Magnetic signal recognition method

Abstract

This invention provides a magnetic signal recognition method, which includes: extracting sample magnetic signals from sample banknotes; determining a magnetic signal threshold based on the sample magnetic signals; if the real-time magnetic signals all conform to the numerical range of the magnetic signal threshold, then a first result is a valid magnetic signal; inputting the sample magnetic signals of the sample banknotes into a network model for iterative training to obtain a training parameter array; determining a final network model based on the training parameter array; inputting the magnetic signals of the banknote to be tested into the final network model to obtain a recognition result; if the recognition result is greater than a preset ideal threshold y... sd (t o If both the first and second results are valid magnetic signals, then the final result of the magnetic signal of the banknote being tested is a valid magnetic signal. This technology, through a dual-layer recognition method, can more accurately identify the validity of the magnetic signal of the banknote.

Description

Technical Field

[0001] This invention relates to the field of signal recognition technology, and in particular to a method for magnetic signal recognition. Background Technology

[0002] Banknotes are an important symbol of value, and how to effectively identify the magnetic signals of banknote images has become a pressing technical problem.

[0003] In banknote magnetic signal recognition devices, various issues such as banknote friction and movement can lead to magnetic signal sampling tilt. Furthermore, vibrations and magnetic signal interference during sampling can cause harmonics and noise interference to be added to the fundamental frequency, making signal recognition difficult. Therefore, effective processing of the magnetic signal before magnetic feature analysis is crucial for subsequent banknote recognition. This application proposes methods such as correcting tilted magnetic signals and removing interference to accurately restore the standard waveform of the magnetic signal.

[0004] Furthermore, the magnetic signal features on banknotes are divided into two types: one is specific magnetic coding features, including magnetic code amplitude, magnetic code width, and magnetic code spacing; the other is the surface magnetic features of the banknote image, where the scanned magnetic signal does not have specific signal characteristics. The magnetic signal of a US dollar banknote consists entirely of surface magnetic features in the image. This application primarily focuses on identifying the validity of the magnetic signal on US dollars.

[0005] For magnetic signal processing of images like the US dollar, which exhibit magnetic characteristics, existing signal processing and logical judgment methods suffer from weak anti-interference capabilities, low accuracy, and poor performance. This application proposes a method that combines existing signal processing methods with an improved BP (Back Propagation) neural network for signal detection. This application makes a holistic judgment on the magnetic signal intensity and distribution, and then uses the improved BP neural network to accurately identify the magnetic signal features. Summary of the Invention

[0006] The purpose of this invention is to provide a magnetic signal recognition method for accurately identifying the magnetic signals of US dollars.

[0007] A first aspect is a magnetic signal recognition method, the method comprising: step S102, extracting sample magnetic signals from a sample banknote, and determining magnetic signal thresholds for each feature region of the sample banknote based on the sample magnetic signals; step S104, if the real-time magnetic signals of each feature region of the banknote to be tested all conform to the numerical range of the magnetic signal thresholds, then the first result ResultSignal of the magnetic signal of the banknote to be tested is a valid magnetic signal; step S106, inputting the sample magnetic signals Data-Result[u,v] of the sample banknote into an improved error backpropagation neural network model for iterative training to obtain a training parameter array; wherein, the training parameter array includes: harmonic weight ω, harmonic frequency k, and harmonic phase. Step S108: Based on the training parameter array, determine the final network model, input the magnetic signal of the banknote to be tested into the final network model, and obtain the recognition result. If the recognition result is greater than the preset ideal threshold y... sd (t o If the first result and the second result are both valid magnetic signals, then the final result of the magnetic signal of the banknote under test is a valid magnetic signal. Step S110: If both the first result and the second result are valid magnetic signals, then the final result of the magnetic signal of the banknote under test is a valid magnetic signal.

