WO2006014095A1 - Method for forming a karhunen-loeve basis for person recognition - Google Patents

Abstract

The invention relates to methods for pattern recognition by producing mathematical characteristics of an entire image by data pre-processing in the course of training. Said invention makes it possible to develop a method for forming a Karhunen-Loeve orthonormalised basis for rapidly and reliably recognising a face or for discriminating said face from a plurality of faces on the basis of the digital image thereof under conditions of different disturbances and distortions in automatic identification, monitoring and access systems, etc. The inventive method uses an information statistical pre-processing which is called a Karhunen-Loeve transformation-KLT consisting in computing covariance matrix eigen vectors and values for image sampling, in iteratively forming a set of images from an initial extended sampling by forming an initial image subset in a random manner, in computing a Karhunen-Loeve basis therefrom, in computing decomposition errors for each image from the initial sampling and in ranging said errors according to the quantities thereof, in consequently selecting a certain number of images having the greatest decomposition errors, in completing the image set with said images and in verifying the iteration end conditions and, thereafter, in computing the final Karhunen-Loeve orthonormalised basis and eigenvalues.

Description

 METHOD FOR KARUNEN-LOEVA BASIS FORMATION FOR NAV ANIA FACE

Technical field

The invention relates to methods for pattern recognition by highlighting the details or characteristics of an image based on data preprocessing during training and can be used to significantly improve the reliability and accuracy of automatic recognition and identification of a person or distinguishing him from many other similar persons by a frontal electronic digitized a photo of his face received from a digital camera or camera in real, specially uncontrolled conditions Barrier-shooting. Currently, there are many ways to recognize a person or his identification among many persons. However, among them it should be noted the methods that are most often used in automatic recognition systems and stand out among other methods for their reliability and simplicity in the implementation and presentation of data.

State of the art

A known method for the detection, recognition and encoding of objects using the space of eigenvectors (US Patent N ° 5710833, CL G06K9 / 00, 1998), which includes the conversion of a two-dimensional normalized image of the object into the space of the orthore-normalized basis of the Karunen-Loeve eigenvectors (KL) obtained as a result of the preliminary statistical processing of information called the Karunen-Loev transform (Kagapep-Loeve Trapsformatiop-KLT). This procedure includes the determination of ownership. vectors and eigenvalues of the covariance matrix for some sample of digital images of the object. It is proved that this image representation is the most optimal from the point of view of information compression among all orthogonal transforms (Fourier, Walsh-Hadamard and others). Moreover, KLTs are also successfully used to create methods of pattern recognition, in particular, to recognize faces that turn out to be more reliable and less dependent on noise and distortion compared to methods based on the use of templates, facial marker points, or normalized correlation.

However, among the many problems that arise in the practical use of such methods for face recognition, there is a problem that has not been given due attention to solving so far. It is known that differences between image classes, or their discriminatory properties, are amplified when these images are presented in the KLT basis. This raises the problem of further enhancing such discriminatory properties by forming the most optimal — from this point of view — sample of images, provided that the sizes of such a sample are finite.

In this case, this is the problem of constructing the Karunen-Loev basis, which is the most optimal from the point of view of reliability of face recognition, by creating a suitable final sample of faces for the subsequent KLT training process. A successful solution to this problem can significantly increase the efficiency and reliability of the automatic recognition systems based on the front face image.

Disclosure of invention

The objective of the invention is to develop an effective method of sampling on a wide variety of face variations in order to construct, based on the use of this method, the optimal, from the point of view of face recognition, Karonen-Loev orthonormalized basis. The technical result of the invention is a significant increase in the reliability and speed of face recognition by its digitized photo image in automatic identification, tracking, access systems, etc. under conditions of various noise and distortion. ^•

Such a technical result is achieved by the fact that in the detection method, which includes converting the visual image into digital signals corresponding to the normalized image, processing the received signals, the digital signals being in the form of an extensive set of frontal images of faces of a certain size with the same distance between the pupils of the eyes, to which according to the invention iteratively apply the procedure of statistical processing of information - KLT, while cores and eigenvalues of the covariance matrix for the current sample of images, which is formed from the initial set of faces by calculating the decomposition errors for each image and ranking them by the values of these errors, then highlighting a certain number of images with the largest decomposition errors, replenishing the set of images with them, and checking the end condition iteration, at the end of which the final desired orthonormal basis and the Karunen-Loeve eigenvalues are calculated. This method of information processing provides a significant increase in the efficiency of face recognition and significantly reduces their dependence on noise and image distortion compared to conventional methods based on KLT, which allows you to create highly efficient and productive automatic systems for identification, tracking, access, etc. Brief Description of the Drawings

The invention is illustrated by the graphic materials.

Figure l shows a general block diagram of an iterative process for calculating the orthogonal basis KLT. Figure 2 presents examples of frontal images of faces from a common sample.

In Fig. 3, examples of images included in the result set. Figure 4 shows the first 6 intrinsic KLT decomposition images. In Fig. 5, examples of images from a test sample of faces. The best embodiment of the invention

The implementation of the method is as follows.

An extensive initial sample of frontal normalized images of faces F is formed (Fig. 1). The formation of the basis is carried out according to an extensive selection of normalized images of faces. By such images we mean images with fixed sizes and with certain eye positions on them, the distance between which is equal to a certain distance in pixels. In addition, images must satisfy certain restrictions on contrast, brightness, and must contain facial images. The generated sample of faces F (figure 2) should reliably reflect the variability of the entire space of faces. A random subset of the images is determined — Fa (FIG. 3). The initial basis of Karunen-Loev is formed on it. Decomposition errors for each face in the sample are calculated, with the exception of images of the subset Fa. Of these, images of persons with the largest errors are selected according to a threshold determined experimentally. Such images are added to the subset Fa and the calculations are cyclically repeated until all decomposition errors are lower than the experimentally determined threshold. The final result of such a procedure is a subset of Fa. Fig. 4 shows the first six intrinsic images of the Karunen - Loev decomposition of the subset Fa.

Industrial applicability

The Karunen-Loev orthonormal basis calculated in this way forms the basis of the front face recognition method image and is widely used in criminal identification systems, automatic control and access, etc.

In FIG. Figure 5 shows a graph of the dependence of the probability of accurate identification on the probability of false identification (Fallsert Raté - FAR), expressed as a percentage obtained as a result of testing the recognition method for a wide selection of images. Numerous test results show the effectiveness of using the computed orthonormal Karunen-Loev basis for face recognition.

Claims

F O R M U L A A method of forming a Karunen-Loev basis for face recognition, including converting a visual image into digital signals corresponding to a normalized image, processing the received signals, the digital signals being represented as a numerical array of values corresponding to the visual image with which preliminary statistical processing is performed, called the development of Karunen-Loev (Karmeme-Loev Trapsformatiop-KLT), in which the eigenvectors and eigenvalues of the covariance are calculated matrix for the selection of images, which is due to the fact that iterative formation of the set of images from the initial large sample was introduced by randomly generating the initial subset of images, calculating the Karunen-Loev basis on its basis, and calculating decomposition errors for each image from the original sample and their ranking according to the values of these errors, with the subsequent selection of a certain number of images with the largest decomposition errors, replenishment of many images and checking the condition for the end of the iteration, at the end of which the finite orthonormal basis and the Karunen-Loeve eigenvalues are calculated