WO2022236874A1

WO2022236874A1 - Banknote quality test method and system based on multi-spectral image, and medium

Info

Publication number: WO2022236874A1
Application number: PCT/CN2021/095958
Authority: WO
Inventors: 康宏伟; 魏东
Original assignee: 广州广电运通金融电子股份有限公司
Priority date: 2021-05-14
Filing date: 2021-05-26
Publication date: 2022-11-17
Also published as: CN113240643A

Abstract

A banknote quality test method and system based on a multi-spectral image, and a storage medium. The banknote quality test method comprises: collecting a multi-spectral image of a banknote to be tested, and pre-processing the image to obtain a plurality of sub-regions; extracting a feature parameter of each sub-region, performing feature fusion on feature parameters of the sub-regions, and then importing same into a trained regression model to obtain a degree of confidence; and determining the quality of said banknote by means of combining the degree of confidence and a regional voting policy. Feature training is performed in advance on banknotes without abnormal features by using a regression model, and when a banknote to be tested needs to be tested, a feature vector of said banknote is imported into the trained regression model, such that graffiti and smudginess detection in any complex background can be completed, thereby achieving the aim of quickly testing the quality of a banknote.

Description

A banknote quality detection method, system and medium based on multispectral images

technical field

The invention relates to the field of banknote quality detection, in particular to a banknote quality detection method, system and storage medium based on multispectral images.

Background technique

At present, paper currency is the current mainstream form of currency circulation, and when paper currency circulates in the market, it is easy to get stains, or someone intentionally or unintentionally scribbles on the paper currency, resulting in frequent graffiti on the surface of the paper currency , dirty, etc., and these paper currencies with graffiti and dirt are not suitable for circulation in the market; in addition, different countries have their own paper currencies, and the surface textures of paper currencies in different countries are not the same. It is difficult to accurately detect the quality of paper money with a complex background; in this situation, a banknote quality evaluation technology is urgently needed to realize the quality rating of banknotes.

Contents of the invention

In order to overcome the deficiencies of the prior art, one of the purposes of the present invention is to provide a banknote quality detection method based on multispectral images, which can perform quality detection on banknotes and improve detection accuracy.

The second object of the present invention is to provide a detection system for implementing the method for detecting the quality of banknotes based on multi-spectral images.

The third object of the present invention is to provide a storage medium for implementing the method for detecting the quality of banknotes based on multi-spectral images.

One of purpose of the present invention adopts following technical scheme to realize:

A banknote quality detection method based on multispectral images, comprising:

Step S1: Collect the multispectral image of the banknote to be tested, and preprocess the image to obtain multiple sub-regions;

Step S2: Extract the characteristic parameters of each sub-region, perform feature fusion on the characteristic parameters of the sub-regions, and import them into the trained regression model to obtain confidence;

Step S3: Combining the confidence level and the regional voting strategy to judge the quality of the banknote to be tested.

Further, the multispectral image is a red, green, and blue three-band image.

Further, the method of the pretreatment is:

Divide the banknote multispectral image into N*N sub-regions on average, wherein N is a non-zero natural number;

The corresponding attention weight is assigned to each sub-region according to the image texture features.

Further, the feature parameters include Hog feature, Gabor filter local energy, texture statistical histogram feature and image wavelet transform coefficient variance.

Further, the method of feature fusion is:

Splicing the feature parameters of the sub-regions into a 26-dimensional primary feature vector;

Multiplying the initial feature vector by the attention weight corresponding to the area to obtain a final area feature vector.

Further, the training method of the regression model is:

A support vector regression model is trained by using the final region feature vectors of multiple banknote images without abnormal features to obtain a trained regression model.

Further, the regional voting strategy is:

Among them, set the sub-region with reliability level greater than or equal to 2 as RS, N _f is the number of RS; N _c is the number of block connected domains of RS in the texture image, and the area of each connected domain is larger than the area of a single sub-region.

Further, the confidence level is divided into three levels:

Among them, the three grades from low to high represent genuine banknotes, suspected abnormal banknotes and abnormal banknotes.

Two of the purpose of the present invention adopts following technical scheme to realize:

A banknote quality inspection system based on multispectral images, including:

The acquisition module is used to acquire the multispectral image of the banknote to be tested;

A preprocessing module for dividing the multispectral image into multiple subregions;

The feature extraction module is used to extract the feature parameters of each sub-region, and perform feature fusion on the feature parameters of the sub-regions to obtain the final region feature vector;

The model analysis module is used to import the final regional feature vector after feature fusion into the trained regression model to obtain confidence;

The post-processing module is used for judging the quality of banknotes to be tested in combination with the confidence level and the regional voting strategy.

