to a preferred embodiment of the present invention.

[95] Figure 9E is a representation of an original image with the search region highlighted (binary) and the pixels used in the approximation highlighted according to a preferred embodiment of the present invention.

[96] Figure 9F is a representation of an original image with the pixels used in the approximation highlighted in accordance with a preferred embodiment of the present invention.

[97] Figure 9G is a representation of an original image with approximations by a straight line and a parabola for the lower eyelid according to a preferred embodiment of the present invention.

[98] Figure 10 is a representation of the mapping of the iris region to a normalized format.

[99] Figure 11A is a representation of an iris region modeled by a circle for a reference image.

[100] Figure 11B is a representation of a region of the iris modeled by a circle for a candidate image.

[101] Figure 12A is a representation of an original image obtained by any image acquisition system in which the method of the present invention was performed, resulting in the pupil and iris regions highlighted, as well as the eyelid line and parabola. upper according to a preferred embodiment of the present invention.

[102] Figure 12B is a representation of an image containing pixels from the iris region sampled in accordance with a preferred embodiment of the present invention.

[103] Figure 12C is a representation of an image containing the pixel samples discarded as noise according to a preferred embodiment of the present invention.

[104] Figure 12D is a representation of an image containing the pixel samples actually used according to a preferred embodiment of the present invention.

[105] Figure 12E is a representation of an original image in which the method of the present invention was performed, resulting in eyelid regions located in accordance with a preferred embodiment of the present invention.

[106] Figure 12F is a representation of an original image with pixels from the iris region sampled in accordance with a preferred embodiment of the present invention.

[107] Figure 12G is a representation of an original image with pixels from the pupil region removed in accordance with a preferred embodiment of the present invention.

[108] Figure 12H is a representation of an iris region extracted from an original image. The pixels of this region will be sampled for the construction of the normalized image.

DETAILED DESCRIPTION OF THE INVENTION

[109] A biometric recognition system normally has the following steps: acquire an eye image, segment the acquired image), normalize the segmented iris region, extract relevant features from the normalized iris), (Registration/Authentication).

[110] The acquisition step, also known in the literature as capture, aims to collect the image of the eye of the person to be identified. For this, optical and/or infrared sensors are preferably used.

[111] The next step is pupil detection and this can be divided into four sub-steps: (b1.1) Filtering and reducing undesirable effects; (b1.2) Edge enhancement of relevant regions; (b1.3) Pupil modeling; and (b1.4) Model verification.

[112] In this way, the original input image undergoes a process of filtering and reducing undesirable effects. Then, in order to enhance the edges, the Gradient technique is applied on the filtered image. In the next step, a cut threshold is applied to the image with the edges highlighted and a binary image is generated. The parameters of the circular model are estimated using the binary image.

[113] For this reason, the parameter estimation procedure can be performed repeatedly until interrupted by one of two conditions: (1) a satisfactory result is obtained; (2) a maximum number of attempts was made; in this case, the best estimate is chosen.

[114] It is important to note that even with glasses and reflex interference, the proposed method is robust enough in successful pupil detection.

[115] The sub-step of filtering and reducing undesirable effects consists of applying, using convolution masks, a median filter over the entire input image. There are three main objectives: — Elimination of noise acquired during image acquisition; — Reduction of reflexes in the pupil region; and — Elimination of iris texture.

[116] Tests carried out for the technology in question have shown that it is essential to reduce or eliminate the reflexes present in the pupil region, otherwise circular modeling may not be efficient.

[117] Another function of the median filter is to eliminate iris texture in the pupil detection process. The presence of iris texture can create false edges when applying the edge enhancement technique.

[118] Figures 1A to 1D illustrate edge enhancement after applying a cut-off threshold, with and without the filtering process. By analyzing them, it is possible to verify the importance of applying the filter to eliminate the iris texture for segmentation purposes, since edge points inside the iris, mainly around the pupil, were eliminated.

[119] The application of the median filter consists of, considering the adjacency of a pixel p of the original input image, ordering in increasing order of brightness value the n x n pixels neighboring p.

[120] After that, to form the filtered image, the brightness value of the position of p in the new image will be the median brightness value among the ordered n x n.

[121] Figure 2 shows the application of the 3 x 3 median filter. Some pixel values of the original image are represented, being the p pixel (central, with brightness equal to 10 in the 5 x 5 matrix). The values of the adjacency brightness of p are ordered and the median value is chosen to form the new image.

[122] Filters with several different sizes were tested. However, the best results were obtained with filters of dimensions 9 x 9, 15 x 15 and 21 x 21. In fact, when necessary, more than one filter can be used in the process. However, preferably you start with the 15 x 15 size filter.

[123] The size of the filter influences the reduction of unwanted effects. It is possible that when using a filter of a certain size, small reflections may persist, while the same reflections are completely eliminated by filters of other sizes. The decision on whether to use 1, 2 or 3 filters will be explained later.

[124] On the other hand, the sub-step of enhancing edges of relevant regions aims to isolate the pupil region from the rest of the image. For this, the best cut-off threshold for the creation of a binary image is estimated. The cutoff threshold is obtained based on the distribution of gray levels of the image filtered by the median filter. In other words, the chosen cut-off threshold should be able to preserve as much as possible the original quality and information of the pupil area.

[125] Thus, the result of empirical investigations, carried out with several databases available in the literature, revealed that the gray level eye images have a histogram format similar to that illustrated in Figure 3.

[126] As can be seen, there are three local maxima located in the following gray level ranges: [0-150], [150-200], [200-255]. The first, second and third maxima of the distribution function are results basically constituted by the pixels located in the pupil, iris and sclera, respectively.

[127] Three sub-steps are performed to enhance the edges of the pupil: (1) on the filtered image, apply a cut threshold, called base threshold, and generate a binary image that contains, among other elements, the region of the pupil; (2) enhance edges; and (3) applying a second cut threshold, generating the binary image that will contain the highlighted pupil edges.

[128] The base threshold value will be the gray level such that, given the total number of pixels with gray level values lower than this, the total is equivalent to 70% of all pixels that have gray levels below 150 Let S be the total number of pixels with values less than 150 (first range of the histogram), the threshold is the lowest gray level that satisfies the equation:

[129] Subsequently, a verification step of the estimated parameters is also performed. Thus, new thresholds from the base threshold can be tested if necessary. The value of 70% was empirically determined after analyzing the test results with several databases available in the literature. This percentage was chosen to counterbalance three criteria: (1) processing time; (2) tested thresholds; and (3) rate of correct pupil location. It is important to emphasize that this value is valid for eye images in gray level, not being dependent on the reader or the database.

[130] Thus, this sub-step aims to separate the objects of interest (pupil) from the rest of the image (background). As previously described, the threshold value used is determined from the analysis on the gray level histogram, which is usually an automatic analysis performed by a control element.