[0008] Further, step S102 includes: S102-1: collecting the initial magnetic signals Data[u,v] of N sample banknotes; where N = 2000-5000; S102-2: preprocessing the initial magnetic signals of the sample banknotes to obtain sample magnetic signals Data-Result[u,v]; S102-3: dividing Data-Result[u,v] into P magnetic feature regions according to a preset division rule; S102-4: determining the magnetic signal thresholds of the P feature regions according to statistical principles and preset threshold determination rules, wherein the magnetic signal thresholds include: magnetic signal intensity threshold, magnetic signal period threshold, and magnetic signal slope threshold.

[0009] Further, step S104 includes: S104-1: Acquiring the initial magnetic signal of the banknote to be tested;

[0010] S104-2: Preprocess the initial magnetic signal of the banknote to be tested to obtain the magnetic signal Real-Data-Result[u,v]; S104-3: Based on Real-Data-Result[u,v], extract the real-time magnetic signal DataFeature of P regions of the banknote to be tested, wherein the real-time magnetic signal includes real-time magnetic signal intensity, real-time magnetic signal period, and real-time magnetic signal slope; S104-4: Compare the numerical relationships between the real-time magnetic signal intensity, real-time magnetic signal period, and real-time magnetic signal slope of the banknote to be tested and the magnetic signal intensity threshold, magnetic signal period threshold, and magnetic signal slope threshold determined in step S102-4; S104-5: If the real-time magnetic signals all meet the magnetic signal threshold range, then the first result ResultSignal of the magnetic signal of the banknote to be tested is a valid magnetic signal.

[0011] Further, step S106 includes: S106-1: Using the sample magnetic signals Data-Result[u,v] of N sample banknotes as t(o), where o = 1, 2, ..., N, this is only an assignment process; using t(o) as the input signal of the neuron's activation function, where o represents the o-th sample banknote, N is the total number of sample banknotes, and N is 2000-5000; S106-2: The activation function of the neural network's neurons is: Where h represents the h-th harmonic order, n is the highest harmonic order, 1 < h < n, and k h , These are the frequency and phase of the harmonics, respectively, and both serve as parameters of the excitation function; they are variable parameters; w is chosen as a fixed constant, w = 1; S106-3: The output of the neural network model is: in, k h ω h These represent the phase, frequency, and weight of the harmonics, respectively; S106-4: Calculate the output error of the neural network, where the output error is: e (o) =y sd (t o )-y out (t o (3), where y sd (t o S106-5: Calculate the performance indicators, which are: Where o represents the oth sample banknote; S106-6: Based on the above N performance indicators V(o), a basic inertial algorithm is introduced as the activation function and weight learning algorithm in the improved BP neural network to correct the weight, frequency and phase of the harmonics. The correction process is as shown in formulas (5)-(7): where η is the learning rate and a is the momentum factor. The appropriateness of the trade-off between the two directly affects the convergence speed; Δω l (o), Δk l (o), are the correction amounts of the weight, frequency, and phase respectively; in these formulas (5)-(7), they are carried out simultaneously without priority; S106-7: Obtain the training parameter arrays ω, k(8).

[0012] Further, step S108 includes: S108-1: Input the training parameter array into formula (2) to obtain the final network model; S108-2: Input the excitation function input signal t of the neurons of the banknote to be measured into the final neural network model to obtain the latest y out ; S108-3: If the latest y out is greater than the preset ideal result value y sd , then the second result of the banknote to be measured is true.

[0013] Further, the process of extracting the sample magnetic signal of the sample banknote in step S102 includes a preprocessing process, and the preprocessing includes: The steps of the above-mentioned rectification process include: the rectification process and the anti-interference processing process.

