RU2756778C1 - Method for image classification - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: invention relates to the field of information technology, in particular, to a method for image classification. The method for image classification consists in the fact that after forming rows of attributes of standards for each class of the classification object and similar attributes of classified images, probability strings corresponding to the attributes of standards and classified images are formed, multiplied by the significance coefficients selected by the operator, and the decisive rule for determining the class for the classification object is formed in the form of calculation of the minimum sum of the squares of the differences between the values of the elements of the probability strings of standards and classified images, and the classified image is attributed to the class of the standard corresponding to said minimum sum.

EFFECT: technical result is a high speed of image classification without a decline in the accuracy.

1 cl, 1 dwg

Description

The proposed image classification method belongs to the field of computing and robotics, and can be used in modern information systems and technologies that simulate or support the process of data mining, including in image recognition processes focused on the classification of objects, which still remains difficult theoretical and technical problem.

The operation of classification is the process of assigning a classified object to a certain class based on the finding of a given formalized attribute of the object, which is present in images and signals. The implementation of the appropriate recognition methods requires a preliminary statistical analysis of the properties of objects in the assumed observation conditions and the formation of reference descriptions of the recognized classes of objects, specific for each applied recognition problem.

At the same time, in practical problems of recognition, the so-called problem of a priori uncertainty takes place, which consists in the fact that objects of any classes are represented by an unlimited set of realizations, in which none of the microstructural properties of the analyzed signal turns out to be a sufficiently reliable indicator for identifying an object and for its identification. When applied to image analysis, such conditions arise for the following reasons:

- the analyzed objects have to be distinguished not by the signal intensity, but by complex geometrical and topological parameters such as the shape;

- the presence of a large structural diversity of the analyzed objects, for example, when recognizing characters - these are different alphabets and fonts; when robotizing production, this is a wide range of parts and operations with them; in systems of observation and control of the situation - these are all kinds of equipment, people, industrial facilities and communications;

- the presence of unlimited variability of conditions for observing objects in terms of scales, angles, underlying surfaces, lighting conditions.

The aforementioned conditions of a priori uncertainty limit the use of recognition systems with specific a priori standards of objects, to overcome which the so-called structural-linguistic recognition methods are developed.

For example, there is a known "Method for classifying the underlying surface by a multichannel image" (RF patent No. 2314565, G06K 9/62, BI No. 1, 01/10/2008, Mironov M.N.), which consists in the formation of attribute lines of the selected type of classification object and the corresponding line the probabilities of the attributes of the classes of the selected type of the classification object with the subsequent assignment of the classified image to one or another class of the selected image type by means of geometric correction and representation of the image components in a single spatial coordinate system, transformation of the probability density of the brightness values to the Gaussian form of those image components in which it differs from Gaussian, determining the class number of each element of the image, and when forming component images for each element of the component image, determining the posterior probabilities of belonging of this element of the component image to the kth class (k = 1, 2, ..., K). Then, for each image element, the posterior probabilities of belonging of this image element to the kth class are determined by adding the corresponding a posteriori probabilities of the component images and their normalization, the class number of the underlying surface (PP) of each element of the multichannel image is determined by the maximum value of the posterior probability of belonging to the given element of the multichannel image to the k-th class.

The disadvantages of this method are its high labor intensity and the dependence of the classification accuracy on the step of dividing the image into components.

Also known "Method and device for recognizing images of objects" (RF patent No. 2361273, G06K 9/62, BI No. 26, 09/20/2008, S. I. Korostelev, etc.), which consists in the following sequence of actions:

- formation of a set of attributes of the selected type of the reference object of classification in the form of a series of flat images of the object at different angles;

- formation of additional class attributes of the selected type of classification object (for example, size, etc.);

- rotation and formation of flat images of the classified object;

- identification of additional attributes of the classified object;

- assignment of the classified image to one or another class of the selected image type by searching for the coincidence of flat images of the classified object with flat images of the reference object, as well as with additional attributes of the class of the reference image.

The disadvantages of this method are the dependence of the classification accuracy on the number of image rotations and the difficulty of obtaining an estimate of the image coincidence, on which the classification correctness depends.

The closest to the claimed technical solution, selected as a prototype, is the classification method described in the article "An algorithm for constructing decision trees in the presence of contradictions in the data", V.N. Vagin, A.V. Krupetskov, M.V. Fomina // Proceedings of the seventeenth national conference on artificial intelligence with international participation KII-2019, Ulyanovsk, October 21-25, 2019, vol. 2, p. 182-191.

The method consists in the following sequence of actions:

- formation of lines of attributes of reference images for each class from the selected type of classification object;

- formation of lines of attributes of the classified image;

- formation of the decision rule for the selection of a class from the selected type of object;

- formation of a training sample of image classes;

- setting up (training) the image class recognition system by using the generated decision rule for classifying the selected type of object;

- assignment of the classified image to one or another class of the selected image type.

The disadvantages of this method are its complexity, the dependence of the classification accuracy on the correctness of the formation of the training sample and the high duration of the classification associated with the time spent on training personnel.

The objective of the invention is to simplify the image classification method and eliminate the influence on the classification accuracy from the correctness of the training sample formation by eliminating the training operation.

The technical result from solving the problem is to achieve a high speed of classification without deteriorating accuracy.

