CN111311696B - A license plate authenticity detection method based on hyperspectral unmixing technology - Google Patents

Abstract

The invention discloses a license plate authenticity detection method based on a hyperspectral unmixing technology, comprising: using a hyperspectral camera to obtain a fake license plate image as an unmixed target image; establishing a linear spectral mixing model according to the unmixed target image; constructing a target endmember matrix set M, wherein the target endmember matrix set M includes a target endmember set M1 extracted from a current license plate image and an endmember set M2 storing a universal real target, and unmixing a given target endmember set M in combination with a least squares method to obtain each endmember m_iThe corresponding abundance result map FMap_i, set the threshold T, and obtain the endmember abundance map FMap of each target_iThe binarized visual result of BMap_i, for the binary visual result BMap_iPerform optimization and obtain the final result FinalMAP of license plate forgery information.

Description

A license plate authenticity detection method based on hyperspectral unmixing technology

technical field

The invention relates to the technical field of hyperspectral unmixing, in particular to a method for detecting the authenticity of a license plate based on hyperspectral unmixing technology.

Background technique

The prosperity of the economy has led to the vigorous development of the automobile industry. In order to facilitate the management of the traffic department, restrict drivers to abide by traffic laws, and quickly determine the specific vehicle and owner information in the event of traffic violations and traffic accidents, it is very important to quickly and accurately identify the license plate. However, there are a large number of forged license plates in the current society, which brings great difficulties to the correct identification of license plates. Therefore, when recognizing license plates, the problem of how to distinguish, accurately identify and locate fake license plate information from real license plates is very important.

At present, the traditional three-color channel RGB image detection technology is difficult to detect forged target information from real license plates, and the detection accuracy is low. Hyperspectral imaging technology is based on many narrow-band image data technologies. It combines imaging technology with spectral technology to detect the two-dimensional geometric space and spectral information of the target, and obtain high-resolution continuous and narrow-band image data. The characteristic of hyperspectral with a large number of continuous narrow-band spectra enables hyperspectral data processing technology to effectively identify targets in complex scenes and distinguish real and fake targets.

At this stage, the use of hyperspectral image processing technology to detect the authenticity of license plates is usually supervised, that is, detection is performed in the case of known forged target spectra. Therefore, when the forged target spectrum is unknown or the target spectral library information is incomplete, it will lead to poor detection performance. In addition, the influence of complex background and lighting will also have a certain impact on the authenticity detection results of the license plate.

Contents of the invention

According to the problems existing in the prior art, the present invention discloses a license plate authenticity detection method based on hyperspectral unmixing technology. The method solves the problem of license plate authenticity discrimination in complex situations, specifically as follows:

Step S1: Use the hyperspectral camera to collect data, and obtain the forged license plate image as the unmixed target image. Compared with the real license plate, the counterfeit license plate has two parts, the blue bottom plate and the white number, which are different from the material of the real license plate;

Step S2: Establishing a linear spectral mixture model according to the above images;

Step S3: Construct the target endmember matrix set M={m ₁ ,m ₂ ,m ₃ ,m ₄ ,m ₅ ,m ₆ }, M is composed of the target endmember set M1={m ₁ ,m ₂ ,m ₃ ,m ₄ } extracted from the current license plate image and the endmember set M2={m ₅ ,m ₆ } for storing universal real targets. The target endmember set M1 is composed of four endmembers m ₁ , m ₂ , m ₃ , and m ₄ extracted by the automatic target retrieval algorithm ATGP, and the target endmember set M2 is composed of the average spectrum m ₅ of the blue bottom plate of the real license plate and the average spectrum m ₆ of the white number;

Step S4: Combine the least square method LS to unmix the given target endmember set M, and give the abundance result map FMap _i corresponding to each endmember m _i , where 1≤i≤6;

Step S5: Set an appropriate threshold T to obtain the binarized visual results BMap _i of each target endmember abundance map FMap _i , 1≤i≤6;

Step S6: optimize the binary visual result BMap _i , 1≤i≤6, and output the final result FinalMAP of license plate forgery information.

Further, the step S1 specifically includes: using a hyperspectral camera to collect data, and obtaining a forged license plate image as a demixing target image. Compared with the real license plate, the counterfeit license plate adds two parts, the blue bottom plate and the white number, which are different from the real license plate material.