[131] The preferred value of 70% can be determined by studying the results obtained by varying the threshold in the 50-100% range in order to preserve the integrity of the pupil region as much as possible.

[132] Preferably, values below the threshold will have their value changed to 0 (black) and will be considered as belonging to some object and above to 255 (white) and will be considered as background pixels. The range for calculating S was set from 0 - 150 due to pupil characteristics. Empirically, observing the pupil, it was verified that they generally do not have pixels with a brightness value greater than 150.

[133] Once the cut-off threshold is determined, a binary image is generated, Ib. On this binary image, the edge enhancement technique is applied using the Sobel operators, represented by the equation below.

. Após estas etapas tem-se uma nova imagem nomeada de Ir. Paralelamente ao realce de bordas, constrói-se uma imagem Ia que contém o valor do ângulo gradiente para cada posição, determinado pela equação

. Sobre a imagem Ib aplica-se um novo limiar cujo valor é dado pela equação

. A imagem resultante conterá as bordas da pupila e outros elementos que não interferem na localização da pupila, enquanto o limiar de corte definido anteriormente tinha por objetivo preservar os pixels da região da pupila, este segundo limiar tem por objetivo preservar os pixels da borda da região da pupila. As figuras 4A a 4F ilustram os resultados obtidos em cada etapa.[134] For edge enhancement, the gradient value used is given by the equation

. After these steps, we have a new image named Ir. Parallel to the edge enhancement, an image Ia is constructed that contains the value of the gradient angle for each position, determined by the equation

. A new threshold is applied to the image Ib, whose value is given by the equation

. The resulting image will contain the edges of the pupil and other elements that do not interfere with the location of the pupil, while the cut threshold defined previously was intended to preserve the pixels of the pupil region, this second threshold aims to preserve the pixels of the edge of the region. of the pupil. Figures 4A to 4F illustrate the results obtained at each stage.

[135] After the pupil detection step, the pupil shaping step is performed. The determination of the boundaries of the pupil region is carried out, preferably, through circular modeling, which consists of estimating the parameters of a circle with center (Xp,Yp) and radius Rp.

[136] For this, images Ib and Ia are used, obtained as described above. The pixels of the highlighted edges, contained in Ib, are initially grouped in 4 groups, namely, S1, S2, S3 and S4. The groups are formed according to the gradient angle values of the highlighted edge pixels. Gradient angle values are contained in Ia. Each group represents one of the four quadrants: Qi: [0.90°); Q2: [90.180°); Q3: [180.270°); Q4: [270,360°]. A highlighted edge pixel will belong to the Qi quadrant, and consequently to the Si group, when its gradient angle has a corresponding value.

[137] Once the four groups have been determined, the estimation of the circumference parameters consists of the search for two points on the circumference capable of providing the most faithful Rp value possible. That is, two points separated by a distance equivalent to the diameter of the circle. This being the case, it is expected that the midpoint indicates the position of the center of the sought-after circle, (Xp,Yp).

[138] Operationally, in order for the estimate to be as accurate and reliable as possible, the search is performed with all possible pairs of points from opposite quadrants. Once all the pairs are formed, the midpoint and the distance between the points of each pair are calculated. In this way, (Xp,Yp) and Rp are determined by the most likely values, that is, those with the highest frequency of occurrence are considered the center and diameter of the sought-after circle. It is trivial to note that the value of Rp will be half the value of the most likely distance.

[139] This parameter estimation method is only susceptible to edge noise when it is not possible to form sets of opposing quadrants. This case is extremely rare since, in general, reflex-type noise reaches only part of the edge of the pupil, partially obstructing the edges of a quadrant.

[140] However, another type of highly harmful interference occurs when the acquired ocular image is obstructed by the upper eyelid covering a considerable part of the pupil. In this case, the previous algorithm is improved with additional capacity to solve this type of problem. Therefore, the resulting algorithm is highly efficient and robust.

[141] The additional capability of the algorithm consists in estimating the unknown parameters using the sets S3 and S4. This, because these are the points of the pupil that are not superimposed by the upper eyelid. Thus, the estimation of the parameters (Xp,Yp) and Rp follows the following sub-steps: 1. Given two points, one belonging to S3 and the other to S4, calculate the midpoint between them; the calculation is performed for all possible combinations of points from S3 and S4 that are on the same line; this second criterion was added in order to reduce the comparison between points at the edge of the pupil with noise points still present. 2. Having performed the calculations for all points, the most probable value will be considered as the best estimate for Xp; 3. Once Xp is determined, for each point of the sets S1, S2, S3 and S4, the distance is calculated in relation to each row of the column Xp. Therefore, for each calculation, two values are generated: the value of the calculated distance, d, for that line i. At the end of the calculations, the pair (i,d) with the highest frequency of occurrence will be considered the estimate for (Yp, Rp)=(i, d).

[142] Initially, the algorithm with the additional capacity was used when the original algorithm did not find the pupil. However, it was observed that the algorithm with the added capability alone is sufficient to detect the pupil in images with and without upper eyelid overlap. Thus, it was decided to use only the algorithm with the additional capacity for all images.

[143] Once the parameters (Xp,Yp) and Rp have been estimated, it is convenient to check whether these estimates are consistent. For this, it is verified if the pixels of the modeled circumference overlap pixels of highlighted edges of the image Ib. Technically, this verification consists of building a circle according to the equations below.

[144] More precisely, for each point (x,y) of the circle calculated using the equations above, it is checked whether its position in the binary image lies on an edge. 360 points are used, varying the angle - from 1°, to cover a complete circle. If 60% or more of the total calculated points correspond to edge points, it is considered to be a good estimate for the parameters. This percentage was chosen in order to counterbalance three criteria: (1) processing time; (2) tested thresholds; and (3) rate of correct pupil location.

[145] If the percentage is not reached, an iterative process starts where new cut-off thresholds are defined. The following thresholds are defined based on the threshold obtained from the histogram of the image as described above. It is worth remembering that the cut-off thresholds can vary from 0 to 255, as grayscale images are used.

[146] Thus, initially, the base threshold is taken as the current threshold. The current threshold is subtracted from 10 units until one of two conditions is satisfied: (1) a coherent estimate has been found or (2) a cut-off threshold value of 20 has been reached. If only the 2nd condition is satisfied , the cut-off thresholds up to 150 are checked by adding 10 units from the base threshold.

[147] If a satisfactory estimate is not found, the cut-off threshold that best satisfies the conditions of this subsection is stored. If necessary, the same procedures are carried out for two other filters of dimensions 9 x 9 and 21 x 21. If a satisfactory result is not found after all attempts, the optimal threshold to be considered for the system to proceed will be the one that produced the best percentage among all.

