RU2848151C1 - Method for signal recognition - Google Patents

Abstract

FIELD: signal recognition.

SUBSTANCE: invention relates to methods for signal recognition. The method is characterised in that the feature vector is additionally fed to a filter system, wherein the number of filters in the filter system and the values of their passbands are determined in accordance with the principle of multiple-scale signal transformation, according to which the passband of each subsequent filter is twice the passband of the previous filter, as a result, a signal feature matrix is obtained at the output of the filter system, and the recognisable signal is identified by comparing its feature matrix with the feature matrices of each of the reference signals, and the decision is made based on the results of calculating the difference between the values of the feature matrix of the recognisable signal and the reference signals, which is averaged in accordance with the bit depth of the feature matrix, while the recognisable signal is considered incidental to the reference signal, the difference module of the feature matrix with which will be minimal.

EFFECT: improved efficiency of object recognition when they are not fully displayed within the processed frame.

1 cl, 11 dwg

Description

The invention relates to methods for recognizing signals and can be used in technical means for recognizing objects depicted in frames of photo and video images.

A "Method for Recognizing Radio Signals" is known (RU Patent No. 2356064, IPC G01S 7/00, Bulletin No. 14, published: 20.05.2009). In this known method, reference radio signals (RS) are pre-defined. Then, for each reference RS, its energy distribution matrix (EDM) is formed. For this purpose, the reference RS are discretized, quantized, and then a frame wavelet transform (FWT) is performed by filtering their quantized samples through filters whose passbands are each doubled with an increase in the ordinal number of the filter. After this, the wavelet coefficients (WC) obtained from the output of each filter are normalized, ranked, and insignificant WCs are excluded. A set of VCs, starting from the smallest, whose total energy is 10-30% of the total energy of the entire set of VCs at the output of each filter, respectively, is selected as insignificant. Then, a matrix of REs is formed from the remaining VCs, where the rows of the RE matrix of each reference PC are the VCs obtained at the filter output. And from the RE matrices of the reference PCs, their feature vectors are formed by row-by-row concatenation of all VCs of the formed RE matrices. After this, a recognizable PC is accepted, and from its quantized samples, a RE matrix and a feature vector are formed in the same way as for the reference PCs. The accepted PC is identified by subtracting its feature vector modulo from the feature vectors of each of the reference PCs. A recognizable PC is considered incident to a reference PC whose feature vector difference with it is minimal.

A disadvantage of the known method is that it is designed to work with signals with a small number of samples, since the signal feature vectors are formed by row-by-row concatenation of the values of their RE matrices, and therefore have a very large number of elements, i.e., a high dimensionality. Reducing the dimensionality of the feature vectors formed in accordance with the known method by averaging them leads to a decrease in recognition reliability.

The “Method for recognizing radio signals” is known (Patent of the Russian Federation No. 2430417, IPC G06K 9/00, Bulletin No. 27, published: 09/27/2011).

In the known method, reference FSs are predefined and a RE matrix is formed for each reference FS. This is done by discretizing and quantizing it, and then performing a FVP operation on the sequence of its quantized reports using filters whose passbands are multiples of two to the power of K, where K is an integer. A feature vector is then formed from the reference FS's VC. Moreover, for each FS's time sample, the maximum VC is selected from among the corresponding VCs at the filter outputs, to which the remaining VCs corresponding to this FS's time sample are normalized. The average VC power values obtained at the output of each filter are selected as the feature vector elements. The recognized FS is then received, and its feature vector is formed similarly to that for the reference FS. The received FS is then identified by successively subtracting the absolute values of its feature vector from the elements of the feature vectors of each reference FS. The recognized FS is considered incident to the reference FS if its feature vector difference with the reference FS is minimal.

The disadvantage of the known method is that the average values of the VC power obtained at the output of each filter are selected as elements of the feature vectors, which leads to a decrease in the reliability of recognition.