Further, the step S2 is specifically:

Based on the above images, a linear spectral mixture model is established, and the specific process includes:

The mixed pixel can be regarded as a linear mixture of endmembers m _i in the image, and the obtained linear mixed model is:

Where: p is the number of endmembers, r is any L-dimensional spectral vector in the image (L is the number of bands), M is an L×p matrix, and each column _mi is an L×1 endmember column vector, α=(α ₁ ,α ₂ ,...,α _p ) ^T is a p×1 abundance vector, and e is an error term.

Further, the step S3 is specifically:

Construct the target endmember matrix set M={m ₁ ,m ₂ ,m ₃ ,m ₄ ,m ₅ ,m ₆ }, M is composed of the target endmember set M1={m ₁ ,m ₂ ,m ₃ ,m ₄ } extracted from the current license plate image and the endmember set M2={m ₅ ,m ₆ } that stores the universal real target. The target endmember set M1 is composed of four endmembers m ₁ , m ₂ , m ₃ , and m ₄ extracted by the automatic target retrieval algorithm ATGP, and the target endmember set M2 is composed of the real license plate blue base plate average spectrum m ₅ and white digital average spectrum m ₆ . The specific process includes:

(1) For the target image to be unmixed, use the automatic target retrieval algorithm ATGP to extract endmembers, and obtain the target endmember set M1. The ATGP algorithm uses the Orthogonal Subspace Projection (OSP) to construct an orthogonal subspace projection to estimate the non-target spectral response, and uses matched filtering to match the target from the data to obtain the endmember of the image. The specific process includes:

Step S31, set the endmember matrix as M;

Step S32, the projection extension space of M is P _M : P _M =M(M ^T M) ^-1 M ^T , then the orthogonal space of M projection space—called the residual projection space, is expressed by express:

Step S33, the projection of any pixel r in M space is The projection residual of any pixel r in M space is r':

r' is a projection residual vector, and its absolute value is the projection length, which is an important criterion for ATGP to select an endmember. The target endmember is obtained by the following formula:

(2) Take the average spectrum m ₅ of the blue base plate of the real license plate and the average spectrum m ₆ of the white digital plate to form the target endmember set M2.

In summary, M1 and M2 together form the target endmember matrix set M.

Further, the step S4 is specifically:

Combining the least square method LS to unmix the given target endmember set M, and give the abundance result map FMap _i corresponding to each endmember m _i , where 1≤i≤6, the specific process includes:

After the data endmember set M is given, the least squares algorithm based on the linear unmixing model is used to find the best matching data of the function by minimizing the square of the error.

Among them, according to the expression of the linear unmixing model, the error value ξ=r-Mα can be obtained, and the expression for finding the optimal solution is as follows:

min{(r-Mα) ^T (r-Mα)} (5)

Further, the unconstrained unmixing abundance α ^LS can be calculated:

α ^LS =(M ^T M) ^T M ^T r (6)

All pixels r _n of the hyperspectral image are traversed according to the above process, so as to obtain the abundance result map corresponding to each end member. Among them, 1≤n≤N, N is the total number of pixels in the hyperspectral image.

Further, the step S5 is specifically:

Set an appropriate threshold T to obtain the binarized visual results BMap _i of each target endmember abundance map FMap _i . The specific process includes:

Step S51 , for the unmixed abundance map generated by the endmember set M1={m ₁ , m ₂ , m ₃ ,m ₄ } extracted by ATGP, set a fixed threshold to t ₁ , perform binarization processing, and obtain a binarized visual result BMap _i , 1≤i≤4.

Step S52, for the endmember set M2={m ₅ ,m ₆ } composed of the real license plate blue base plate average spectrum m ₅ and white digital average spectrum m ₆ , the unmixed abundance image FMap _i is obtained, where 5≤i≤6, and its threshold is automatically determined by Otsu. The Otsu process is as follows:

According to the grayscale characteristics of the image, the image is divided into two parts, the background and the target.

(T)＝W _A (μ _a -μ) ² +W _B (μ _b -μ) ² (7)

Among them, (T) is the maximum variance between the two classes, W _A is the probability of class A, μ _a is the average gray level of class A, W _B is the probability of class B, μ _b is the average gray level of class B, and μ is the overall average gray level of the image. That is, the threshold T divides the image into two parts, A and B, so that T, which takes the maximum value of the total variance of the two types, is the optimal segmentation threshold.