[148] This procedure does not have a high computational cost, since the amount of pixels remaining after applying the cutoff threshold is small. In this way, the algorithm estimates the necessary parameters using a minimum amount of pixels, which makes the process extremely fast. In addition to this factor, on average few threshold values are tested. Once the boundary of the pupil region is determined, the location of the border of the iris region begins.

[149] Thus, after the acquisition step, the iris segmentation step, the main focus of the present invention, is performed.

[150] To determine the location of the pupil, iris and eyelids, there are some methods most used in the literature, which will be exemplified by way of comparison. Basically, they fall into two categories: (1) model the pupil and iris by non-concentric circumferences; (2) use of training methods, such as Adaboost classifier cascade.

[151] Methods that rely on approximating the iris and pupil to a circumference most often use the Hough Transform. This transform is a widely known algorithm in the image processing area. First, it is necessary to point out that all the processing performed during the Hough Transform takes place on a binary image. In other words, before applying the transform, it is necessary to perform edge enhancement, using Sobel operators, and later to create a binary image that has only two components: object edges and the background.

onde (xc,yc) representa a coordenada do centro da circunferência e r o raio.[152] The Hough transform consists of determining the parameters of a defined geometric figure, that is, a geometric figure which it is possible to equate, such as: parabolas, lines, circles. For this, the basic idea is: given n unknown parameters, equate in order to allow the calculation of the nth parameter from the remaining (n-1) parameters. Thus, to estimate the value of the nth parameter, all possible values for the remaining (n-1) are used. In the case of the circumference, in two-dimensional space, the object can be represented by the equation below,

where (xc,yc) represents the coordinate of the center of the circle and the radius.

[153] The equation above presents three unknown parameters (xc, yc and r), where (x,y) is a coordinate of an image of size N x M. Thus, the idea of the Hough Transform is: for each edge point from the image of the binary image N x M, vary the possible values of two unknown parameters (xc,yc) and calculate the third, r. The values used, (xc,yc), and the calculated, r, are grouped into some data structure, for example a three-dimensional matrix (x, y, r). After performing the calculations, the coordinate set of values (x, y, r) with the highest frequency of occurrence will contain the estimated values for the center and radius of the circle. Operationally, we have: 1. Suppose there is prior knowledge that the object to be identified has a radius whose variation is between 5 and 15 pixels; 2. Initialize with zero all positions of a matrix M1 of size N x M x 15 (the same dimensions of the binary image used and where the circumference to be determined is; note that yc can vary in the interval [0,N]; xc can assume values in the interval [0,M]; the third dimension of the matrix corresponding to the possible size of the ray. 3. For each point (x,y) of the binary image in which the corresponding position has an edge pixel, replace in the above equation the values (x,y), and all the values of xc and yc; for each substitution, r is calculated; for each calculation, the position value (xc,yc,r) is incremented by one of M1; 4. The procedure is performed for all (x,y) edge points of the binary image; 5. After the calculations, the position of M1 that contains the highest accumulated value will be considered the estimate for (xc,yc, r).

[154] There are variations of the Hough Transform. Basically, these variations seek to reduce the search space of the parameters to reduce the computational time. In the case of the circumference, equations in polar coordinates are usually used.

[155] Some methods even use the Hough Transform in their methods to approximate the iris and pupil by a circumference. In addition, they also use the transform to locate the upper and lower eyelids. However, in this case, the eyelids are approximated by parabolas.

[156] A clear disadvantage of the Hough Transform is the high computational cost. Therefore, in real-time iris recognition systems implemented in software, the Hough Transform is not recommended.

[157] Therefore, to overcome this problem, the technique presented in this project aims to propose a fast detection of the iris without using mathematical equations with numerous parameters, as in the Hough Transform, reducing the computational time. With this, allow the use of the final system in real-time applications.

onde I(x,y) representa a imagem do olho em tons de cinza; r o raio a ser estimado; representa um filtro de suavização gaussiano; s é o contorno do círculo de raio r e centro (x0,y0).[158] Another technique that can be used is the integral-differential operator. Daugman, one of the pioneers in the use of the iris as a biometric feature, proposed a method for locating the iris. The Daugman technique uses an integro-differential operator to locate and extract the visible iris. His method treats the problem where the eyelids cover part of the iris region. The equation below defines the operator used by Daugman,

where I(x,y) represents the grayscale image of the eye; r radius to be estimated; represents a Gaussian smoothing filter; s is the contour of the circle of radius r and center (x0,y0).

[159] The operator models the iris by a circle of center (x0,y0) and radius r. For this, the maximum value given by the application of the operator on the image is sought. This procedure is performed iteratively, varying the parameters (x0,y0) and r. Thus, at the end of the process, the border of the iris has been delimited.

[160] The same procedure is performed to locate the pupil. A similar procedure is used to locate the eyelids: the operator integration path is changed from a circumference to an arc.

[161] Another segmentation method is the use of classifiers, proposed by Zhaofeng He et al. In this technique, a method is described in order to be possible to extract the iris even with low quality images. The proposed method is divided into 4 steps: (1) reflections removal and iris detection; (2) pupil location and iris boundary delimitation; (3) location of the eyelids; (4) location of eyelashes and shadows.

[162] To remove the reflections, the authors basically transform the original image into a binary image, starting from an adaptive cut-off threshold. The binary image aims to highlight the regions of reflections. Once highlighted, for each reflex point, four other points are defined, according to specific equations. From the brightness values of the original image of the positions of these four points, using interpolation, the new brightness value for the reflection point is defined. Once the reflections are eliminated, the method detects a region that contains the iris.

[163] To determine the region that contains the iris, the method proposes training in a cascade of Adaboost classifiers. In this training, positive samples (contains the iris region) and negative samples (does not contain) are used. During learning, the detector searches the original image for subwindows at different scales, using 346 features and 5 layers. If there is at least one subregion validated in the 5 layers as positive, the input image is considered as an iris image.

[164] The second step is to determine the pupil and iris boundaries, using the image validated in the previous step. In locating the pupil, the authors use the idea of the traditional Hooke's law (law of physics related to the elasticity of bodies). The proposed method initially assumes a point O located in the center of the pupil. This point is connected by imaginary springs to several other points on the edge of the pupil. In this way, through an iterative process, point O moves inside the pupil due to the forces exerted by the springs to an equilibrium position. Once this position is determined, the pupil is said to be located. After the pupil boundaries are determined, the iris boundaries are delimited. For this, the authors sample edge pixels according to previously defined criteria. Once sampled, it is approximated by a spline curve.

[165] In the third step, the eyelids are located. The proposed procedure applies several filters, and using histograms of regions defined a priori, the sampling of pixels is performed. Subsequently, using the sampled pixels, each eyelid is modeled by a parabola.

[166] The last step consists of locating eyelashes and shadows. This procedure is performed by performing histogram analyses. Histograms are constructed for two regions: a region where it is known a priori that there are no cilia and shadows; and a region where you want to determine whether these effects are present. Through the comparative analysis of the histograms of these two regions, the presence of eyelashes and shadows is determined.