The problem is solved due to the fact that in the proposed method, as in the method adopted as a prototype, lines of reference attributes are formed for each class of the object of classification and similar attributes of classified images. After that, the classified images are attributed to one or another class of the image reference by applying the generated decision rule. In contrast to the prototype, after the formation of strings of reference attributes for each class of the classification object and similar attributes of the classified images, probability lines corresponding to the attributes of the reference and classified images are formed, multiplied by the significance coefficients selected by the operator, and the decision rule for determining the reference class for the classification object is formed in the form of a definition the sum of the squares of the differences of the elements of the probability lines, multiplied by the coefficients of the significance of the standard and the classified image for the first class of the standard, the sums of the squares of the differences of the elements of the lines of probabilities multiplied by the coefficients of the significance of the standard and the classified image for the second class of the standard and all subsequent classes of the standard are determined, then the minimum of the above amounts and classify the image to the class of the standard corresponding to this minimum amount.

The essence of the proposed method is illustrated by a drawing, where in FIG. 1 shows a block diagram of the device.

The block diagram contains a control unit 1, a block 2 containing a database of standards, a computer vision unit 3, a unit 4 for forming a line of standards attributes, a unit 5 for forming a line of attributes of classified images, a unit 6 for forming a line of probabilities of attributes of standards, a unit 7 for forming a line of probabilities of attributes classified images, block 8 for selecting lines of standards, block 9 for selecting lines of classified images, block 10 containing a database of significance coefficients, block 11 for multiplying by significance coefficients of reference attributes, block 12 for multiplying by significance coefficients of attributes of classified images, block 13 for calculating the squares of value differences elements of lines of probabilities of attributes of standards and classified images, block 14 for calculating the sum of squares of differences in values of elements of lines of probabilities of attributes of standards and classified images, block 15 containing a database of sums of squares of differences of values of elements of lines of probabilities values of the attributes of standards and classified images, block 16 for calculating the minimum sum of squares of the differences of the values of the elements of the lines of probabilities of the attributes of standards and classified images and block 17 for designating the class of the image corresponding to the minimum sum.

The outputs of block 1 are connected to the inputs of blocks 2, 3, 8, 9, 10, 15 and 16. The first output of block 2 is connected to the input of block 4, the second - to the input of block 6, and the first output of block 3 is connected to the input of block 5, the second - to the input of block 7. The output of block 4 is connected to the input of block 6, the output of which is connected to the input of block 8. The output of block 5 is connected to the input of block 7, the output of which is connected to the input of block 9. The output of block 8 is connected to the input of block 11, the other whose input is connected to the output of block 10. The output of block 9 is connected to the input of block 12, the other input of which is connected to another output of block 10. The output of block 11 is connected to one input of block 13, the other input of which is connected to the output of block 12. The output of block 13 is connected with the input of block 14, and its output with the input of block 15, the other input of which is connected to another output of block 1. The output of block 15 is connected to the input of block 16, the other input of which is connected to another output of block 1. The output of block 16 is connected to the input of block 17 , and its output with block 1.

Method description

In accordance with the diagram shown in the drawing, the control unit 1 generates commands for unit 2 to select the first class of the standard of the classified image and for unit 3 to select the attributes of the image with attributes formed in the computer vision system. Block 2 transfers the attributes of the first template to block 4, which generates a string of reference attributes, then transfers the probabilities of these attributes to block 6, which generates a string of probabilities corresponding to the attribute string. Block 3 transfers the attributes of the classified image to block 5, which generates a string of attributes of this image of the same dimension as that of the reference line, then the probabilities of these attributes are fed to block 7, which generates a string of probabilities corresponding to the attribute string. Unit 8, on command from control unit 1, selects a line of probabilities from unit 6 and transmits it to unit 11. Unit 9, on command of control unit 1, selects a line of probabilities from unit 7 and transmits it to unit 12. Unit 10, on command of control unit 1, transmits it to blocks 11 and 12 are the significance factors of the attributes, where they are multiplied by the corresponding probabilities. Blocks 11 and 12 are transmitted to block 13 of the row of probabilities for calculating the squares of the differences of the elements of the rows of the template and the classified image, which are supplied from block 13 to block 14, to determine the sum of the squares of the differences for the first class of the standard. This amount is transferred from block 14 to block 15 for storage.

Then, the control unit 1 generates a command for unit 2 to select the second class of the standard of the classified image, and then, similarly to the previous one, the sum of the squares of the differences for the second class of the standard is calculated, which is also transferred from unit 14 to unit 15 for storage.

Then, the control unit 1 generates a command for unit 2 to select the next class of the standard of the classified image, and then, similarly to the previous one, the sum of the squares of the differences for this class of the standard is calculated, which is also transferred from unit 14 to unit 15 for storage.

The described actions continue until the moment until all classes of the standard are exhausted. In this case, block 15 will contain the sum of the squares of the differences for all classes of the standard. After that, the control unit 1 generates a command for the unit 16 to determine the minimum of all the stored amounts, which is transmitted from the unit 16 to the unit 17 to determine the class number of the reference to which the classified image belongs. This completes the classification process for this image. Unit 17 transmits to control unit 1 a signal that the system is ready to classify the next image.

Claims (1)

A method for image classification, which consists in generating strings of reference attributes for each class of object of classification and similar attributes of classified images and assigning classified images to one or another class of an image reference in accordance with the formed decision rule, characterized in that after the formation of lines of reference attributes for each class of the object of classification and similar attributes of the classified images, the probability lines corresponding to the attributes of the standards and classified images are multiplied by the coefficients of significance selected by the operator, and the decision rule for determining the class of the standard for the object of classification is formed in the form of determining the sum of the squares of the differences of the elements of the lines of probabilities, multiplied by the coefficients of the significance of the standard and the classified image for the first class of the standard, the sums of the squares of the differences of the elements of the probability lines multiplied by the coefficients of the significance of the standard and the classified image for the second class of the standard and all subsequent classes of the standard, then determine the minimum of the above amounts and refer the classified image to the class of the standard corresponding to this minimum amount.

Images