Further, the step S6 is specifically:

Optimize the binary visual result BMap _i , 1≤i≤6, and output the final result FinalMAP of license plate forgery information. The specific process includes:

(1) Optimize the binary visualization results. Due to the interference of factors such as angle, exposure and noise in the hyperspectral image shooting and imaging process, the unmixing and detection of the license plate will be affected to a certain extent. Therefore, this method optimizes the binarized result BMap _i from the following two aspects, and obtains the optimized binary result

Step S61, for the small-area interference information of the BMap _i image due to exposure or noise in the image itself, delete objects with an area smaller than p in the binary image BW by the following formula to obtain an optimized binary result 1≤i≤6:

Among them, p is set as the lower limit threshold of the area, conn is the corresponding neighborhood search method, and the default is 8, which means 8 neighborhood search.

Step S62, the wrong segmentation of narrow and long areas caused by poor image shooting angles and tilted license plates, the processing process is as follows:

First, mark the connected domain D in the binary image, calculate the minimum circumscribed rectangle of each connected domain, and obtain its length (L) and width (W). Second, set the fault-tolerant parameters NL and NW to limit the maximum scale of the distribution of narrow and long regions. Finally, compare L, NL and W, NW respectively. When L>α(NL) or W<β(NW), the connected region D does not meet the character characteristics of the license plate, and it is deleted to obtain an optimized binary result 1≤i≤6. Among them, α and β are elastic coefficients, 0≤α, β≤1, which control the size of the actual license plate character features.

(2) Output license plate forgery information FinalMAP, the process is as follows:

Considering the fact that the target endmember set M1 obtained by the automatic target retrieval algorithm ATGP may contain both real targets and forged target endmembers, while the endmembers in the target endmember set M2 are all real license plate endmembers, consider comparing the difference between the two binary results to determine whether the license plate contains forged information, and output the final result FinalMAP of the license plate forged information. The specific process is as follows:

Take the optimized binary result corresponding to the target endmember set M1 The union of get FinalMap(M1), 1≤i≤4, as follows:

Wherein, FinalMap(M1) represents the detection result of the license plate forgery information corresponding to the target endmember set M1; Indicates the optimized binarized detection results corresponding to each end member.

Take the optimized binary result corresponding to the target endmember set M2 The union of get FinalMap(M2), 5≤i≤6, as follows:

Wherein, FinalMap(M2) represents the detection result of the license plate forgery information corresponding to the target endmember set M2; Indicates the optimized binarized detection results corresponding to each end member.

FinalMap＝FinalMap(M1)-FinalMap(M2) (11)

Among them, FinalMAP is the final detection result of license plate forgery information.

The present invention applies the hyperspectral unmixing technology to the field of authenticity detection of license plates. Compared with the traditional RGB image detection technology, the authenticity detection method of the license plate based on the hyperspectral unmixing technology can accurately identify fake license plate information, and the accuracy of the detection result is higher than that of the traditional detection effect when the material of the authentic license plate is relatively similar and cannot be distinguished by human eyes. The license plate authenticity detection method based on hyperspectral unmixing technology does not need to know the prior information of the target, and it is an unsupervised detection method. In the case that the forged target spectrum is unknown, the target spectral information can be automatically obtained from the image for subsequent spectral unmixing. In addition, various optimization processes have been carried out on the interference information generated by factors such as shooting, illumination and noise, so as to make the detection results more accurate. Finally, the detection results are visualized by means of the result difference.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or prior art. Obviously, the accompanying drawings in the following description are only some embodiments recorded in the application. For those of ordinary skill in the art, other accompanying drawings can also be obtained based on these drawings without creative work.

Fig. 1 is method flowchart of the present invention;

Fig. 2 is a license plate image in an embodiment of the present invention;

Fig. 3 is a schematic diagram of the extraction results of license plate endmembers in an embodiment of the present invention;

Fig. 4 is a schematic diagram of license plate unmixing results in an embodiment of the present invention;

Fig. 5 is a schematic diagram of the binarization result of the license plate in an embodiment of the present invention;

Fig. 6 is a schematic diagram of the optimized binarization result of the license plate in an embodiment of the present invention;

Fig. 7 is a schematic diagram of the final detection result of the license plate in an embodiment of the present invention.