[167] As seen, existing models require high computation, have a large processing time, and still require previous training for correct operation.

[168] Thus, the objective of this step is to determine the boundaries of the iris, assuming that it can be modeled by a circle with center (Xi,Yi) and radius Ri. Some factors make this procedure difficult, such as: reflections at the borders of the iris; possibility of a smooth transition between the iris region and the sclera, especially when using grayscale images; the texture itself: If there is a representative texture, using edge enhancement filters can result in false edges.

[169] Thus, the iris segmentation method proposed by the present invention can be divided into four distinct sub-steps: (b2.1) Determination of search regions and filtering process; (b2.2) Edge enhancement of relevant regions; (b2.3) Modeling the iris; (b2.4)Choose the best model.

[170] Preferably, the edge of an iris is modeled by a circle with center (Xi, Yi) and radius (Ri). Therefore, the presented step aims to estimate these parameters.

[171] The first sub-step is the determination of search regions and filtering process. This step has three objectives: (1) reduction of computational processing time by discarding pixels that will not contribute to the correct and precise location of the iris; (2) elimination of noise acquired during image acquisition and (3) reduction of iris texture.

[172] To achieve the first objective, two search regions are defined. These regions are rectangular in shape and defined by four points described in the table in Figure 5 and in the areas illustrated in Figure 5. In the table, we have: Lm corresponds to the total number of lines in the image; Cm corresponds to the total number of columns in the image.

[173] The other two objectives of this step are achieved by applying median filters of different sizes. The application, using convolution masks, of the median filters occurs on the pixels inside the search regions of the input image. The process of applying a median filter was described earlier during pupil location. Three filters of different dimensions are applied: 9x9, 15x15 and 21x21. In this way, three filtered images are generated.

[174] It is important to note that if the median filter of one of the defined dimensions has already been applied to the pixels of the entire image during the pupil location, it is not applied again. Operationally, a copy of the filtered image is kept in memory during pupil location.

[175] After determining the search regions and filtering process, a sub-step of highlighting the edges of relevant regions is performed.

[176] This subsection aims to highlight the edges within the search regions. To achieve this goal, Sobel operators, as described above, are preferably used. Once the Sobel operators are applied to the pixels inside the search regions, a cut threshold is applied in order to create a binary image, which will contain the highlighted edges.

[177] Experimental results have shown that the chosen cut-off threshold is dependent on the resolution of the acquired image. A threshold equal to the average gradient, calculated as previously described, is used for images with resolutions greater than 400x400 pixels. For lower resolution images, the threshold used is equal to three times the average gradient. The average gradient value is calculated with the pixels inside the search regions.

[178] Subsequently, the iris modeling sub-step is performed. This step consists of estimating the parameters of the center (Xi,Yi) and the radius Ri. The developed algorithm estimates the searched parameters using the pixels of the edges of the search regions of the three filtered and binarized images. Furthermore, for each image used, the parameter Ri is estimated in three different ways, obtaining Ri1, Ri2 and Ri3.

[179] The estimation for (Xi,Yi) and the first value of Ri, occurs in a similar way to the algorithm with additional capacity to locate the pupil. When applying trimming threshold and edge enhancement, pixels are divided into four sets according to their respective gradient angle values.

[180] The sets S3 and S4 are used to estimate Xi. Once the parameter Xi is estimated, all sets are used to estimate (Yi,Ri), as described above. At the end of this first procedure, we have the parameters (Xi,Yi) and Ri1 estimated.

[181] The second estimate for the Ri parameter consists of: calculating the Euclidean distance between each edge pixel of the search regions and the estimated point (Xi,Yi). The distance with the highest frequency of occurrence will be considered the estimate for Ri2.

[182] The third estimate for the Ri parameter involves two operations: (1) a smoothing process is performed on the highlighted edges inside the search regions, in which operationally this process consists of maintaining the state of an edge pixel just in case it has all its neighbors as well as edge pixels; (2) once the smoothing process is performed, Ri3 is estimated in the same way as Ri2, that is, by calculating the Euclidean distance and the most probable value.

[183] Once the parameters (Xi,Yi) and the three radii have been estimated, for each of the three filtered images, the next step is to calculate average iris radii. Three mean radii are obtained, generally defined by the equation below.

2. A melhor estimativa para os parâmetros (Xi,Yi) serão os valores calculados para estes parâmetros utilizando a imagem filtrada pelo filtro da mediana de dimensão z que fornecer a maior distância dz; 3. Utilizando-se os três valores de raios estimados, Ri1, Ri2 e Ri3, para a imagem filtrada pelo filtro da mediana de dimensão z, são calculadas três distâncias, definidas pela equação abaixo. Note que são utilizados os raios médios calculados na subseção anterior;

4. O valor de Ri será o valor de Rix que produzir o menor valor de dx.[184] The last step consists of choosing the best estimate for the parameters sought, (Xi,Yi) and Ri, that is, choosing the best model. The decision involves four sub-steps, defined as follows: 1. The modules of three distance values are calculated. The calculation of each distance involves the estimated ray values using each of the filtered and binarized images, and are generally defined by the equation below;

2. The best estimate for the parameters (Xi,Yi) will be the values calculated for these parameters using the image filtered by the z-dimensional median filter that provides the greatest distance dz; 3. Using the three estimated radii values, Ri1, Ri2 and Ri3, for the image filtered by the z-dimensional median filter, three distances are calculated, defined by the equation below. Note that the average radii calculated in the previous subsection are used;

4. The value of Ri will be the value of Rix that produces the smallest value of dx.

[185] During testing and analysis of the results, it was observed that the algorithm used to estimate the iris radius to a smaller value than expected when iris texture was strongly present; when the iris edge was not well defined, the algorithm estimated the iris radius to a higher value than expected. Thus, the calculations performed by the algorithm seek exactly to avoid these two types of errors and find a satisfactory estimate.

[186] The general idea of the calculations performed is to make an estimate for the iris radius (Ri) that is supposed to be the best; after that, make two other estimates, one being a smaller Ri and one larger than the supposedly best. Thus, dz and dx are calculated and search for the values (Xi,Yi) and Ri in which the three estimates for Ri are closest.

[187] Tests were performed using more than three filters, as well as filters with dimensions different from those shown. However, a greater number of filters did not present a significant improvement in the segmentation rate that would justify the resulting computational expense.

[188] A problem to be solved in iris border detection is that the acquired eye image does not always have a clear contrast between the iris and the sclera (white area of the eyeball). Thus, the present invention solves such problems by taking this factor into account.

[189] Before normalization is carried out, a detection step of the upper and lower eyelids is also foreseen. This procedure may be necessary as part of the iris may be overlapped by the eyelids.