Three of the purpose of the present invention adopts following technical scheme to realize:

A storage medium stores a program, and when the program is executed by a processor, the above method for detecting banknote quality based on multi-spectral images is realized.

Compared with the prior art, the beneficial effects of the present invention are:

The present invention uses the regression model to pre-train banknotes that do not contain abnormal features. When the banknotes to be tested need to be detected, the feature vectors of the banknotes to be tested are imported into the trained regression model, and the graffiti in any complex background can be completed. , Dirt detection, to achieve the purpose of quickly detecting the quality of banknotes.

Description of drawings

Fig. 1 is the schematic diagram of the banknote quality detection method based on the multispectral image of the present invention;

Fig. 2 is the specific flowchart of the banknote quality detection method based on the multispectral image of the present invention;

Fig. 3 is a structural block diagram of the multi-spectral image-based banknote quality detection system of the present invention.

Detailed ways

Below, the present invention will be further described in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of not conflicting, the various embodiments described below or the technical features can be combined arbitrarily to form new embodiments. .

Embodiment one

This embodiment discloses a banknote quality detection method based on multispectral images, which can detect graffiti and dirt on the banknote surface under any complex background, and improve the accuracy of banknote quality evaluation.

As shown in Figure 1 and Figure 2, the banknote quality detection method of the present embodiment comprises the following steps:

Step S3: Perform regional voting strategy processing in combination with the confidence level to judge the quality of the banknote to be tested.

In this embodiment, before the banknotes to be tested are subjected to quality inspection, a regression model needs to be constructed in advance, and the training method of the regression model is:

Step a: Collect multi-spectral images of banknote images without abnormal features. Specifically, the banknote images without abnormal features refer to the original banknote images without dirt or graffiti. The image information of different bands of the same target is obtained at different times, and in this embodiment, the multi-spectral image specifically refers to the image information of red, green and blue bands. The acquisition method of the multispectral image is an existing technology, and will not be described in detail here.

Step b: Divide the multispectral image into N*N sub-regions on average, and in the N*N sub-regions, assign different attention weights α _j according to the texture depth and density in the original banknote image, where j=1 ,2,...,N*N. In this embodiment, allocating different attention weights can be used to emphasize or select important information of the target processing object, suppress irrelevant detailed information, reduce interference of subsequent feature extraction, and improve feature extraction accuracy.

Step c: Extract feature parameters for each sub-region, where feature parameters include but not limited to Hog features, Gabor filter local energy, texture statistical histogram features, and image wavelet transform coefficient variance.

Among them, the Hog feature is the histogram of the gradient distribution of the direction, and the feature is formed by calculating and counting the histogram of the gradient direction of the local area of the image; in this method, 8 gradient directions are used, and the calculated gradient distribution histogram is normalized After this normalization, better effects on illumination changes and shadows can be obtained, and finally an 8-dimensional Hog feature vector is obtained.

And the Gabor filter bank in 4 directions

2 wavelengths to choose from. The local energy here is represented by the global gray level of the Gabor filtered sub-region. Due to the complexity of the texture, the direction and wavelength can be set according to the specific texture, and finally an 8-dimensional feature vector is obtained.

Among them, the texture statistical histogram feature can be represented by the following six metrics:

(1) The variance of the histogram, as shown in the following formula:

(2) Probability sum of squares, as shown in the following formula:

(3) Average entropy, as shown in the following formula:

(4) Relative smoothness, as shown in the following formula:

(5) Skewness, as shown in the following formula:

(6) Kurtosis, as shown in the following formula:

Among them, μ is the average value of the gray level in the region, Z _i refers to the value corresponding to the i-th gray level, L means that the sub-region has L gray levels, and p(Z _i ) means the value of the i-th gray level Frequency, σ represents the gray standard deviation in the region, and the finally generated texture statistical histogram feature is (s ₁ , s ₂ , s ₃ , s ₄ , s ₅ , s ₆ ), which is a 6-dimensional feature vector.

The variance of the image wavelet transform coefficients in the feature parameters is a four-dimensional vector, representing the variance of the coefficients calculated from the low-pass decomposition filter. These four coefficients are approximate, horizontal, vertical, and diagonal coefficients.

Step d: Concatenate the above feature parameters into a 26-dimensional initial feature vector V _i , and then multiply the initial feature vector by the regional attention weight α _i to obtain the final regional feature vector V _f . Different weights are added to the final regional feature vector _Vf , which improves the detection ability of graffiti and dirty banknotes in weak texture areas, reduces the probability of mistaking genuine banknotes for graffiti and dirty banknotes in strong texture areas, and further improves the accuracy of banknote quality detection. sex.