Detailed ways

In order to make the technical solutions and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the drawings in the embodiments of the present invention:

A kind of license plate authenticity detection method based on hyperspectral unmixing technology as shown in Figure 1, specifically comprises the following steps:

Step S1, using a hyperspectral camera to collect data, and obtaining a forged license plate image as a demixing target image. Compared with the real license plate, the counterfeit license plate has two parts, the blue bottom plate and the white number, which are different from the material of the real license plate;

Step S2, establishing a linear spectral mixture model according to the above images;

Step S3: Construct the target endmember matrix set M={m ₁ ,m ₂ ,m ₃ ,m ₄ ,m ₅ ,m ₆ }, M is composed of the target endmember set M1={m ₁ ,m ₂ ,m ₃ ,m ₄ } extracted from the current license plate image and the endmember set M2={m ₅ ,m ₆ } for storing universal real targets. The target endmember set M1 is composed of four endmembers m ₁ , m ₂ , m ₃ , and m ₄ extracted by the automatic target retrieval algorithm ATGP, and the target endmember set M2 is composed of the average spectrum m ₅ of the blue bottom plate of the real license plate and the average spectrum m ₆ of the white number;

Step S4, combine the least square method LS to unmix the given target endmember set M, and give the abundance result map FMap _i corresponding to each endmember m _i , where 1≤i≤6;

Step S5. Set an appropriate threshold T to obtain the binarized visual results BMap _i of each target endmember abundance map FMap _i , 1≤i≤6;

Step S6 , optimize the binary visual result BMap _i , 1≤i≤6, and output the final result FinalMAP of license plate forgery information.

In this embodiment, the step S1 is specifically:

The hyperspectral camera is used to collect data, and the forged license plate image is obtained as the unmixed target image. The forged license plate is shown in Figure 2. Compared with the real license plate, the counterfeit license plate adds two parts, the blue bottom plate and the white number, which are different from the real license plate material. The red area in Figure 2 is the part of the fake license plate, where the fake license plate I is made of metal material, and the fake license plate C is made of sticker material.

In this embodiment, the step S2 is specifically:

Based on the above images, a linear spectral mixture model is established, and the specific process includes:

The mixed pixel can be regarded as a linear mixture of endmembers m _i in the image, and the obtained linear mixed model is:

Where: p is the number of endmembers, r is any L-dimensional spectral vector in the image (L is the number of bands), M is an L×p matrix, and each column _mi is an L×1 endmember column vector, α=(α ₁ ,α ₂ ,...,α _p ) ^T is a p×1 abundance vector, and e is an error term.

In this embodiment, the step S3 is specifically:

Construct the target endmember matrix set M={m ₁ ,m ₂ ,m ₃ ,m ₄ ,m ₅ ,m ₆ }, M is composed of the target endmember set M1={m ₁ ,m ₂ ,m ₃ ,m ₄ } extracted from the current license plate image and the endmember set M2={m ₅ ,m ₆ } that stores the universal real target. The target endmember set M1 is composed of four endmembers m ₁ , m ₂ , m ₃ , and m ₄ extracted by the automatic target retrieval algorithm ATGP, and the target endmember set M2 is composed of the real license plate blue base plate average spectrum m ₅ and white digital average spectrum m ₆ . The specific process includes:

For the target image to be unmixed, use the automatic target retrieval algorithm ATGP to extract endmembers, and obtain the target endmember set M1={m ₁ ,m ₂ ,m ₃ ,m ₄ }. Ideally, the endmembers m ₁ , m ₂ , m ₃ , and m ₄ can represent four types of spectra: ① the real blue bottom plate spectrum of the current license plate; ② the real white digital spectrum of the current license plate; ③ the fake blue bottom plate spectrum of the current license plate; ④ the fake white digital spectrum of the current license plate.