[190] Thus, in the later normalization step, when sampling points from the iris, we should disregard points that are above the top and below the bottom, if they are covered by the eyelids.

[191] We can divide this eyelid detection step into three sub-steps, namely: (b3.1) Determination of eyelid search regions; (b3.2) Edge enhancement of the regions determined in step (b3.1); and (b3.3) Determination of eyelid boundary curves. Preferably, this step is performed twice, once for the location of the upper eyelid and another for the lower eyelid.

[192] The first sub-step of such a step is the determination of eyelid search regions. In this step, to determine the eyelid search region, rectangular windows are delimited that cover the regions where the most relevant pixels for their locations are contained. These rectangular windows are determined through the parameters estimated to model the pupil and the iris (Xp,Yp,Rp,Xi,Yi and Ri).

[193] The table in figure 7 shows the calculation of these limits, where: α = 0.05 x Ri and α1 = 0.05 x Rp. Figures 8A and 8B illustrate the windows for the upper and lower eyelids.

[194] In the case of upper eyelid detection, not all points within the search region are used. The search region itself is divided into three subregions, described in the table in Figure 7, and illustrated in Figure 9B. Thus, in the case of upper eyelid detection, only information contained in sub-regions 1 and 3 is used. The central region (2nd region) has the presence of a significant amount of eyelashes, which can reduce the efficiency of the algorithm in locating the eyelid.

[195] Next, the sub-step of enhancing edges of the regions determined in step (b3.1) is performed. This sub-step consists of applying a median filter, preferably 15x15 in size, within the regions of interest determined in the previous sub-step. If the median filter has already been applied to the entire image during pupil location, the filter does not need to be applied again.

[196] Once the median filter is applied, an edge enhancement process is performed, preferably using Sobel operators on the pixels inside the search region. For the creation of a binary image, the value of the cutoff threshold is equal to the value of the average gradient, as previously explained. At the end of this step, there is a binary image with the edges of the regions of interest highlighted. Figures 9B and 9C illustrate the eyelid regions of interest, with the edges highlighted.

[197] Finally, the sub-step determination of the eyelid boundary curves is performed. This sub-step consists of determining the curves of the upper and lower eyelid boundaries. The boundaries are approximated by a linear segment and a parabolic curve, as illustrated in figures 9D and 9G. Figures 9C and 9E illustrate the pixels used to perform the approximation, marked with a brightness value equal to gray within the search region. Note that the sampled points are inside the search region and the iris modeled by a circle.

[198] The approximation made is by the least squares criterion, as represented by the system of equations below.

, ou seja, a soma dos ,

, ou seja, a soma dos

, ou seja, a soma dos

, ou seja, a soma dos

, ou seja, a soma dos

, ou seja, a soma dos

ou seja, a soma dos .[199] Where n is the number of points, i.e. the number of pixels,

, that is, the sum of ,

, that is, the sum of the

, that is, the sum of the

, that is, the sum of the

, that is, the sum of the

, that is, the sum of the

that is, the sum of .

[200] In this way, the models (line and parabola) are specified by determining the coefficients of the above equations. When observing figures 9D and 9G, it is noted that two curves were approximated. Both curves provide important information about the eyelids and, therefore, can be used together. The straight line provides better details about the central region located over the pupil, but little information about the sides. The parable provides information about the sides and little about the central part.

[201] After this step, a normalization step is performed, which consists of an iris recognition step that aims to compensate for deformations that occurred during the acquisition step, and thus, perform a better feature extraction.

[202] This step is necessary as eye images acquired using different sensors may vary in quality and size. In addition to the sensors, the image collection environment can also cause changes in the images, such as: variation in lighting, distance (which can cause pupil dilation/erosion).

[203] Thus, a better extraction will be possible since the normalization will allow the sampling of points in the iris region from the same spatial location.

[204] Deformations are understood to be the acquisition of different sized iris regions and pupil dilation/erosion. These factors are mainly caused by: shooting in environments that have variations in light levels, focal lengths, head tilts, camera rotations.

[205] Normalization also aims to address non-concentricity between the pupil and iris regions, as can be seen in the specialized literature on the subject. Among the methods most used in the literature to address the problems mentioned, through the normalization of iris regions, are Daugman, Wildes et al. and Boles et al.

[206] The model proposed by Daugman aims to normalize the iris region to correct or minimize the effects caused by pupil dilation, as well as variations in the scale of the iris region, as can be seen in Figure 10. The effects of rotation of the iris image are corrected later, more precisely in the comparison step.

[207] The proposed method performs a mapping from Cartesian coordinates to polar coordinates. Operationally, this procedure consists of: for each point (x,y), belonging to the iris region, it is mapped to a point represented in polar coordinates by the pair (r, ' ), according to the equations below:

[208] Where I is the matrix that contains the pixels of the captured image, I(x,y) is a point of this matrix, I1 is the normalized matrix, (r, ^ ) is the position corresponding to (x,y) in the normalized matrix, represented in polar coordinates, (xp,yp) and (xi,yi) are the coordinates of the centers of the circles that model the pupil and the iris, respectively, where r ε [0,1] and ' ε [ 0.2π].

[209] In this way, a Cartesian point of a detected image comprising a first ray (r1) with a first angle (1) is transformed into a polar point with a second ray (r2) and a second angle 2).

[210] The technique proposed by Wildes et al. consists of comparing two images through the geometric adjustment (translation, rotation and scale change) of one of them.

[211] Operationally, we have: suppose two images that already contain the extracted iris region, called Ia and Id, the first being a candidate input image and the second already stored in an iris database.

[212] An iterative process is carried out in order to align Ia and Id. This procedure is performed through a mapping function (u(x,y), v(x,y)) chosen in order to minimize the equation:

[213] This procedure seeks to correct or minimize the effects of scaling and rotation.

[214] The third most cited technique in the literature was proposed by Boles et al, as can be seen in Figures 11A and 11B. The proposed method uses two iris regions, one being a reference and one to be normalized. In the reference region, a maximum boundary limit of the iris region is determined. In this way, the diameter of the iris (d) is fixed, as can be seen in figure 11A.

[215] Then, the maximum point of the border of the iris region to be normalized is determined in order to have the same diameter as the reference, that is, observing figure 11B, the maximum limit (d1) must be fixed, which is identical to the diameter of the iris (d).

[216] Once the maximum limit (d1) is established within the iris region of the candidate image, concentric circles are constructed whose distances from each other are the same.

[217] During the normalization process, the pixels on the virtual rings are sampled. This technique addresses the problems of pupil dilation/erosion, as well as variations in the scale of the iris region. It is not mentioned by the authors if the method considers rotation effects.

[218] After the normalization of the iris region, two interconnected processes are required: feature extraction and encoding of the extracted data. These processes aim to generate data that will allow the comparison between different iris.