A "Method for Recognizing Radio Signals" is known (RU Patent No. 2423735, IPC G06K 9/00, Bulletin No. 19, published: 10.07.2011). In this known method, reference signals are pre-defined, and a radio frequency matrix is formed for each reference signal. This matrix is discretized, quantized, and a frequency-domain transformation (FWD) operation is performed on the sequence of its quantized reports using filters whose passbands are multiples of two to the power of K, where K is an integer. After this, the FWD VC is normalized relative to its maximum value. Then, the VCs of the reference signal, obtained in each frequency band, are ranked, and a feature vector of the reference signal is formed. Moreover, the average values of the FWD power obtained at the output of each filter are selected as the elements of the feature vectors. Afterwards, the recognized signal is received and its feature vector is formed in the same way as for the reference signal. The received PC is identified by successively subtracting the absolute values of its feature vector elements from the feature vectors of each reference PC. The recognized PC is considered incident to the reference PC whose feature vector difference is minimal.

The disadvantage of the known method is that the average values of the VC power obtained at the output of each filter are selected as elements of the feature vectors, which leads to a decrease in the reliability of recognition.

A "Method for Recognizing Radio Signals" is known (RU Patent No. 2533651, IPC G06K 9/00, published: 20.11.2014, Bulletin No. 32). In the known method, reference RS of equal duration are pre-defined, which are discretized and quantized, then the FVP operation is performed for the sequence of quantized reports of each reference RS and a feature vector is formed for each of them from the VC of the obtained RE matrices, then the recognized RS is received and its feature vector is formed in the same way as for the reference RS, after which the recognized RS is identified by subtracting its feature vector by modulus from the feature vector of each of the reference RS, the recognized RS is considered incident to the reference RS, the difference in feature vectors with which is minimal. In this case, the feature vectors are constructed by row-by-row concatenation of the VC of the obtained RE matrices, and the VC in the feature vectors are normalized and ranked, and the recognized RS is preliminarily divided into fragments, the duration of each of which coincides with the duration of the reference RS, and a fragment of the average values of the recognized RS is calculated, from which a feature vector for the recognized RS is then formed.

A disadvantage of this method is that it is designed to work with signals with a small number of samples, since the signal feature vectors are formed by row-by-row concatenation of the values of their RE matrices, and therefore have a very large number of elements, i.e., a high dimensionality. Reducing the dimensionality of feature vectors formed in accordance with the known method by averaging them leads to a decrease in recognition reliability.

The closest analogue in technical essence to the claimed one is the “Signal Recognition Method” (RU Patent No. 2809021, IPC G06F 18/00, published: 06.12.2023, Bulletin No. 34)

In the prototype method, reference signals of equal duration are pre-set, a feature vector is formed for each of them, then a recognizable signal is received and its feature vector is formed in the same way as for the reference signal, after which the recognizable signal is identified by comparing its feature vector with the feature vectors of each of the reference signals, and a decision is made based on the results of calculating the difference in the values of the parameters of the recognizable signal and the reference signals, the recognizable signal is considered incident to the reference signal when the modulus of the difference in the parameters of the feature vectors with which is minimal, wherein a signal consisting of elements sequentially read from a video image frame is used as the reference and recognizable signals, and the feature vector is formed from the values of the histograms of the distribution of the frequency of brightness manifestation by gradations of levels, in accordance with the bit depth, which are considered the parameters of the feature vectors, after which the calculated difference in the values of the parameters of the recognizable signal and the reference signals is averaged in accordance with the bit depth of the feature vector.

The disadvantage of the prototype method is that it does not provide effective recognition of objects when they are not fully displayed within the processed frame.

The objective of the invention is to develop a method that would allow, through secondary processing of feature vectors obtained from the values of histograms of the distribution of the frequency of brightness manifestation by gradations of levels, in accordance with the bit depth, to increase the contrast of the result of identification procedures for a recognizable signal.

The technical result is an increase in the efficiency of recognizing objects when they are not fully displayed within the processed frame.