[0014] [[ID=2​​Further, the interference removal process includes: 1) performing a Fast Fourier Transform (FFT) on the torsional magnetic signal data Data-rotate[u,v] to obtain approximate distribution data S(x,y) of different harmonic frequencies in all channels of the magnetic signal, where x is the amplitude and y is the frequency; 2) determining a suitable wavelet function and the number of wavelet decomposition levels f based on the distribution data S(x,y) of a large number of samples, and performing f-level wavelet separation on the signal Data_rotate[u,v] to obtain f groups of low-frequency wavelet distribution arrays and high-frequency wavelet distribution arrays; wherein the wavelet function is db4 and f = 3; 3) utilizing The wavelet denoising principle is to set all coefficients of the high-frequency wavelet distribution array obtained by wavelet decomposition to zero to remove high-frequency noise; 4) the low-frequency wavelet distribution array obtained by wavelet decomposition is used with asynchronous sampling method - windowed interpolation FFT algorithm to obtain f groups of accurate low-frequency wavelet distribution arrays; 5) using the inverse FFT transform principle, the f groups of accurate low-frequency wavelet distribution arrays are subjected to inverse FFT transform to obtain f groups of magnetic signal arrays of low-frequency steady state, and the f groups of magnetic signal arrays are superimposed to obtain the sample magnetic signal Data-Result[u,v].

[0016] The beneficial effects of the embodiments of the present invention are as follows:

[0017] 1. By utilizing the straightening process, tilted banknotes can be straightened, thereby extracting more accurate magnetic signals.

[0018] 2. By utilizing the interference removal process, noise reduction is achieved in the magnetic signal of banknotes, thereby extracting more accurate magnetic signals.

[0019] 3. By dividing the magnetic signal region and setting the threshold, the validity of the magnetic signal in the first layer can be identified.

[0020] 4. Using an improved BP neural network recognition method, the validity of the magnetic signal in the second layer is recognized.

[0021] 5. Combining a two-layer process yields more accurate results. Specifically, this scheme combines existing signal processing methods with an improved BP neural network. The first layer first divides the magnetic signal feature region and extracts magnetic signal feature points such as signal intensity, signal variation, and signal width, using signal processing, logical calculation, and threshold comparison for judgment. The second layer extracts N training samples and uses the improved BP neural network for identification and judgment. If both layers determine that the magnetic signal to be tested is a valid signal, then the magnetic signal is ultimately considered a valid magnetic signal; otherwise, it is considered an invalid magnetic signal. Through two-layer recognition, the accuracy of magnetic feature signal detection in image tables is improved.

[0022] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 A flowchart of a magnetic signal recognition method provided in an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of the magnetic signal corresponding to the magnetic signal image provided in an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the tilted magnetic signal provided in an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of the distorted magnetic signal provided in an embodiment of the present invention;

[0029] Figure 5 A magnetic feature region division diagram provided for an embodiment of the present invention;

[0030] Figure 6 A schematic diagram illustrating the operation of the improved BP neural network model provided in this embodiment of the invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0032] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0033] With the widespread use of banknotes, how to accurately identify the validity of magnetic signals in banknotes has become an urgent technical problem to be solved.

[0034] Based on this, the present invention provides a magnetic signal identification method. This technology is applied to the technical scenario of validating the magnetic signals of US dollars.

[0035] Example 1

[0036] This invention provides a method for magnetic signal recognition. The table below shows the correspondence between each letter and its physical meaning:

[0037]

[0038]

[0039] like Figure 1 As shown, the magnetic signal identification method includes:

[0040] Step S102: Extract the sample magnetic signal of the sample banknote, and determine the magnetic signal threshold of each feature area of ​​the sample banknote based on the sample magnetic signal.

[0041] Specifically, the banknotes in this patent are generally US dollars, and this technology is an effective method for identifying the magnetic signals of US dollars. The magnetic signals of US dollars are represented by the magnetic characteristics shown in the full-page image. Figure 2 This is a schematic diagram of the sample magnetic signal.

[0042] Specifically, whether extracting the sample magnetic signal of a sample banknote or the test magnetic signal of the banknote to be tested, preprocessing of the banknote is required during the extraction process to obtain a more accurate magnetic signal. The preprocessing process includes a straightening process and an interference processing process. Here, the sample magnetic signal of the preprocessed sample banknote is defined as Data-Result[u,v]; and the test magnetic signal of the preprocessed test banknote is defined as Real-Data-Result[u,v].