[219] The feature extraction process consists of sampling relevant pixels from the iris region, that is, pixels that contain information about the iris pattern. Once sampled, a coding process can be carried out: the transformation of the extracted data (brightness values) to a new format, in order to allow a quick and accurate comparison between the generated codes.

[220] Despite the conceptual division between the normalization, feature extraction and coding steps, preferably the sampling process (feature extraction) already occurs together with the normalization process. Thus, for the construction of the image that contains the normalized iris region, only the relevant pixels are sampled, simplifying the computational need and reducing the processing time.

[221] Preferably, a review of the main methods in the literature for extracting features and creating biometric samples (coding) is performed.

[222] Basically, three methods are the most cited in the literature: Daugman, Wildes et al and Boles et al. Daugman and Boles et al. use the Wavelet Transform technique, widely known in the area of Digital Signal Processing. Wildes presents a different proposal, using a Laplacian pyramid. An overview of each step is presented below.

[223] Wavelets can be used to analyze iris region data at different scales. Wavelets have the advantage, in relation to the traditional Fourier Transform, of analyzing the data locally. Normally, a bank of Wavelets is used, that is, Wavelet filters at different scales. Boles et al. use Wavelets for feature extraction.

[224] In general, first, there is the normalization of the iris region, in which 16 virtual rings are sampled. Each ring is transformed into a vector of 256 values. After that, each of the vectors is treated as a one-dimensional periodic signal. A transformation with dyadic Wavelets is applied to each signal, that is, Wavelets that vary the scale in powers of 2. It is possible to conclude that only some resolution levels really represent the iris.

[225] After the application of the transform, there is a second processing on the data, called crossing by zeros, to generate a final data set that represents the iris information and allows subsequent comparison.

[226] On the other hand, Gabor filters are able to represent the signal both in space and in frequency. For this, the input signal is modulated by an ejw function, combined with a Gaussian function.

[227] There are several versions of Gabor filters, the best known being the one-dimensional (1D), two-dimensional (2D) and logarithmic scale (Log-Gabor) Gabor filters.

[228] The input signal is decomposed using a pair of quadrature Gabor filters: a real part, specified by a cosine function modulated by a Gaussian; and an imaginary part, specified by a sine function, also modulated by a Gaussian.

onde: (x0,y0) representam uma posição na imagem;

especificam a largura e o comprimento do filtro e (u0,v0) representam a modulação, com frequência espacial dada por:

[229] The real and imaginary filters are also known in the literature as even and odd symmetric components, respectively. The equation below represents a Gabor filter in two dimensions,

where: (x0,y0) represent a position in the image;

specify the width and length of the filter and (u0,v0) represent the modulation, with spatial frequency given by:

[230] To extract the iris characteristics and represent it in a way that allows a future comparison, Daugman uses Gabor filters. In his works, he demonstrates that 2D Gabor filters (in two dimensions) are suitable for representing texture. In this way, they can be used to represent iris features.

[231] Apply the 2D Gabor filter on the image through the convolution process. The filtering result gives complex coefficients. In this way, the real and imaginary values determine the phase in the complex plane. These values are used to create a binary vector, the final format of the iris representation.

[232] The construction process of this vector consists of: for each complex output value after applying the 2D Gabor filter on the normalized image, the phase value Φ is decoded in one of the four possibilities presented in Figure 6a, that is, , given a phase value, check which quadrant it is in and then replace it with two bits. At the end of the process, a vector of bits is generated. This vector is called the iris code, it is illustrated in figure 6b.

[233] Differently, the 2D Gabor Filter has a DC component referring to the even part of the filter as a disadvantage. This occurs whenever the bandwidth is greater than one octave. However, one can obtain a DC component equal to zero for any bandwidth using the Gabor filter on a logarithmic scale.

onde: f0 representa a frequência central e 7 a largura de banda.[234] The frequency response of the Log-Gabor filter is given by the equation below:

where: f0 represents the center frequency and 7 the bandwidth.

[235] Once the Log-Gabor transform is applied to the normalized image, the demodulation process proposed by Daugman can be carried out to generate the iris code.

onde: 7 é o desvio padrão e P é a distância entre um dado ponto e o centro do filtro.[236] The Laplacian pyramid method, proposed by Wildes et al, uses a normalized image decomposition through a Laplacian pyramid. In general, from the normalized image, four other images are produced by applying, by convolution, the filter represented by the equation below:

where: 7 is the standard deviation and P is the distance between a given point and the center of the filter.

[237] Thus, the iris image can be decomposed into a set of four matrices, forming a Laplacian pyramid. The original proposal by Wildes et al uses images of sizes 64x512, 32x256, 16x128 and 8x64.

[238] After the step of extracting the iris features and converting to a format that allows comparison between samples, a step is necessary to verify the degree of similarity between samples. Subsequently, it is necessary to establish decision criteria to define whether two samples belong to the same individual or not. According to bibliographic surveys, there are also three most used techniques in the literature, Hamming Distance (HD - Hamming Distance); Euclidean Distance (WED - Weighted Euclidean Distance); and Normalized Correlation.

[239] The Hamming distance provides a measure of how different two bit patterns are. For this, an XOR (or exclusive) operation is performed between each bit of both patterns. Given two vectors of bits, nominally X and Y, the equation below presents the calculation of the Hamming distance between X and Y. Note that the calculation is normalized to a value in the interval [0,1],

onde: Xi e Yi são dois vetores binários resultantes da extração de características após a demodulação e e y são os vetores ruídos correspondentes.[240] The Hamming distance is then used to check the degree of similarity between two iris codes. As mentioned, there is the presence of noise in the iris region. Thus, it is necessary to disregard the bits of the binary vector referring to noise. For this, the total number of bits is subtracted from the total number of bits considered as noise in one or both of the sampled vectors, represented by the equation below. Normally, works in the literature define as noise the bits that represent the pixels that, although they belong to the iris region, are covered by the eyelids. Also considered as noise are the reflection pixels of the iris region,

where: Xi and Yi are two binary vectors resulting from feature extraction after demodulation and y are the corresponding noise vectors.

[241] After calculating the Hamming distance, it is necessary to compare it with a previously established cut-off threshold. If the Hamming value is less than the threshold value, the result is considered positive. If not, a bit shift is performed to the right or left of the binary vectors of a sample (iris code and noise). That is, the displacement is carried out in one of the samples while the other remains constant. After the displacement, a new calculation of the Hamming distance is carried out and a new comparison with the threshold. The total displacements is a design choice.

onde fi é a i-ésima característica da íris capturada (a íris candidata que pretende-se encontrar uma similar), é i-ésima característica da amostra k armazenada, e < é o desvio-padrão da i-ésima característica da amostra k.[242] The Euclidean distance provides a measure of how similar two samples are. When the comparison between samples is performed with integer values, the Euclidean distance is chosen. Thus, this metric can be specific using the equation below,

where fi is the i-th characteristic of the captured iris (the candidate iris that we intend to find a similar one), is the i-th characteristic of the stored sample k, and < is the standard deviation of the i-th characteristic of the sample k.

