An image recognition method based on ultra-deep adversarial learning

Abstract

The present invention relates to a method of ultra-deep adversarial learning in the field of information processing, which is characterized in that when evaluating the characteristic elements of the probability distribution of a logged probability space to which the characteristic elements of the recognition object belong, the final result is obtained through adversarial learning established by the characteristic elements of the probability distribution of the two closest probability spaces, and the adversarial learning is evaluated based on the fuzzy event probability measure constructed based on the distance across different spaces. The implementation effect of the present invention is: adversarial learning based on the strict fuzzy event probability measure relationship established between a data and two probability distributions can obtain higher classification accuracy, can make the accuracy of pattern recognition reach the best accuracy, and provide an algorithm with higher accuracy learning effect for unsupervised machine learning.

Description

Technical Field

[0001] The present invention belongs to an image recognition method of ultra-deep adversarial learning in the field of artificial intelligence. Background Art

[0002] With the brilliant achievement of AlphaGo, developed by Google, defeating all human chess players, the world has once again been in a frenzy of pursuing deep learning. In the past year, the number of patent applications related to artificial intelligence has almost exceeded the total number of patents related to artificial intelligence in all previous years.

[0003] In this regard, the famous Japanese company Furukawa Electric has published a patent application for "Image Processing Method and Image Processing Device" (Patent Document 1). The patent proposes to select the processing threshold of the image through the algorithm of artificial intelligence neural network, so as to extract the contour of the image with high precision.

[0004] In the application of autonomous driving, Toyota, a well-known Japanese company, has published a patent for a "driving direction estimation device" (Patent Document 2). This patent proposes that in the process of autonomous driving, in response to sudden situations, even if the driver does not respond, the machine learning algorithm of the artificial intelligence's inverse transfer neural network can automatically select the driving state to avoid traffic accidents.

[0005] Application of image analysis: Hosei University in Japan published a patent for "Plant disease diagnosis system, plant disease diagnosis method, and plant disease diagnosis program" (Patent Document 3). The patent proposes to introduce a deep learning CNN model to identify and analyze images of plant leaves to diagnose plant lesions.

[0006] Fuji Xerox of Japan, one of the world's largest copier manufacturers, has applied for a patent for "Anti-counterfeiting device and method using tiny anti-counterfeiting marks" from the perspective of anti-counterfeiting identification (Patent Document 4). This patent attempts to use artificial intelligence algorithms to construct tiny anti-counterfeiting marks that can record information in order to achieve the purpose of solving product anti-counterfeiting identification.

[0007] [Patent Literature]

[0008] [Patent Document 1] (Japanese Patent Application Laid-Open No. 2013-109762)

[0009] [Patent Document 2] (Japanese Patent Application Laid-Open No. 2008-225923)

[0010] [Patent Document 3] (Japanese Patent Application Laid-Open No. 2016-168046)

[0011] [Patent Document 4] (Japanese Patent Application Laid-Open No. 2012-124957)

[0012] The above-mentioned (Patent Document 1), (Patent Document 2) and (Patent Document 3) all mention the neural network algorithm using artificial intelligence. However, in the neural network algorithm, the information of the target function is mainly carried on a large number of weighted parameters through "training". In the process of learning the weighted value W and the threshold T, in order to obtain the best solution, all states need to be tested. The total number of combinations required is {(W×T) n}×P, where n is the number of nodes in a layer of neural network, and P is the number of layers of the neural network. Such a high exponential computational complexity makes the computational workload huge, and the hardware overhead required is huge. The probability gradient descent method used in the loss function of the learning effect of deep learning is referred to as SGD. The training value obtained is a local optimal solution, so the "black box" problem will inevitably appear. In addition, the threshold in the neural network model is artificially defined, which is very different from the mechanism of the neural network of the human brain. The principle of the stimulation signal of the brain nerve cannot be fully reflected in the traditional neural network model. The mechanism that the human brain makes different judgments based on the different excitement levels of the neural signals of the neurons cannot be reflected in the current neural network model, etc. The current neural network model can only be academic, representing a directional theory, and the gap with the level of practical application is very large. Now that we have entered the stage of deep learning, compared with the traditional neural network, only the number of hidden layers has been increased, which further increases the complexity of the calculation. Although some optimization algorithms have been introduced in the learning, they have not deviated from the original neural network foundation. The fatal problems of the traditional neural network cannot be solved, and the prospect of widespread application is difficult to expect.

[0013] The above (Patent Document 4) proposed to generate tiny anti-counterfeiting marks through artificial intelligence is far-fetched. First of all, it did not propose any special means to use artificial intelligence? What outstanding results have been achieved? The method of recording information by the size, color and different positions of the dot matrix used in the micro-dot matrix information recording method of the tiny anti-counterfeiting mark is already a public means. After the introduction of artificial intelligence, the technical characteristics of the traditional method have not changed, so it is not advanced. In the method of representing the characteristic symbols or reference symbols of the direction and position of the information dot matrix, the image of the non-information dot matrix graphic that is much larger than the size of the information dot matrix is ​​used, so that the tiny anti-counterfeiting mark is easy to be discovered by the lawbreakers in terms of concealment, which does not conform to the composition rules of the invisible anti-counterfeiting code image. The biggest problem is that it cannot solve the problem of the lawbreaker copying the tiny anti-counterfeiting mark through the scanner, because in reality the accuracy of the scanner is much greater than the printing accuracy. Although the patent application once proposed to form a tiny anti-counterfeiting mark by laser processing, anti-counterfeiting through tiny graphics can be achieved through many methods. At present, the most concerned issue is how to achieve mass anti-counterfeiting identification through consumers' mobile phones, which this patent application does not touch at all.

