WO2006014097A1 - Method for automatically determining the conformity of a normalised electronic image with a face image - Google Patents

Abstract

The invention relates to methods for pattern recognition by selecting certain image characteristics and can be used for automatically determining the conformity of an electronic digitised normalised pattern with a face image by matching characteristics thereof with reference characteristics which are obtainable and modifiable by an adaptive method in the course of training. Said invention makes it possible to reduce a visual image processing time and to increase the certainty of a person's identification in a real-time mode. The inventive method for automatically determining the conformity of an electronic digitised normalised pattern with a face image consists in converting a visual image into a digital two-dimensional array of a certain size, in carrying out a training process on the basis of an extended sample (equal to or greater than 10000) of normalised face images, wherein said process consists in image processing by a high-frequency filter in horizontal and vertical directions in order to comput vertical and horizontal derivatives, in determining the function of a quantity value distribution of the derivative matrixes at each point and in determining maximum and minimum threshold values of a range within which the majority of corresponding derivative values are located. Said method also consists in classifying the identifiable image by processing said image with the aid of a high-frequency filter in horizontal and vertical directions in order to compute the vertical and horizontal derivatives of the identifiable image, in verifying the conformity of the derivative values with the threshold values obtainable in the course of training, in assigning a determined value to each point, and, if the sum of said values is greater than a certain threshold determined by experiment, in coming to a conclusion with respect to the presence of the face image on a current image, if not, in drawing an opposite conclusion.

Description

 METHOD OF AUTOMATIC

DEFINITIONS OF THE ACCORDANCE OF THE ELECTRONIC DIGITAL BATH

NORMALIZED IMAGE

FACE PICTURE

Technical field

The invention relates to methods for pattern recognition by highlighting certain characteristics of an image and can be used to automatically determine whether an electronic digitized normalized image matches a face image by comparing its characteristics with the reference ones obtained and modified by the adaptive method during training. Quick and reliable determination of the correspondence of an electronic image to a normalized face image is one of the most important tasks of preliminary image processing and serves as a necessary basis for the successful implementation of further transformations in comparison and identification of a person. State of the art

Currently, there are many ways to recognize and classify images, including the solution of the problem of recognition of the face and its elements. Among them, we note the main ones: methods that use the Method of Fundamental Components (Ріпсірлі Сомропепт Апалусіс - PCA), methods of applying templates of face elements and methods that use neural networks. Each of these methods has both strengths and weaknesses.

A known method of statistical training for detecting an object in an image (Patent US N ° 5710833, CL G06K9 / 00, 1998), in which A PCA method for detecting a face and its elements in an image. In the method, the visual image is converted into digital signals corresponding to the normalized image, the eigenvectors of these signals and the eigenvalues of the covariance matrix of the obtained set corresponding to the given image are calculated. It is proved that the eigenvectors of such a system form an orthonormal set, the expansion of which leads to significant data compression. Based on this representation of the image, methods for detecting objects in the image are constructed that are superior in their reliability and efficiency to methods using templates and normalized correlation. However, it is known that calculations in this way require considerable time, which complicates its use in practical problems, where the image processing time plays a decisive role.

There is also a method of detecting faces in a digital image (Patent US Na 6128397, class G06K9 / 00, 2000), based on the use of neural networks (lane patworks). In this method, the visual image is converted into digital signals corresponding to the normalized image, the received signals are processed, and the normalized digital image is checked for the presence of facial images on it using neural networks that are pre-trained on face images and background images. It is known, however, that the use of methods based on neural networks is associated with large time costs that impede their widespread use in practical real-time problems.

The closest to the proposed method is a method for detecting a face and its elements using deformable standards (Goldstep AJ, Narmop LD, Lesk AB, "Ideptufopf Numap Faces", - Roc. IEEE, Mau 1971, vol 59), in which the visual image is transformed into digital signals corresponding to the normalized image process the received signals, while digital signals are presented in the form of a numerical array of values corresponding to the visual image, which after corresponding normalization is compared in a suitable metric with a standard representing deformed variations of the face or its individual elements. However, this method is also inferior in terms of detection reliability to methods for detecting a face and its elements based on PCA and also requires significant time costs. Thus, all known methods for detecting faces in an image require significant computational and time consuming, are quite complex and time-consuming.

Disclosure of invention

The task to which the invention is directed is the creation of a simple and effective method for automatically determining the correspondence of an electronic digitized normalized image to a face image.