Step e: use the final region feature vector V _f to train the support vector regression model M ₁ to complete the training step of the regression model.

The specific method of training the support vector regression model is as follows:

(1) Set the training sample set:

Set the final region feature vector V _f as the training sample set T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,(x _k ,y _k ),...,(x _N ,y _N )}, x _k ∈ R ⁿ , R ⁿ is the input space, and n represents the dimension of the input space. The value of the output space y _k ∈ R, R is the output space. (x _k , y _k ) means the kth sample point, N means there are N sample points in the training set, and set the insensitive loss factor ε and penalty factor C.

(2) Calculate the kernel function matrix K and initialize it:

Calculate the kernel function matrix K, k _pq =K(x _p ,x _q )=<φ(x _p ),φ(x _q )>, where k _pq represents the kernel of the pth sample point and the qth sample point Function product, φ(x) is kernel function; initialize Lagrangian multiplier constant λ=(λ ₁ ,λ ₂ ,...,λ _k ,...,λ _N )=0, decision function bias b is equal to 0. Among them, λ _k represents the Lagrangian multiplier corresponding to the kth sample x _k in the decision function, and b represents the bias of the decision function.

(3) Use the SMO algorithm to optimize the objective function until all samples in the training set satisfy the KKT condition. The objective function to be optimized is as follows:

Among them, _λi represents the Lagrangian multiplier corresponding to the _i -th sample point, and kij represents the kernel function product of the i-th sample point and the j-th sample point. ε is the insensitivity loss factor.

(4) Calculate the final regression decision function:

At the end of the training, the final Lagrangian vector λ=(λ ₁ ,λ ₂ ,...,λ _k ,...,λ _N ) is calculated, the final decision bias is b, and the calculated regression decision function :

Among them, λ _i represents the i-th Lagrangian multiplier in the final Lagrangian vector, x _i is the i-th sample in the training sample set, x is a new sample, and K( _xi , x) is the training sample set The kernel function product of the i-th sample and the new sample x, y _x represents the output of the regression decision function corresponding to the new sample x.

After the regression model training is completed, the multi-spectral image of the banknote to be tested can be collected, and the image of the banknote to be tested is partitioned according to the above steps a to step d to obtain the final regional feature vector V _f of the banknote to be tested for each sub-region. The final region feature vector V _f is input as a test sample set into the already trained regression model for calculation, and the output value _of the regression model M1 is used as the confidence degree Conf that the sub-region contains abnormal features; in this embodiment, Conf is divided into 3 A grade (Grade), as shown in the following formula:

The three levels from low to high represent genuine banknotes, suspected abnormal banknotes, and abnormal banknotes.

In this embodiment, considering that the abnormal features of banknotes often appear concentrated in pieces, it is necessary to perform regional voting strategy processing during the final two classifications, and finally judge the quality of banknotes based on the results of regional voting strategy processing; this embodiment is performing regional voting During policy processing, it is also necessary to consider the relative positions of each region and their respective confidence levels. The regional voting strategy is as follows:

Wherein, set the sub-area whose reliability level is greater than or equal to 2 as RS, and N _f is the number of RS. Grade _i represents the confidence level of the _i -th sub-region in Nf. According to whether RS is connected in space, the concept of block connected domain is introduced. N _c is the number of block-connected domains of RS in the texture image, and the minimum is 1. The area of each connected domain must be larger than the area of a single subregion. When Grade _i is constant, the larger _Nc means that the distribution of abnormal banknotes is relatively discrete, and the score of regional voting will be low; when _Nc is constant and Grade _i is smaller, it means that banknotes are abnormally distributed The degree is relatively low, and the score of regional voting will also be low at this time.

When Y<3, there is no abnormality in the banknote;

When 3<=Y<=5, the banknote is slightly abnormal;

When 6<=Y<=9, the banknote is seriously abnormal;

When Y>10, the banknote is seriously abnormal.

In this embodiment, the regression model is used to perform feature training on the banknotes without abnormal features in advance. When the banknotes to be tested need to be detected, the feature vectors of the banknotes to be tested are imported into the trained regression model, and the detection of any complex background can be completed. Graffiti and dirt detection to achieve the purpose of quickly detecting the quality of banknotes.