The ATGP algorithm uses the Orthogonal Subspace Projection (OSP) to construct an orthogonal subspace projection to estimate the non-target spectral response, and uses matched filtering to match the target from the data to obtain the endmember of the image. The specific process includes:

Step S31, set the endmember matrix as M;

Step S32, the projection extension space of M is P _M : P _M =M(M ^T M) ^-1 M ^T , then the orthogonal space of M projection space—called the residual projection space, is expressed by express:

Step S33, the projection of any pixel r in M space is The projection residual of any pixel r in M space is r':

r' is a projection residual vector, and its absolute value is the projection length, which is an important criterion for ATGP to select an endmember. The target endmember is obtained by the following formula:

Finally, the extraction results of the four target endmembers in the target endmember set M1: m ₁ , m ₂ , m ₃ , and m ₄ are shown in FIG. 3 .

(2) In a complex background, the ATGP algorithm may not be able to extract the corresponding real blue cards and white digital spectra, so as to reduce the unreliability of the ATGP algorithm to extract endmembers and improve the accuracy of subsequent detection. Therefore, take the average spectrum _m5 of the blue base plate of the real license plate and the average spectrum _m6 of the white number as fixed end members to form the target end member set M2.

In summary, M1 and M2 together form the target endmember set M.

In this embodiment, the step S4 is specifically:

Combining the least square method LS to unmix the given target endmember set M, and give the abundance result map FMap _i corresponding to each endmember m _i , where 1≤i≤6, the specific process includes:

After the data endmember set M is given, the least squares algorithm based on the linear unmixing model is used to find the best matching data of the function by minimizing the square of the error.

Among them, according to the expression of the linear unmixing model, the error value ξ=r-Mα can be obtained, and the expression for finding the optimal solution is as follows:

min{(r-Mα) ^T (r-Mα)} (5)

Further, the unconstrained unmixing abundance α ^LS can be calculated:

α ^LS =(M ^T M) ^T M ^T r (6)

All pixels r _n of the hyperspectral image are traversed according to the above process, so as to obtain the abundance result map corresponding to each end member. Among them, 1≤n≤N, N is the total number of pixels in the hyperspectral image.

In this embodiment, the step S5 is specifically:

Set an appropriate threshold T to obtain the binarized visual results BMap _i of each target endmember abundance map FMap _i . The specific process includes:

Step S51, for the unmixed abundance map generated by the endmember set M1={m ₁ ,m ₂ ,m ₃ ,m ₄ } extracted by ATGP, set a fixed threshold value of t ₁ =0.6, perform binarization processing, and obtain the binarized visual result BMap _i , 1≤i≤4, the results are shown in Figure 5(a)-(d).

Step S52, for the endmember set M2={m ₅ , m ₆ } composed of the real license plate blue base plate average spectrum m ₅ and white digital average spectrum m ₆ , obtain the unmixed abundance image FMap _i , where 5≤i≤6, the results are shown in Figure 5(e)-(f). Its threshold is automatically determined by Otsu. The Otsu process is as follows:

According to the grayscale characteristics of the image, the image is divided into two parts, the background and the target.

(T)＝W _A (μ _a -μ) ² +W _B (μ _b -μ) ² (7)

Among them, (T) is the maximum variance between the two classes, W _A is the probability of class A, μ _a is the average gray level of class A, W _B is the probability of class B, μ _b is the average gray level of class B, and μ is the overall average gray level of the image. That is, the threshold T divides the image into two parts, A and B, and (T) makes the total variance of the two types take the maximum T, which is the optimal segmentation threshold.

In this embodiment, the step S6 is specifically:

Optimize the binary visual result BMap _i , 1≤i≤6, and output the final result FinalMAP of license plate forgery information. The specific process includes:

(1) Optimize the binary visualization results. Due to the interference of factors such as angle, exposure and noise in the hyperspectral image shooting and imaging process, the unmixing and detection of the license plate will be affected to a certain extent. Therefore, this method optimizes the binarized result BMap _i from the following two aspects, and obtains the optimized binary result

Step S61, for the small-area interference information formed by the image due to exposure or noise in the image itself, delete objects with an area smaller than p=10 in the binary image BW by the following formula to obtain an optimized binary result

Among them, p is set as the lower limit threshold of the area, conn is the corresponding neighborhood search method, and the default is 8, which means 8 neighborhood search.