The technical result of the invention is to reduce the processing time of visual images, and increase the reliability of personal identification in real time. This technical result is achieved by the fact that at the first stage the visual image is converted into digital signals corresponding to the normalized image, and these signals are presented in the form of a numerical matrix corresponding to the visual image, after which, according to the invention, the learning process is carried out over an extensive sample (at least 10,000) normalized images of faces, which includes: the calculation of vertical and horizontal derivatives for which the image is processed using a filter in high frequencies (HPF) horizontally and vertically, determining the distribution function of the values of the matrices of derivatives at each point and determining the minimum and maximum threshold values of the interval in which most of the values of the corresponding derivatives are located. Thus, the threshold values of the derivatives at each point in the image are determined, which are then used in the classification. At the second stage, the classification problem is solved, that is, the determination of the presence or absence of normalized image of the face on the incoming image for identification. For this image, vertical and horizontal derivatives are successively calculated and their values are checked for threshold values obtained during training, as a result of which a certain value is assigned to each image point. If the sum of such values exceeds a certain threshold determined empirically, a conclusion follows about the presence of a face image in the current image, otherwise, the opposite conclusion is made.

This method of image classification provides a significant increase in processing speed with high face detection efficiency, which allows the proposed method to be used in real-time identification and access control systems.

Brief Description of the Drawings

The invention is illustrated by the drawings.

Figure l shows typical examples of complete (left) and normalized (right) images from a sample of faces. Figure 2 shows the results of applying a high-pass filter (HPF) horizontally and vertically for the image of Fig.l. In Fig.3 - masks obtained as a result of training in an extensive sample for images processed by the HPF horizontally and vertically. Figure 4 shows a graph of the probability of an incorrect acceptance of an object (Falls Assert Rat-FAR) on the threshold level, and Figure 5 is a graph of the dependence of the probability of skipping an object (Fall Asset Rat-FRR) on the threshold level.

The best embodiment of the invention

The implementation of the method is carried out in two stages. At the first stage, the visual image is converted into digital signals corresponding to the normalized image, the received signals are processed, and the digital signals are presented in the form of a numerical array of values corresponding to the visual image, after which the learning process is carried out using an extensive selection of normalized entities representing the variability of many faces. Figure l shows an example of a complete image from a sample and a normalized image of a face. The learning process consists in determining the maximum value at each point of the processed HPF images from a sample of individuals. FIG. 2 shows the results of applying a horizontal and vertical high-pass filter for the image of FIG. Fig. 3 shows masks obtained in the learning process.

At the second stage, the incoming image is directly classified, which is carried out as follows. The image is processed by the HPF horizontally and vertically. Then, the values at certain points of the image corresponding to the maximum values of the corresponding mask and exceeding a certain threshold determined experimentally are summarized. In the same way, the values at the image points corresponding to the minimum values of the mask, and whose values are less than a certain minimum threshold, are summed. In this way, two coefficients corresponding to the image processed by the horizontal high-pass filter and two image coefficients processed by the vertical high-pass filter are obtained. The end result is determined as the weighted sum of all four coefficients. The threshold was determined empirically, when exceeded, the image belongs to the class of persons, otherwise the opposite conclusion is made.

Industrial applicability

Such a simple approach to solving the classification problem can significantly reduce processing time and increase the reliability of finding people. Computations were performed on an extensive sample of test images. In FIG. Figures 4 and 5 show graphs of the statistical characteristics of the method for the probability of incorrect acceptance of an object (FAR) and the probability of skipping an object (Falls Relay Ratt - FRR) on the threshold level. These characteristics were obtained after testing the method for an extensive sample of images.

Claims

 CLAIM A method for automatically determining the correspondence of an electronic digitized normalized image to a face image, which consists in converting a visual image into a digital two-dimensional array of a certain size, which is done by preliminarily conducting an extensive sampling training procedure (at least 10,000) normalized face images, including: image processing by a high-pass filter horizontally and vertically to calculate vertical and horizontal derivatives, determination distribution function of the values of matrices of values derivatives at each point, and determine the minimum and maximum threshold interval in which the majority of the values of the corresponding derivatives; classify the identifiable image by calculating the horizontal and vertical derivatives of the identifiable image; check the correspondence of the values of derivatives to threshold values obtained in the learning process; each point is assigned a specific value, and if the sum of such values exceeds a certain threshold determined experimentally, a conclusion follows about the presence of a face image in the current image, otherwise, the opposite conclusion is drawn.