Embodiment two

This embodiment provides a banknote quality detection system based on multispectral images. The system executes the banknote quality detection method based on multispectral images described in Embodiment 1. As shown in FIG. 3, the detection system of this embodiment specifically includes the following module:

The collection module is used to collect the multispectral image of the banknote to be tested, wherein the multispectral image is the image information of three bands of red, green and blue;

The preprocessing module is used to divide the multi-spectral image into multiple sub-regions. The division rule is to divide the multi-spectral image into N*N sub-regions on average. In the N*N sub-regions, according to the texture depth in the original banknote image , assigning different attention weights α _j to the degree of density;

The feature extraction module is used to extract the feature parameters of each sub-region, perform feature fusion on the feature parameters of the sub-regions to obtain the initial feature vector V _i , and then multiply the initial feature vector by the regional attention weight α _i to obtain the final regional feature vector V _f ; Wherein, characteristic parameters include but not limited to Hog feature, Gabor filter local energy, texture statistical histogram feature and image wavelet transform coefficient variance;

The model analysis module is used to import the final regional feature vector V _f into the trained regression model to obtain the degree of confidence;

Embodiment Three

This embodiment provides a banknote quality detection device, including:

program;

a memory for storing the program;

The processor is configured to load the program to execute the multi-spectral image-based banknote quality detection method described in Embodiment 1.

In addition, this embodiment also provides a storage medium, which stores a program, which is characterized in that, when the program is executed by a processor, the method for detecting the quality of banknotes based on multi-spectral images as described above is realized.

The device and storage medium in this embodiment are based on the two aspects of the same inventive concept as the method in the previous embodiment. The implementation process of the method has been described in detail above, so those skilled in the art can understand clearly from the foregoing description To understand the structure and implementation process of the device in this implementation, for the sake of brevity of the description, details will not be repeated here. The above-mentioned embodiment is only a preferred embodiment of the present invention, and cannot be used to limit the protection scope of the present invention. Any insubstantial changes and substitutions made by those skilled in the art on the basis of the present invention belong to the scope of the present invention. Scope of protection claimed.

The above-mentioned embodiment is only a preferred embodiment of the present invention, and cannot be used to limit the protection scope of the present invention. Any insubstantial changes and substitutions made by those skilled in the art on the basis of the present invention belong to the scope of the present invention. Scope of protection claimed.

Claims

A method for detecting the quality of banknotes based on multispectral images, characterized in that it includes:

Step S1: Collect the multispectral image of the banknote to be tested, and preprocess the image to obtain multiple sub-regions;

Step S2: Extract the characteristic parameters of each sub-region, perform feature fusion on the characteristic parameters of the sub-regions, and import them into the trained regression model to obtain confidence;

Step S3: Combining the confidence level and the regional voting strategy to judge the quality of the banknote to be tested.
The banknote quality detection method based on multi-spectral images according to claim 1, wherein the multi-spectral images are red, green and blue three-band images.
The method for detecting the quality of banknotes based on multispectral images according to claim 1, wherein the preprocessing method is:

Divide the banknote multispectral image into N*N sub-regions on average, wherein N is a non-zero natural number;

The corresponding attention weight is assigned to each sub-region according to the image texture features.
The banknote quality detection method based on multi-spectral images according to claim 1, wherein the feature parameters include Hog features, Gabor filter local energy, texture statistical histogram features and image wavelet transform coefficient variance.
The banknote quality detection method based on multispectral images according to claim 3, wherein the method of feature fusion is:

Splicing the feature parameters of the sub-regions into a 26-dimensional primary feature vector;

Multiplying the initial feature vector by the attention weight corresponding to the area to obtain a final area feature vector.
The banknote quality detection method based on multispectral images according to claim 1, wherein the training method of the regression model is:

Using the final region feature vectors of multiple banknote images without abnormal features to train the support vector regression model to obtain a trained regression model.
The banknote quality detection method based on multispectral images according to claim 1, wherein the regional voting strategy is:

Among them, set the sub-region with reliability level greater than or equal to 2 as RS, N f is the number of RS; N c is the number of block connected domains of RS in the texture image, and the area of each connected domain is larger than the area of a single sub-region.
The banknote quality detection method based on multispectral images according to claim 1, wherein the confidence is divided into three levels:

Among them, the three grades from low to high represent genuine banknotes, suspected abnormal banknotes and abnormal banknotes.
A banknote quality detection system based on multispectral images, characterized in that it includes:

The acquisition module is used to acquire the multispectral image of the banknote to be tested;

A preprocessing module for dividing the multispectral image into multiple subregions;

The feature extraction module is used to extract the feature parameters of each sub-region, and perform feature fusion on the feature parameters of the sub-regions to obtain the final region feature vector;

The model analysis module is used to import the final regional feature vector after feature fusion into the trained regression model to obtain confidence;

The post-processing module is used for judging the quality of banknotes to be tested in combination with the confidence level and the regional voting strategy.
A storage medium storing a program, characterized in that, when the program is executed by a processor, the multi-spectral image-based banknote quality detection method according to any one of claims 1-8 is realized.