[0043] The preprocessing procedure for sample banknotes is as follows:

[0044] I. Correction Process

[0045] This process is also known as the tilting process. The main steps are to first calculate the tilt angle of the banknote (i.e., paper money), and then to translate and correct each channel of the magnetic signal according to the tilt angle.

[0046] The steps of the above-mentioned rectification process include:

[0047] 1) Assume that the initial magnetic data matrix of the banknote magnetic signal is Data[u, v], where u is the number of channels and v is the number of points collected per channel.

[0048] 2) Extract the left boundary LB[i] and the right boundary RB[i] of the banknote magnetic signal for each channel, where (0 <= i < u), and here, i is the i-th channel.

[0049] 3) Select the left or right boundary of two channels with complete magnetic signals, and obtain the slope DSlope of the magnetic signal according to the selected boundary.

[0050] 4) According to the slope DSlope of the magnetic signal, the left boundary LB[i] and the right boundary RB[i], perform a translation transformation on each channel of the data Data[u, v] to achieve rectification, and obtain the rectified magnetic signal matrix Data-rotate[u, v].

[0051] As Figure 3 shown is the inclined magnetic signal Data[u, v], Figure 4 and Data-rotate[u, v] is the rectified magnetic signal.

[0052] II. Anti-interference processing

[0053] The processing method for signal interference is generally as follows: Use algorithms such as Fourier transform and wavelet to perform anti-interference processing on the signal.

[0054] Specifically:

[0055] 1) Perform a fast Fourier transform (Fast Fourier Transform, FFT) on the rectified magnetic signal data Data-rotate[u, v] to obtain the approximate distribution data S(x, y) of the frequencies of different harmonics in the magnetic signals of all channels, where x is the amplitude and y is the frequency.

[0056] 2) According to the distribution data S(x, y) of a large number of samples, confirm a suitable wavelet function and the decomposition level f of the wavelet. On this basis, perform f-layer wavelet separation on the signal Data_rotate[u, v] to obtain f groups each of the low-frequency wavelet distribution array and the high-frequency wavelet distribution array. In this invention, the wavelet function db4 is selected and f is equal to 3.

[0057] 3) Using the wavelet denoising principle, set all the coefficients of the obtained high-frequency wavelet distribution array of the wavelet decomposition to zero to remove high-frequency noise.

[0058] 4) The low-frequency wavelet distribution array obtained by wavelet decomposition is used to obtain f sets of accurate low-frequency wavelet distribution arrays by using the asynchronous sampling method-windowed interpolation FFT algorithm. The windowing method in this application is the Hanning window.

[0059] Specifically, this step can reduce the shortcomings of traditional FFT algorithms, such as spectral leakage and picket fence effect.

[0060] 5) Using the principle of inverse FFT, perform inverse FFT on the f groups of accurate low-frequency wavelet distribution arrays to obtain the f groups of magnetic signal arrays of the low-frequency steady-state part. After superimposing the f groups of magnetic signal arrays, the final low-frequency steady-state magnetic signal Data-Result[u,v] is obtained, which is also called the sample magnetic signal.

[0061] At this point, the preprocessing is complete. Using this preprocessing method, the sample magnetic signal of the sample banknote can be obtained, namely Data-Result[u,v].

[0062] Specifically, the same preprocessing method can be used to obtain the magnetic signal Real-Data-Result[u,v] of the banknote to be tested.

[0063] Then, based on the sample magnetic signal Data-Result[u,v] obtained from preprocessing, the magnetic signal threshold of each feature region of the sample banknote is determined.

[0064] The specific operation of "determining the magnetic signal threshold of each feature region of the sample banknote based on the sample magnetic signal" in step S102 is as follows:

[0065] S102-1: Collect the initial magnetic signal Data[u,v] of N sample banknotes.

[0066] Where N = 2000 - 5000.

