RU2625523C1 - Method for determining protectability of designations as trademarks - Google Patents

Abstract

FIELD: physics.

SUBSTANCE: method uses an artificial neural network consisting of artificial neurons with nonlinear activation functions, as well as an algorithm for converting the input image to the R, G, B and Gray matrices for input to the neural network input.

EFFECT: improving the quality of determining the similarity or difference of images to the degree of mixing and the entry of image elements into the composition of comparative samples.

3 dwg

Description

The invention relates to computer technology and can be used in computer systems to determine the protection of designations as trademarks.

Establishing the protection of a designation by comparing it with an array of previously registered trademarks is a difficult task of recognizing and classifying images, currently resolved using a manual system for classifying image elements (Vienna classification), the correct use of which is possible only with a sufficiently high qualification of a specialist specialist .

Existing methods and algorithms for machine pattern recognition cannot be correctly applied to the task of establishing the protection of a designation due to specific conditions and limitations, such as the need to establish the similarity of images to the degree of confusion and the inclusion of image elements in the composition of comparative samples.

A known method and system for searching for copyright infringement on images (patent WO 2014031022 A1), based on the calculation and comparison of descriptors for the compared images. However, this approach does not allow non-linear analysis of the compared images, which affects the quality of image comparison in practice.

The technical result of the present invention is to improve the quality of determining the similarity or difference of images to the extent of mixing and the inclusion of image elements in the composition of the comparative samples.

The proposed method consists of calculating the degree of similarity of the image being tested for the presence of a sign of eligibility, with each image from the database of registered images.

The degree of similarity of the two images is determined in two stages.

Stage 1.

At this stage, the image is read from the file and prepared for further processing. The read image must be converted by scaling to the resolution x × y pixels specified by the user (x is the image width, y is the image height). For this, the bicubic interpolation algorithm is used.

For the resulting x × y pixel array, the following x × y matrices are formed: R is the matrix of the red level of the image pixels (the Red component in the RGB representation); G - matrix of the green level of the pixels of the image (Green component in the RGB representation); B - matrix of the level of blue color of the image pixels (Blue component in the RGB representation); Gray - a matrix of the gray level of the pixels of the image (the brightness level of the pixel when displaying the image in Grayscale grayscale), calculated by the formula Gray = 0.2989R + 0.5870G + 0.1140B.

Elements of the prepared matrices R, G, B, and Gray have unsigned integer values (unsigned integer 8 bit format - integer data that varies in the range from 0 to 255).

After that, the elements of the matrices R, G, B and Gray are converted to the real format by dividing by 256, that is, the converted matrix elements vary from 0 to 1. The resulting converted matrices are fed to the input of the neural network.

2 stage.

Two groups of matrices R, G, B and Gray of the same size, formed for two compared images using the sequence of steps described in step 1, are input to the neural network.

A neural network of elementary components - neurons depicted in FIG. 1. The neuron depicted in Figure 1 operates as follows. The inputs of the neuron receive signals from inputs 1-1, 1-2, ..., 1-n. The signal from each input is multiplied by the corresponding weight coefficient 2-1, 2-2, ..., 2-n using multiplication blocks 3-1, 3-2, ..., 3-n. All the obtained multiplication results are added up in the adder 4. The multiplication result is converted by the activation function 5. The result is sent to the output of the neuron 6. All inputs of the neuron, weighting coefficients and the output of the neuron must have real values, varying from 0 to 1.

The activation function may have one of the following types.

A. Linear activation function

B. Threshold transfer function

where T is the threshold value (real) specified by the user.

B. Sigmoid function

where A, B are real parameters specified by the user.

G. Hyperbolic tangent

where A, B are real parameters specified by the user.

Weighting factors 2-1, 2-2, ..., 2-n are determined before the start of work, in the process of training a neural network.

In FIG. 2 shows the architecture of a neural network. The neural network architecture depicted in Figure 2 contains an input layer of neurons 8, consisting of neurons 9-1, 9-2, ..., 9s; the first hidden layer of neurons 10, consisting of neurons 11-1, 11-2, ..., 11th; the second hidden layer of neurons 12, consisting of neurons 13-1, 13-2, ..., 13-z; and the output layer of neurons 14, consisting of one neuron 15. Each neuron 9-1, 9-2, ..., 9th, 11-1, 11-2, ..., 11th, 13-1, 13-2, ..., 13-z is the device shown in FIG. 1. Each layer of neurons is associated with only two layers: the previous and the next. The connection between two adjacent layers is as follows: the only output of each neuron of the previous layer is fed to the input of each neuron of the next layer. The number of neurons in the input layer, as well as in each of the hidden layers, is determined by the user before using the work of the invention. The output 16 of neuron 15 is considered the output of the entire network and is the result of its work. At the inputs of neurons 9-1, 9-2, ..., 9th input layer of neurons 8 receives information about the two compared images in accordance with the circuit shown in Figure 3.

Figure 3 depicts in detail the description of the input layer 8 and its interaction with the input image pixel arrays of the module images generated at the first stage. All neurons of the input layer 8 are divided into 4 groups: a group of neurons 21 for processing the matrices R of two compared images supplied through the inputs 17-1 and 17-2, respectively; a group of neurons 22 for processing matrixes G of two compared images supplied through inputs 18-1 and 18-2, respectively; a group of neurons 23 for processing the matrices B of two compared images supplied through the inputs 19-1 and 19-2, respectively; a group of neurons 24 for processing Gray matrices of two compared images supplied through inputs 20-1 and 20-2, respectively. The number of neurons of each group is determined by the user individually before using the work of the invention. Neurons of each group must process all elements of two matrices of two compared images. This means that each neuron from group 21 receives at the input all the elements of the matrix R of the first image and all the elements of the matrix R of the second image; each neuron from group 22 receives at the input all the elements of the matrix G of the first image and all the elements of the matrix G of the second image; each neuron from group 23 receives at the input all the elements of the matrix B of the first image and all the elements of the matrix B of the second image; each neuron from group 24 receives at the input all the elements of the Gray matrix of the first image and all the elements of the Gray matrix of the second image. If the matrices R, G, B, Gray prepared at the first stage of each image had x × elements, then each neuron of groups 21-24 should have 2xy inputs and 2xy weighting factors for them.

The result of this stage is a value of 16 obtained at the output of the neural network, which varies from 0 to 1.

The result of the use of a neural network is a numerical value from 0 to 1, showing the degree of similarity of the compared images. To determine the image protection, it is necessary to compare the tested image with each image from the database of registered images. An image is considered to be protectable if, for any pair of compared images, the threshold value of the degree of similarity set before using the invention was not exceeded.

The proposed method allows non-linear comparison of images due to an artificial neural network, the neurons of which have non-linear activation functions. The result of this approach is to improve the quality of determining the similarity or difference of images, including the establishment of similarity of images to the degree of mixing and the inclusion of image elements in the composition of comparative samples.

Claims (1)

A method for determining the similarity of images, which consists in using an artificial neural network consisting of layers of neurons with non-linear activation functions, the input of which receives the values of the matrix elements of components of red, green, blue and gray colors of two compared images of the same dimension, which process the neurons of the input layer of an artificial neural networks, the outputs of which are connected to the inputs of the neurons of the first hidden layer of the artificial neural network, the outputs of which are connected to the inputs of the neur of the second hidden layer of the artificial neural network, the outputs of which are connected to the neuron inputs of the output layer of the artificial neural network, the output of which, which is the output of the artificial neural network, is represented by a number varying from 0 to 1 and showing the degree of similarity of the two compared images.

Images