[243] Therefore, the characteristics of the unknown iris are compared with the characteristics of the known iris. If the Euclidean Distance between this and the sample k is minimal, an iris similar to the candidate will be found.

[244] The third technique described in the literature and proposed by Wildes is represented by the equation below,

[245] where p1 and p2 are images of size m x n, ;:i and '-7i are the mean and standard deviation of the pixel intensity of p1, and ;<2 and '-y2 are the mean and '-y2 p2 pixel intensity pattern.

[246] Through a linear discriminant, it is possible to minimize the intra-class variance and maximize the inter-class variance and, in this way, make the choice between accept/reject for the candidate iris sample.

[247] In this way, preferably, each binary vector obtained in the step prior to storage is saved in a test file. In this way, a binary representation of its iris is generated for each image of each individual, which is stored. Subsequently, these data are read, stored in vectors and compared. For comparison, the Hamming distance is preferably used.

[248] According to a preferred embodiment of the present invention, the feature extraction step aims to extract features from the iris and convert it to a format that allows subsequent comparison quickly and effectively. Thus, this step was implemented preferably using the Log-Gabor function, widely described in the state of the art.

[249] It should be noted that, as mentioned in the previous subsection, there are hidden pixels belonging to the iris region and, therefore, they are considered as noise.

[250] Thus, to perform the transform, the values of these pixels are assigned as zero, instead of the brightness value of the original image. Such procedure can be seen in figures 12A to 12H.

[251] The highlighted points of figure 12A are sampled to generate the normalized image of figure 12B, which has the pixels of the iris region sampled. The highlighted points in figure 12C are considered noise and marked in the noise image, as can be seen in figure 12D. In a preferred configuration of the present invention, the following procedures are performed in this feature extraction step: 1. Apply the Fourier Transform to each row of the normalized matrix, that is, one row at a time; the resulting image is named If and consists of complex values for each position. 2. Convolute the Log-Gabor filter with the If image; convolution is performed for each row separately, ie first row of If convoluted with first row of filter and so on; 3. After the convolution, the demodulation process is carried out, preferably using the method proposed by Daugman, previously described and widely demonstrated in the state of the art. 4. After the demodulation process, the output is a binary vector representing the iris, known as IrisCode.

[252] It should be noted that to perform the Fourier Transform, the values of the positions in the normalized image that correspond to noise have their brightness value changed to zero.

[253] Figures 12E to 12H illustrate the completion of the entire process for extracting a characteristic iris, in which the parabolas and lines previously found are used to find the iris region, as can be seen in figures 12E and 12F. In this case, the pupil region is also removed, as can be seen in Figure 12G, resulting in normalized and noise-free iris segmentation, as can be seen in Figure 12H.

[254] Thus, briefly the present invention describes an iris segmentation method comprising the steps and sub-steps of: (b1) Determining the search regions and filtering process; (b2) Edge enhancement of relevant regions; (b3) Modeling of the iris; and (b4) Choice of the best model.

[255] The proposed segmentation method provides the creation of a method for locating and extracting an iris region that comprises the steps and sub-steps of: (a) Acquiring an image of an iris; (b) Segment the acquired iris image; (c) Normalize the segmented iris image; and (d) Extract relevant features from the standardized iris;

[256] in which the iris segmentation step comprises the sub-steps of: (b1) Determination of search regions and filtering process; (b2) Edge enhancement of relevant regions; (b3) Modeling of the iris; and (b4) Choice of the best model.

[257] In this method of locating and extracting an iris region, the iris image acquisition step preferably comprises the sub-steps of: (a) Acquiring an iris image; (a1) Filtering and reducing undesirable effects; (a2) Edge enhancement of relevant regions; (a3) Pupil modeling; and (a4) Model verification.

[258] Furthermore, the present invention describes a biometric authentication method comprising the steps of: (re)registration; and (au) authentication.

[259] The registration step comprises the steps and sub-steps of: (a) Acquiring an image of an iris; (b) Segment the acquired iris image; (c) Normalize the segmented iris image; (d) Extract relevant features from the normalized iris; and (e) Store the relevant characteristics. The authentication step comprises the steps and sub-steps of: (f) Acquiring the iris image; (g) Segment the acquired iris image; (h) Normalize the segmented iris image; (i) Extract relevant features from the normalized iris; and (j) Compare the relevant features extracted in step d of the authentication sub-step with the features stored in step e of the registration sub-step.

[260] Stage b of segmentation of the acquired iris image comprises the sub-steps of: (b1) Determination of search regions and filtering process; (b2) Edge enhancement of relevant regions; (b3) Modeling of the iris; and (b4) Choice of the best model.

[261] Obviously, such a biometric authentication method authenticates a user if the comparison in step (f) is positive, and does not authenticate a user if the comparison in step (f) is negative.

[262] Finally, the present invention also provides for a biometric authentication system that comprises a means for acquiring an iris image, a controller configured in order to segment the acquired iris image, normalize the segmented iris image, extract relevant iris features from the normalized iris image and store the relevant features in a memory, then acquire a new image of an iris, segment the acquired iris image, normalize the segmented iris image, extract relevant iris features from the normalized iris and compare the relevant extracted features with the features stored in memory.

[263] Again, the biometric authentication system authenticates a user if the match is positive and does not authenticate a user if the match is negative.

[264] Thus, the invention consists of a set of techniques with the objective of segmenting the iris region of an ocular image. Segmenting means locating, isolating and extracting the iris region of an image.

[265] The input ocular image will contain the following regions: pupil, iris, upper and lower eyelids, and sclera (white part of the eye).

[266] Other elements, considered as adverse factors, since they are not of interest to the problem, will also be present, such as: eyelashes, reflections caused by ambient light during capture and the eyelids over the iris. In the latter case, the iris region will not be fully visible.

[267] The set of techniques of the method described here involves three steps: determining the boundary of the pupil region, determining the boundary of the iris region; determine the location of the eyelids. In all stages, the numerous adverse factors mentioned (lashes, reflexes, occlusions) are considered.

[268] The proposed methods were tested on a set of 26,225 images from the database, as follows: (1) 2,631 images from the Interval set; (2) (16,208) images from the Lamp set; and (3) 7,386 images from the Thousand set.

[269] Table I presents the results obtained in terms of segmentation error rate, detailed for each subset, being images captured from both eyes. In the tables presented, the following notation was adopted: the letter E indicates images of the left eye; the letter D indicates images of the right eye. To verify that the iris was correctly located, each image was visually analyzed. A segmentation error indicates that the iris was not located correctly, either because of an error in the location of the pupil, iris, or eyelids. A non-correct location means that it was not possible to extract the iris region from the original image to perform the authentication, either because: (1) the algorithm incorrectly located the iris/pupil/eyelids, that is, it indicated a region that did not correspond the truth; or (2) despite the iris/pupil/eyelids being located, the algorithm did not precisely determine its boundaries, which did not allow the system to perform biometric authentication. Comparisons were performed with the entire database, although prior art generally only makes comparisons with the best quality ones, for example, the Interval subset.