[0067] S102-2: Preprocess the initial magnetic signal of the sample banknote to obtain Data-Result[u,v].

[0068] S102-3: According to the preset partitioning rules, divide Data-Result[u,v] into P magnetic feature regions.

[0069] Specifically, the characteristic regions of banknotes differ depending on their denomination and version. Taking a specific banknote as an example, it can be divided into 7 rectangular characteristic regions, i.e., P=7, such as... Figure 5 As shown.

[0070] S102-4: Based on statistical principles and preset threshold determination rules, determine the magnetic signal thresholds for P feature regions. The magnetic signal thresholds include: magnetic signal intensity threshold, magnetic signal period threshold, and magnetic signal slope threshold.

[0071] Specifically, by collecting a large number of sample banknotes from different regions, the magnetic signal threshold can be determined based on statistical principles, thus establishing the normal threshold range. That is, based on the statistical analysis of the collected batch data, the threshold ThresholdGg for different detection points in each region is determined (G represents feature regions one, two, three...P, and g represents detection points 1, 2, 3... in each region).

[0072] Step S104: If the real-time magnetic signals of each feature area of ​​the banknote to be tested all meet the magnetic signal threshold range, then the first result ResultSignal of the magnetic signal of the banknote to be tested is a valid magnetic signal.

[0073] Specifically, step S104 above includes:

[0074] S104-1: Collect the initial magnetic signal of the banknote to be tested.

[0075] S104-2: Preprocess the initial magnetic signal of the banknote to be tested (the same process as the preprocessing of the sample banknote) to obtain Real-Data-Result[u,v].

[0076] S104-3: Based on Real-Data-Result[u,v], extract the real-time magnetic signal DataFeature of P regions of the banknote to be tested. The real-time magnetic signal includes the real-time magnetic signal intensity, real-time magnetic signal period and real-time magnetic signal slope.

[0077] Specifically, in S104-3, the magnetic signal data of the real-time magnetic signal feature regions I, II, III...P are extracted and defined as DataFeature1, DataFeature2..., and data processing is performed on each magnetic signal feature region. The processing includes signal extraction and signal statistics, and outputs the real-time data of each detection point, such as real-time magnetic signal intensity, real-time magnetic signal period, and real-time magnetic signal slope.

[0078] S104-4: Compare the numerical relationships between the real-time magnetic signal intensity, real-time magnetic signal period, and real-time magnetic signal slope of the banknote under test and the magnetic signal intensity threshold, magnetic signal period threshold, and magnetic signal slope threshold determined in step S102-4.

[0079] S104-5: If the real-time magnetic signals all meet the magnetic signal threshold range, then the first result ResultSignal of the magnetic signal of the banknote under test is a valid magnetic signal, that is, ResultSignal = 1.

[0080] For S104-5, specifically, the real-time magnetic signals DataFeature1, DataFeature2, etc. of each detection point in each feature region are compared with the magnetic signal threshold ThresholdGg to obtain the detection result of each detection point. If DataFeature is within the corresponding ThresholdGg range, the detection result is 1. If the detection results of all regions and all detection points are 1, the first result of signal processing, ResultSignal = 1, is obtained, that is, the magnetic signal of the sample banknote is determined to be a valid signal for the first time.

[0081] Step S106: Input the sample magnetic signal Data-Result[u,v] of the sample banknote into the improved backpropagation (BP) neural network model for iterative training to obtain a training parameter array; the training parameter array includes: harmonic weight ω, harmonic frequency k, and harmonic phase.

[0082] Specifically, step S106 includes:

[0083] S106-1: The sample magnetic signals Data-Result[u,v] of N sample banknotes are used as t(o), where o = 1, 2, ..., N. This is only an assignment process. t(o) is used as the input signal of the activation function of the neuron, where o represents the o-th sample banknote, and N is the total number of sample banknotes, which is generally 2000-5000. t(o) is Data-Result[u,v].