The technical result is achieved in that in the method of recognizing signals, which consists in pre-setting reference signals of equal duration, consisting of elements sequentially read from a video image frame, forming for each of them a feature vector, from the values of the histograms of the distribution of the frequency of manifestation of brightness by gradations of levels, in accordance with the bit depth, then receiving a recognizable signal, consisting of elements sequentially read from a video image frame, and forming its feature vector in the same way as for the reference signal, after which identifying the recognizable signal by comparing its feature vector with the feature vectors of each of the reference signals, and making a decision based on the results of calculating the difference between the feature vectors of the recognizable signal and the reference signals, which is averaged in accordance with the bit depth of the feature vector, the recognizable signal is considered incident to the reference signal, the modulus of the difference of the feature vectors with which will be minimal, in addition, the feature vector of both the reference signals and the recognizable signal is fed to a filter system, wherein the number of filters in the filter system and the values of their bandwidths are determined in accordance with the principle of multiple-scale signal transformation, according to which the bandwidth of each subsequent filter is twice the bandwidth of the previous filter, as a result of which a matrix of signal features is obtained at the output of the filter system, and the recognizable signal is identified by comparing its matrix of features with the matrices of features of each of the reference signals, and the decision is made based on the results of calculating the difference in the values of the matrix of features of the recognizable signal and the reference signals, which is averaged in accordance with the bit depth of the matrix of features, while the recognizable signal is considered incident to the reference signal, the modulus of the difference in the matrix of features with which will be minimal.

The claimed method is explained by drawings showing:

Fig. 1 - a frame with the first recognizable object, used to form the first reference signal;

Fig. 2 - a frame with a second recognizable object used to form a second reference signal;

Fig. 3 - the generated feature vector for the first reference signal;

Fig. 4 - the generated feature vector for the second reference signal;

Fig. 5 - frame with a recognizable object for forming a received signal;

Fig. 6 - formed feature vector of the signal received for recognition;

Fig. 7 - matrix of discrete samples of the image frame of the first reference signal;

Fig. 8 - matrix of discrete samples of the image frame of the second reference signal;

Fig. 9 - matrix of discrete samples of the image frame received for signal recognition;

Fig. 10. Modules of the difference between the values of the matrix of features of the recognized signal and the first reference signal;

Fig. 11 Modules of the difference between the values of the matrix of features of the recognized signal and the second reference signal.

All frames are presented in the same format, and all feature vectors and matrices are presented in the same scale.

The implementation of the claimed method is explained as follows.

P.1. Set reference signals of equal duration. These signals consist of elements sequentially read from a video frame containing the objects to be recognized.

As an example, Fig. 1 and Fig. 2 show video image frames that serve as reference for recognizing objects depicted in them. Frame dimensions are determined by image standards, the resolution of which is represented by the number of pixels. For the presented examples, frames measuring 1280 × 720 pixels in grayscale mode were used. In this case, each pixel of the image will contain a brightness value in the range from 0 (black) to 255 (white), the encoding of which requires 8 bits. Consequently, signals consisting of elements sequentially read from the video image frame will represent an 8-bit data stream, the length of which is determined by the dimensions of the original frame. In the presented case, one frame will contain 921,600 pixels, for a frame volume of 7,372,800 bytes.

P.2. A feature vector is formed for each reference signal. The feature vector is formed from the histogram values of the luminance frequency distribution across gradations of levels, in accordance with the bit depth, which are considered the parameters of the feature vector.

As an example, Fig. 3 and Fig. 4 show the feature vectors of the reference signals. Fig. 3 shows the feature vector corresponding to the object depicted in Fig. 1, and Fig. 4 shows the feature vector corresponding to the object depicted in Fig. 2.

For an 8-bit stream, the feature vector dimension will be 256 values. The feature vector can be formed as follows. The bitstream is fed to a comparator, the number of whose inputs is determined by the number of bits per pixel (8 bits for this example). Each comparator output, corresponding to the number of possible combinations (256 for this example), has its own counter, which counts the number of combinations per brightness gradation occurring in the frame. The counter values represent the parameters of the feature vector.

As a result of carrying out the specified technical operations, the feature vector will represent a histogram of the distribution of the frequency of brightness manifestation by gradations of levels, in accordance with the bit depth.

P.3 Receive the signal to be recognized and form its feature vector in the same way as for the reference signal.

As an example, Fig. 5 shows a video frame, the generated signal from which is fed for recognition. Fig. 6 shows a feature vector corresponding to the object depicted in Fig. 5.

P.4 The feature vector of both the reference signals and the recognizable signal is fed to a filter system, whereby the number of filters in the filter system and the values of their passbands are determined in accordance with the principle of multiple-scale signal transformation, according to which the passband of each subsequent filter is twice the passband of the previous filter.