Step S62, the wrong segmentation of narrow and long areas caused by poor image shooting angles and tilted license plates, the processing process is as follows:

First, mark the connected domain D in the binary image, calculate the minimum circumscribed rectangle of each connected domain, and obtain its length (L) and width (W). Second, set the fault-tolerant parameters NL and NW to limit the maximum scale of the distribution of narrow and long regions. Finally, compare L, NL and W, NW respectively. When L>α(NL) or W<β(NW), the connected region D does not meet the character characteristics of the license plate, and it is deleted to obtain an optimized binary result 1≤i≤6. Among them, α and β are elastic coefficients, 0≤α, β≤1, which control the size of the actual license plate character features. When the elastic coefficient is small, a large area of narrow and long areas can be eliminated to accurately detect the results. But in practice, the elastic coefficient is not as small as possible. Because if the elastic coefficient is too small, the real target will be removed as interference information. In the present invention, α=β=1/5 is set to ensure maximum extraction of effective information and improve the accuracy of results. Each end member corresponds to the optimization result of the binarization result BMap _i /> 1≤i≤6, as shown in Figure 6.

(2) Output license plate forgery information FinalMAP, the process is as follows:

Considering the fact that the target endmember set M1 obtained by the automatic target retrieval algorithm ATGP may contain both real targets and forged target endmembers, while the endmembers in the target endmember set M2 are all real license plate endmembers, consider comparing the difference between the binary results of the two to determine whether the license plate contains forged information, and output the final result FinalMAP of the license plate forged information.

The specific process is as follows:

Take the optimized binary result corresponding to the target endmember set M1 The union of get FinalMap(M1), 1≤i≤4, as follows:

Wherein, FinalMap(M1) represents the detection result of the license plate forgery information corresponding to the target endmember set M1; Indicates the optimized binarized detection results corresponding to each end member.

Take the optimized binary result corresponding to the target endmember set M2 The union of get FinalMap(M2), 5≤i≤6, as follows:

Wherein, FinalMap(M2) represents the detection result of the license plate forgery information corresponding to the target endmember set M2; Indicates the optimized binarized detection results corresponding to each end member.

Since the target endmember set M1 extracted by ATGP contains both the spectral features of forged information and the real license plate, while the target endmember set M2 only contains the real license plate spectral features, FinalMap(M1) can be regarded as the comprehensive detection result of the fake target and the real target, and FinalMap(M2) can be regarded as the real target detection result. Furthermore, the result values corresponding to the fake targets and real targets in the FinalMap(M1) and FinalMap(M2) graphs should contain the following characteristics, as shown in the table:

Calculate the difference between FinalMap(M1) and FinalMap(M2) by formula (11) to determine whether the license plate contains forged information. The difference is shown in the table below. The pixel points with a median value of 1 in the FinalMAP result are judged as fake license plate points, and the points with a median value of -1 or 0 are judged as real license plate points. Finally, the final result FinalMAP of license plate forgery information is output, and the detection results are shown in the table.

FinalMap＝FinalMap(M1)-FinalMap(M2) (11)

Among them, FinalMAP is the final detection result of license plate forgery information.

In this embodiment, hyperspectral unmixing technology is used to detect license plates containing forged objects, which effectively solves the problem that traditional RGB image detection techniques are difficult to distinguish between genuine and fake license plates of similar materials. In this embodiment, first, the ATGP algorithm is used to automatically extract the target end-meta information for subsequent unmixing without prior information, which is a non-supervised detection method. Secondly, a variety of optimization processes have been performed on the interference information generated by some external factors to reduce its impact on the detection results. Finally, in order to reduce the instability of the automatic endmember extraction method ATGP to extract endmembers, this embodiment compares the difference between the binary results of the ATGP endmember set and the real license plate endmember set to determine whether the license plate contains forged information, thereby enhancing the reliability of the detection results.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person familiar with the technical field within the technical scope disclosed in the present invention, according to the technical solution of the present invention and its inventive concept to make equivalent replacements or changes, should be covered within the scope of protection of the present invention.