[0084] S106-2: The activation function of a neuron in a neural network is: Where h represents the h-th harmonic order, n is the highest harmonic order, 1 < h < n, and k h , These are the frequency and phase of the harmonics, respectively, which together serve as parameters of the excitation function and are variable parameters; w is chosen as a fixed constant, and in this application w = 1.

[0085] Specifically, These represent the activation functions of neurons 1 through n, respectively. Figure 6 As shown.

[0086] S106-3: The output of the neural network model is: in, k h ω h These represent the phase, frequency, and weight of the harmonics, respectively.

[0087] S106-4: Calculate the output error of the neural network, where the output error is: e(o) =y sd (t o )-y out (t o (3), where y sd (t o ) represents the preset ideal result value.

[0088] S106-5: Calculate performance indicators, which are as follows: Where o represents the oth sample banknote.

[0089] S106-6: Based on the above N performance indicators V(o), a basic inertial algorithm is introduced as the activation function and weight learning algorithm in the improved BP neural network to correct the weights, frequencies and phases of the harmonics. The correction process is as shown in formulas (5)-(7):

[0090]

[0091]

[0092]

[0093] Where η is the learning rate and a is the momentum factor, the appropriateness of their selection directly affects the convergence speed. Δω l (o), Δk l (o) These are the correction amounts for weight, frequency, and phase, respectively; in these formulas 5-7, the corrections are performed simultaneously, without any order.

[0094] Taking formula (5) as an example, the iterative parameters are explained, including explanations 1)-3). The parameter iteration rules of formulas (6) and (7) are the same as those of formula (5), and will not be repeated here.

[0095] η ω a ω η and η represent the adjustable parameters of learning rate and momentum factor, respectively. The modified law of this invention is (i.e., adjusting η) ω a ω The pattern is as follows:

[0096] Note 1: When o = 1 and 2, the values ​​of η1 and η2 are both assigned to the preset value of η0 = 0.005, and the values ​​of a1 and a2 are both assigned to the preset value of a0 = 0.003.

[0097] Note 2: When o > 2, the definition is: This is a mathematical assignment process and has no physical meaning.

[0098] Description 3: If the signs of sw(o-2), sw(o-1), and sw(o) are the same, then η ω (o + 1) = pη ω (o - 1)(12); a ω (o + 1) = qa ω (o - 1)(13); where p > 1, 0 < q < 1; p and q are preset constant parameters.

[0099] Here, set η ω (o), a ω the upper threshold and the lower threshold of η ω (o), a ω When η

[0100] S106-7: Obtain the trained parameter ω according to formula (5), where ω = [ω1, ω2,..., ω n (8). [[ID=2​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

[0108] Specifically, the process of calculating the second result (i.e., steps S106-S108) is the second-layer recognition process, such as... Figure 6 The diagram shown is a schematic of the improved BP neural network model, which is a schematic of the second layer of recognition.

[0109] Figure 6 In this context, 'c' refers to the excitation function, and 'ω' is the weight of the harmonics.

[0110] Step S110: If both the first result and the second result are valid magnetic signals, then the final result of the magnetic signal of the banknote to be tested is a valid magnetic signal.

[0111] The table below compares the recognition rate and the false negative rate. Based on experimental data analysis, the improved algorithm shows an increase in the recognition rate of approximately 4% and a false negative rate of 0%. The experimental data is as follows:

[0112]

[0113] The beneficial effects of the embodiments of the present invention are as follows:

[0114] The beneficial effects of the embodiments of the present invention are as follows:

[0115] 1. By utilizing the straightening process, tilted banknotes can be straightened, thereby extracting more accurate magnetic signals.

[0116] 2. By utilizing the interference removal process, noise reduction is achieved on the magnetic signal of banknotes, thereby extracting more accurate magnetic signals.

[0117] 3. By dividing the magnetic signal region and setting the threshold, the validity of the magnetic signal in the first layer can be identified.

[0118] 4. Using an improved BP neural network recognition method, the validity of the magnetic signal in the second layer is recognized.