Thus, for a feature vector consisting of 256 elements, the filter system will include 8 digital filters with bandwidths equal to 1, 2, 4, 8, 16, 32, 64, 128 reports.

The implementation of digital filters is known, see the description of the invention to the USSR Author's Certificate No. 758166 "Digital Filter", application publication date: 08/23/1980 Bulletin No. 24. Published: 08/25/1980.

As an example, Fig. 7 shows the matrix of discrete samples of the image frame, which is the matrix of features of the signal shown in Fig. 1, obtained at the output of the filter system. Fig. 8 shows the matrix of discrete samples of the image frame, which is the matrix of features of the signal shown in Fig. 2, obtained at the output of the filter system. And Fig. 9 shows the matrix of discrete samples of the image frame, which is the matrix of features of the signal shown in Fig. 5, obtained at the output of the filter system.

P.5 Identify the recognizable signal by comparing its feature matrix with the feature matrices of each of the reference signals.

As an example, Fig. 10 shows the absolute values of the difference between the feature matrix V02 _k,m of the recognizable signal, see Fig. 5, and the reference signal, see Fig. 1. And Fig. 11 shows the absolute values of the difference between the feature matrix V12 _k,m of the recognizable signal, see Fig. 5, and the reference signal, see Fig. 2.

P.6 The calculated difference between the values of the feature matrix of the recognized signal and the reference signals is averaged in accordance with the bit depth of the feature matrix, while the recognized signal is considered incident to the reference signal, the modulus of the difference between the feature matrix with which will be minimal

For the presented example, the average value of the modulus of the difference between the feature matrix of the recognized signal and the first reference signal, see Fig. 1, is R02ср = 2.102. And the difference with the feature vector of the second signal, see Fig. 2, was R12ср = 2.819.

In accordance with the claimed method, the video image frame of the recognized signal will be assigned to the first reference frame of the video image.

Moreover, the implementation of recognition procedures in accordance with the prototype method resulted in the video frame of the recognized signal being assigned to the second reference video frame. The calculated difference values, respectively, were R02cp = 4.898 and R12cp = 4.891.

To validate the effectiveness of the proposed method, 200 experiments were conducted, during which video frames were recognized, each capturing various parts of an object (car) ranging from 100% to 2% of its total area. It was found that the proposed method consistently performs recognition if the video frame captures at least 3-5% of the object, while maintaining constant illumination of the area shown in the video frame (Fig. 2).

Thus, thanks to a new set of essential features in the claimed method, the claimed technical result is achieved: an increase in the efficiency of recognizing objects when they are not fully displayed within the processed frame.

Claims (1)

A method for recognizing signals which consists in pre-setting reference signals of equal duration, consisting of elements sequentially read from a video image frame, forming for each of them a feature vector from the values of the histograms of the distribution of the frequency of occurrence of brightness by gradations of levels, in accordance with the bit depth, then receiving a recognizable signal consisting of elements sequentially read from the video image frame, and forming its feature vector in the same way as for the reference signal, after which identifying the recognizable signal by comparing its feature vector with the feature vectors of each of the reference signals, and making a decision based on the results of calculating the difference between the feature vectors of the recognizable signal and the reference signals, which is averaged in accordance with the bit depth of the feature vector, the recognizable signal is considered incident to the reference signal, the modulus of the difference in the feature vectors with which will be minimal, characterized in that the feature vector of both the reference signals and the recognizable signal is fed to a filter system, wherein the number of filters in the filter system and the values of their passbands are determined in in accordance with the principle of multi-scale signal transformation, according to which the bandwidth of each subsequent filter is twice the bandwidth of the previous filter, as a result, at the output of the filter system, a matrix of signal features is obtained, and the recognizable signal is identified by comparing its matrix of features with the matrices of features of each of the reference signals, and the decision is made based on the results of calculating the difference in the values of the matrix of features of the recognizable signal and the reference signals, which is averaged in accordance with the bit depth of the matrix of features, while the recognizable signal is considered incident to the reference signal, the modulus of the difference in the matrix of features with which will be minimal.