Tabela II: Comparação com o estado-da-arte - CASIAv3 Interval.

[270] Table II presents the comparison of the proposed method with 3 state-of-the-art methods available in the literature.Table I: Performance in relation to the correct segmentation rate (location and extraction of the iris region).

Table II: State-of-the-art comparison - CASIAv3 Interval.

[271] By way of comparison, using the same database and the same hardware conditions, iris, pupil and eyelid detection techniques provided by previous techniques were tested. Regarding the processing time, the method proposed here provides a gain of 15 times in relation to the most used one.

[272] In the comparison of hit rates, in the segmentation stage, the proposed method presents a gain of 10% in relation to the most used one, maintaining practically the same detection rates at the end of all stages of the recognition process (FRR and FAR), being less than 1% if we discard non-segmented images correctly and 12% if they are not discarded.

[273] Daugman, the lead author in the field, in several documents gives FRR and FAR errors of less than 1%, discarding non-segmented images correctly (the first of his techniques, in one of the first articles). Comparing with this rate, the method proposed here provides a high gain in number of hits, keeping the final recognition rate below 1% (by discarding incorrectly segmented images). Thus, our algorithms improve the segmentation step, which has the advantage of requiring a smaller number of image captures for comparative recognition purposes. And yet, they maintain the recognition rates presented by the main authors in the area.

[274] Finally, the comparison of ROC curves obtained with the results at the end of the recognition process was performed considering the images not correctly detected in the segmentation step, and ROC curves disregarding the incorrectly segmented images. Comparing both curves, it was possible to verify that the segmentation step is important in the whole process, since an image not segmented correctly can lead to an error in the final recognition process.

[275] By correctly discarding unsegmented images, it is possible to maintain the error rates presented in the literature. Thus, as the method proposed here obtains a gain in the segmentation step, by discarding a smaller amount of images, this makes it better than the systems presented in the literature.

[276] In addition, it is also necessary to analyze the execution time factor, since the method proposed by the present invention to segment is faster than those presented in the literature, as discussed.

[277] Correct segmentation is critical to the recognition process. With the new proposed methodology, in addition to reducing segmentation times, increasing the hit rate in this step, maintaining the high accuracy of the following steps in the recognition. With this, it is possible to conclude that the iris detection method is more appropriate to be used in an iris recognition system, being able to replace the current methods, solving the problems of processing capacity, speed and training pointed out in the state of the art. .

[278] Although the invention has been extensively described, it is obvious to those skilled in the art that various changes and modifications may be made without said changes falling outside the scope of the invention.

Claims (15)

1. Iris segmentation method comprising the steps and sub-steps of: (a) Acquiring an ocular image; and (b) Segment the acquired image; wherein the segmentation step comprises the sub-steps of: (c) ) Determining the boundaries of the pupil region; and (d) ) Determine the boundaries of the iris. 2. Iris segmentation method, according to claim 1, CHARACTERIZED by the fact that the determination of the boundaries of the pupil region, step (b1), comprises the following steps: (b1.1) Filtering and reduction of undesirable effects ; (b1.2) Edge enhancement of relevant regions; (b1.3) Modeling of the pupil, and said modeling comprises determining, first, the parameter Xp, and later, the parameters (Yp, Rp); and (b1.4) Model verification. 3. Iris segmentation method, according to claim 2, CHARACTERIZED by the fact that in the filtering and undesirable effects reduction step (b1.1), a 15x15 dimension filter is preferably used. 4. Iris segmentation method, according to claim 2, CHARACTERIZED by the fact that in pupil modeling, step (b1.3), the determination of parameter Xp is performed using sets S3 and S4. 5. Iris segmentation method, according to claims 2 and 4, CHARACTERIZED by the fact that in pupil modeling, step (b1.3), the determination of the parameters (Yp, Rp) is carried out from the parameter Xp and from sets S1, S2, S3 and S4. 6. Iris segmentation method, according to claim 2, CHARACTERIZED by the fact that the model verification, step (b1.4), comprises building a circle using the parameters (Yp, Xp) and Rp; and verify that a minimum percentage of points on the circumference correspond to edge pixels in the image. 7. Iris segmentation method, according to claim 6, CHARACTERIZED by the fact that when a minimum percentage of points of the circumference corresponding to edge pixels in the image is not verified, step (b1.2) is returned using New values for the cutoff threshold used to create the binary image are added. 8. Iris segmentation method, according to claim 7, CHARACTERIZED by the fact that the cut-off threshold comprises values between 20 and 150. 9. Iris segmentation method, according to claim 6, CHARACTERIZED by the fact that when a minimum percentage of points of the circumference corresponding to the edge pixels of the image is not verified, alternatively, step (b1.1) ), using at least one filter selected from a filter with dimensions 9x9 and a filter with dimensions 21x21. 10. Iris segmentation method, according to claim 1, CHARACTERIZED by the fact that the determination of the iris boundaries, step (b2), comprises the following steps: (b2.1) Determination of the search regions and process of filtering; (b2.2) Edge enhancement of relevant regions; (b2.3) Modeling of the iris; and (b2.4) Choice of the best model. 11. Iris segmentation method, according to claim 10, CHARACTERIZED by the fact that the determination of the relevant regions, step (b2.1), is carried out using the parameters (Xp, Yp) and Rp. 12. Biometric authentication method comprising the steps and sub-steps of: (a) Acquiring an eye image; (b) Segment the acquired image; wherein the segmentation step comprises the sub-steps of: (b1) Determining the boundaries of the pupil region; (b2) Determine the boundaries of the iris; and (c) Normalize the segmented iris region; (d) Extract relevant features from the normalized iris; (e) (re) Registration; and (f) (au) Authentication; 13. Biometric authentication system CHARACTERIZED by the fact that it comprises a means of acquiring an eye image; a controller configured to segment the acquired eye image for pupil and iris location; extract the iris region, normalize the extracted iris region; sample the relevant features of the normalized iris region; and store the relevant features in a memory, then acquire a new eye image, segment the acquired eye image, normalize the segmented iris image, extract relevant iris features from the normalized iris image, and compare the relevant extracted features with the stored features in the memory. 14. Biometric authentication system, according to claim 13, CHARACTERIZED by the fact that it authenticates a user if the comparison is positive; and not authenticate a user if the comparison is negative. 15. Biometric authentication method, CHARACTERIZED by the fact that it comprises a segmentation method as defined in claims 1 to 11.

Images