[0119] 5. Combining a two-layer process yields more accurate results. Specifically, this scheme combines existing signal processing methods with an improved BP neural network. The first layer first divides the magnetic signal feature region and extracts magnetic signal feature points such as signal intensity, signal variation, and signal width, using signal processing, logical calculation, and threshold comparison for judgment. The second layer extracts N training samples and uses the improved BP neural network for identification and judgment. If both layers determine that the magnetic signal to be tested is a valid signal, then the magnetic signal is ultimately considered a valid magnetic signal; otherwise, it is considered an invalid magnetic signal. Through two-layer recognition, the accuracy of magnetic feature signal detection in image tables is improved.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for magnetic signal recognition, characterized in that, The method includes: Step S102: Extract the sample magnetic signal of the sample banknote, and determine the magnetic signal threshold of each feature region of the sample banknote based on the sample magnetic signal. Step S104: If the real-time magnetic signals of each feature area of ​​the banknote to be tested all meet the numerical range of the magnetic signal threshold, then the first result ResultSignal of the magnetic signal of the banknote to be tested is a valid magnetic signal. Step S106: Input the sample magnetic signal Data-Result[u,v] of the sample banknote into the improved error backpropagation neural network model for iterative training to obtain a training parameter array; wherein, the training parameter array includes: harmonic weight ω, harmonic frequency k, and harmonic phase. Step S108: Based on the training parameter array, determine the final network model, input the magnetic signal of the banknote to be tested into the final network model, and obtain the recognition result. If the recognition result is greater than the preset ideal threshold y... sd If the second result of the magnetic signal of the banknote to be tested, ResultNetwork, is the valid magnetic signal; Step S110: If both the first result and the second result are valid magnetic signals, then the final result of the magnetic signal of the banknote to be tested is a valid magnetic signal. Step S102 includes: S102-1: Collect the initial magnetic signal Data[u,v] of N sample banknotes, where N = 2000-5000; S102-2: Preprocess the initial magnetic signal of the sample banknote to obtain the sample magnetic signal Data-Result[u,v]; S102-3: Divide Data-Result[u,v] into P magnetic feature regions according to the preset division rules; S102-4: Based on statistical principles and preset threshold determination rules, determine the magnetic signal thresholds for P feature regions, wherein the magnetic signal thresholds include: magnetic signal intensity threshold, magnetic signal period threshold, and magnetic signal slope threshold.

2. The magnetic signal recognition method according to claim 1, characterized in that, Step S104 includes: S104-1: Acquire the initial magnetic signal of the banknote to be tested; S104-2: Preprocess the initial magnetic signal of the banknote to be tested to obtain the magnetic signal Real-Data-Result[u,v] to be tested; S104-3: Based on Real-Data-Result[u,v], extract the real-time magnetic signal DataFeature of P regions of the banknote to be tested, wherein the real-time magnetic signal includes real-time magnetic signal intensity, real-time magnetic signal period and real-time magnetic signal slope; S104-4: Compare the numerical relationships between the real-time magnetic signal intensity, real-time magnetic signal period, and real-time magnetic signal slope of the banknote under test and the magnetic signal intensity threshold, magnetic signal period threshold, and magnetic signal slope threshold determined in step S102-4. S104-5: If the real-time magnetic signals all meet the magnetic signal threshold range, then the first result ResultSignal of the magnetic signal of the banknote under test is a valid magnetic signal.