Claims (6)

1. A license plate authenticity detection method based on hyperspectral unmixing technology, characterized in that it comprises: Use a hyperspectral camera to obtain a forged license plate image and use it as the unmixed target image; Build a linear spectral mixing model based on the unmixed target image; Construct the target endmember matrix set M, wherein the target endmember matrix set M includes the target endmember set M1 extracted from the current license plate image and the endmember set M2 storing the ubiquitous real target, wherein the target endmember set M1 includes four endmembers m ₁ , m ₂ , m ₃ , and m ₄ extracted by the automatic target retrieval algorithm ATGP, and the target endmember set M2 includes the real license plate blue bottom plate average spectrum m ₅ and white digital average spectrum m _6; Combine the least squares method to unmix the given target endmember set M, and obtain the abundance result map FMap _i corresponding to each endmember m _i , where 1≤i≤6; Set the threshold T to obtain the binarized visual results BMap _i of each target endmember abundance map FMap _i , 1≤i≤6; optimize the binary visual results BMap _i to obtain the final result FinalMAP of license plate forgery information. 2. The method according to claim 1, further characterized in that: the automatic target retrieval algorithm ATGP is used to carry out endmember extraction to the target endmember matrix set M to obtain the target endmember matrix set M1: the orthogonal subspace projection method OSP is used to construct the orthogonal subspace projection for estimating the non-target spectral response, and the matched filter is used to match the target from the data so as to obtain the endmember of the image. 3. The method according to claim 2, further characterized in that: when the automatic target retrieval algorithm ATGP is adopted to obtain the target endmember matrix set M1: Let the projection expansion space of the target endmember matrix set M be P _M : P _M ＝M(M ^T M) ^-1 M ^T , then the orthogonal space of the M projection space is called the residual projection space. express: The projection of any pixel r in M space is The projection residual of any pixel r in M space is r': r' is a projection residual vector, and its absolute value is the projection length, which is an important criterion for ATGP to select an endmember. The target endmember is obtained by the following formula: 4. The method according to claim 1, further characterized in that: the binarized visual result BMap _i of each target endmember abundance map FMap _i is specifically obtained in the following manner: Set a fixed threshold t ₁ , perform binarization on the unmixed abundance map generated by the target endmember matrix set M1={m ₁ ,m ₂ ,m ₃ ,m ₄ } to obtain the binarized visual result BMap _i , 1≤i≤4; Binarize the unmixed abundance image FMap _i generated by the endmember set M2={m ₅ ,m ₆ }, where 5≤i≤6, the threshold _t2 is automatically determined by Otsu, The Otsu process is: according to the grayscale characteristics of the image, the image is divided into two parts, the background and the target (T)＝W _A (μ _a -μ) ² +W _B (μ _b -μ) ² (4) Among them, (T) is the maximum variance between the two classes, W _A is the probability of class A, μ _a is the average gray level of class A, W _B is the probability of class B, μ _b is the average gray level of class B, and μ is the overall average gray level of the image, that is, the threshold T divides the image into two parts, A and B, so that the total variance of the two classes takes the maximum T, which is the optimal segmentation threshold. 5. The method according to claim 1, further characterized in that: when the binary visual result BMap _i is optimized: for the small-area interference information formed by exposure or the noise of the image itself, the binary image BW is deleted by the following formula to delete the object _whose area is less than p in the binary image BW, to obtain the optimized binary result 1≤i≤6: Among them, p is set as the area lower limit threshold, and conn is the corresponding neighborhood search method; The mis-segmentation of long and narrow areas caused by poor image shooting angles and inclined license plates is handled as follows: mark the connected domain D in the binary image, calculate the minimum circumscribed rectangle of each connected domain, and obtain its length L and width W, set the error tolerance parameters NL and NW to limit the maximum scale of the distribution of narrow and long areas, and compare L, NL, W, and NW respectively. Optimizing Binary Results Among them, α and β are elastic coefficients, 0≤α, β≤1, which control the size of the actual license plate character features. 6. The method according to claim 1, further characterized in that: obtaining the final result FinalMAP of license plate forgery information specifically adopts the following methods: Take the optimized binary result corresponding to the target endmember set M1 The union of get FinalMap(M1), 1≤i≤4, as follows: Wherein, FinalMap(M1) represents the detection result of the license plate forgery information corresponding to the target endmember set M1; Indicates the optimized binarized detection results corresponding to each end member. Take the optimized binary result corresponding to the target endmember set M2 The union of get FinalMap(M2), 5≤i≤6, as follows: Wherein, FinalMap(M2) represents the detection result of the license plate forgery information corresponding to the target endmember set M2; Indicates the optimized binary detection results corresponding to each end member; FinalMap＝FinalMap(M1)-FinalMap(M2) (8) Among them, FinalMAP is the final detection result of license plate forgery information.