3. The magnetic signal recognition method according to claim 2, characterized in that, Step S106 includes: S106-1: Take the sample magnetic signal Data-Result[u,v] of N sample banknotes as t(o), where o = 1, 2, ..., N. This is only the assignment process; take t(o) as the input signal of the activation function of the neuron, where o represents the o-th sample banknote, N is the total number of sample banknotes, and N is 2000-5000; S106-2: The activation function of a neuron in a neural network is: Where h represents the h-th harmonic order, n is the highest harmonic order, 1 < h < n, and k h , These are the frequency and phase of the harmonic, respectively, and both serve as parameters of the excitation function, which are variable parameters; w is chosen as a fixed constant, w = 1; S106-3: The output of the neural network model is: in, k h ω h These are the phase, frequency, and weight of the harmonics, respectively. S106-4: Calculate the output error of the neural network, where the output error is: e (o) =y sd -y out (t o Formula 3, where y sd The preset ideal threshold; S106-5: Calculate performance indicators, which are as follows: Where o represents the o-th sample banknote, and the number of sample banknotes is N; S106-6: Based on N performance indicators V(o), a basic inertial algorithm is introduced as the activation function and weight learning algorithm in the improved BP neural network to correct the weights, frequencies, and phases of the harmonics. The correction process is shown in Equations 5-7. Where η is the learning rate and a is the momentum factor, the appropriateness of their selection directly affects the convergence speed; Δω l (o), Δk l (o) These are the correction amounts for weight, frequency, and phase, respectively; Equations 5-7 are performed simultaneously, without any order. S106-7: Obtain the training parameter arrays ω, respectively, according to formulas 5-7. k.

4. The magnetic signal recognition method according to claim 3, characterized in that, Step S108 includes: S108-1: Input the training parameter array into Formula 2 to obtain the final network model; S108-2: Substitute the activation function input signal t of the neurons of the banknote under test into the final neural network model to obtain the latest y. out ; S108-3: If the latest y out Greater than the preset ideal threshold y sd If the second result of the banknote being tested is true, then the banknote being tested is true.

5. The magnetic signal recognition method according to claim 1, characterized in that, The process of extracting the sample magnetic signal of the sample banknote in step S102 includes a preprocessing process, and the preprocessing process includes: a rectification process and an anti-interference processing process.

6. The magnetic signal identification method according to claim 5, characterized in that, The rectification process includes: 1) Assume that the initial magnetic data matrix of the banknote magnetic signal is Data[u, v], where u is the number of channels and v is the number of points collected per channel; 2) Extract the left boundary LB[i] and the right boundary RB[i] of the banknote magnetic signal for each channel, where (0 <= i < u), and here, i is the i-th channel; 3) Select the left boundary or the right boundary of two channels with complete magnetic signals, and obtain the slope DSlope of the magnetic signal according to the selected boundary; 4) According to the slope DSlope of the magnetic signal, the left boundary LB[i] and the right boundary RB[i], perform a translation transformation on each channel of the data Data[u, v] to achieve rectification, and obtain the rectified magnetic signal matrix Data-rotate[u, v].

7. The magnetic signal recognition method according to claim 6, characterized in that, The anti-interference processing process includes: 1) Perform a fast Fourier transform (Fast Fourier Transform, FFT) on the rectified magnetic signal data Data-rotate[u, v] to obtain the approximate distribution data S(x, y) of the frequencies of different harmonics in the magnetic signals of all channels, where x is the amplitude and y is the frequency; 2) According to the distribution data S(x, y) of a large number of samples, confirm a suitable wavelet function and the decomposition level f of the wavelet. On this basis, perform f-level wavelet separation on the signal Data_rotate[u, v] to obtain f groups of low-frequency wavelet distribution arrays and f groups of high-frequency wavelet distribution arrays; where the wavelet function is db4 and f = 3; 3) Use the wavelet denoising principle to set all the coefficients of the high-frequency wavelet distribution array obtained by wavelet decomposition to zero to remove high-frequency noise; 4) For the low-frequency wavelet distribution array obtained by wavelet decomposition, use the asynchronous sampling method - windowed interpolation FFT algorithm to obtain f groups of accurate low-frequency wavelet distribution arrays; 5) Use the FFT inverse transform principle to perform an FFT inverse transform on the f groups of accurate low-frequency wavelet distribution arrays to obtain f groups of magnetic signal arrays of the low-frequency steady-state part, and superimpose the f groups of magnetic signal arrays to obtain the sample magnetic signal Data-Result[u, v].

Images