[0014] The professional terms involved in this patent application are defined below, and the following definitions include the content of the present invention.

[0015] Probability Space:

[0016] Based on the theory of Soviet mathematician Andrey Kolmogorov that "probability theory is based on measure theory", the so-called probability space is a measurable space with a total measure of "1".

[0017] Probability Distribution:

[0018] The so-called probability distribution is the arrangement of probability functions according to the likelihood of their occurrence.

[0019] Probability Scale:

[0020] For any probability distribution in the probability space, there must be a probability scale that can calibrate the degree of the probability distribution.

[0021] Probability Density:

[0022] The probability value of an event occurring in a given area.

[0023] Probability Measure of Fazzy Event

[0024] In the Euclidean space S, including the probability space, if p(x) satisfies the additivity of the probability measure, μ A (x) is a membership function and also satisfies the additivity of fuzzy measure. Then the measure P(A) of the set A of fuzzy event probabilities is:

[0025] Formula 1

[0026] P(A)=∫ S μ A (x)p(x)dx

[0027] Its discrete expression formula is:

[0028] Formula 2

[0029]

[0030] Intelligent System:

[0031] An intelligent system is a system implemented based on a deterministic algorithm. It is a system that implements a certain objective function according to an algorithm, and its processing result is a deterministic system.

[0032] Artificial Intelligence:

[0033] What is artificial intelligence? Simply put, it is to use computers to realize the functions of human brains, that is, to use computers to realize the effects of human brain thinking. The problems to be solved are often uncertain, or unpredictable in advance.

[0034] It can also be said that artificial intelligence is the modeling and standardization of the effects of human intervention. A more specific definition is that artificial intelligence is the most in-depth application of probabilistic models through machine learning algorithms.

[0035] Clustering:

[0036] Based on the scale of Euclidean space, data can be migrated without a target, and ultimately only data results within a given range can be obtained.

[0037] Self-organization:

[0038] Based on the scale of the probability space, data can be autonomously migrated to the direction of high probability, and ultimately an unpredictable objective function can be obtained.

[0039] Definition of Machine Learning:

[0040] A model in which computers autonomously learn regularities from data.

[0041] Self-organizing based on a Probability Scale:

[0042] Assume that there exists a probability distribution set G in the probability space, which contains ζ data:

[0043] Formula 3

[0044] g f ∈G(f=1,2,…,ζ)

[0045] This probability distribution g in the probability space f (f = 1, 2, ..., ζ), there must be an eigenvalue A(G). Since the probability space is a measure space, there must be a probability scale M[G, A(G)] for the eigenvalue A(G). When the following probability scale self-organization conditions are met, the probability scale can be used as a reference to let the set G(n) Migrate in the direction of maximum probability.

[0046] Formula 4

[0047] A (n) =A(G (n) )

[0048] M (n) =M[G (n) , A(G (n) )]

[0049] G (n) =G{[A(G (n-1) ), M[G (n-1) , A(G (n-1) )]]

[0050] When n≥β (β is a value greater than 4), we can set A(G (n) ) is the maximum probability eigenvalue, M[G (n) , A(G (n) )] is the maximum probability scale centered on the maximum probability eigenvalue.

[0051] Maximum Probability:

[0052] The results go beyond those predicted by traditional statistics, that is, the predicted results that are closest to the parent population.

[0053] Probability Space Distance:

[0054] Assume that there are two probability distributions w in the probability space j ∈W,v j ∈V, and the corresponding v j The maximum probability scale M of each element of the probability distribution vector ∈V j , where (j=1, 2, ..., n) then w j to v j The distance between the maximum probability scales of two probability distributions is defined as follows:

[0055] Formula 5

[0056]

[0057] Super Deep Learning:

[0058] A neural network model that is composed of a perception layer, a neural layer, and a cerebral cortex, and is connected by nodes between layers using self-organizing machine learning on a probabilistic scale. It can directly perform autonomous and unsupervised machine learning on randomly distributed data. Summary of the invention

[0059] The first purpose of the present invention is to propose a multi-scale probability scale self-organizing algorithm, construct an adversarial learning model, and maximize the use of all probability distribution information and fuzzy information in the probability space through ultra-deep learning, so as to achieve the best level of accuracy in image recognition and speech recognition.

[0060] The second purpose of the present invention is to provide a more rigorous definition of distance in different spaces including probability space based on the self-organization of multi-scale probability scales.

[0061] The third object of the present invention is to provide a definition of fuzzy event probability measure that can traverse different spaces including probability space based on the self-organization of multi-scale probability scales.

[0062] The fourth object of the present invention is to provide an ultra-deep regression analysis learning model based on probability scale self-organization.

[0063] The fifth object of the present invention is to provide a nonlinear probability scale self-organizing ultra-deep manifold learning model.

[0064] The sixth objective of the present invention is to provide an ultra-deep and strong adversarial learning model that can distinguish authenticity logos through popular mobile phones.

[0065] In order to achieve at least one of the above purposes, a method for image recognition using ultra-deep adversarial learning is proposed. The present invention proposes the following technical solutions:

[0066] A method for image recognition using ultra-deep adversarial learning, which has the following characteristics:

[0067] (1) When evaluating the characteristic elements of the probability distribution of the registered probability space to which the characteristic elements of the recognition object belong, the final result is obtained through adversarial learning established by the characteristic elements of the probability distribution of the two closest probability spaces;

[0068] (2) The above adversarial learning is evaluated based on the fuzzy event probability measure constructed based on the distance across different spaces.

[0069] Moreover, the adversarial learning refers to: when calculating the fuzzy event probability measure of the feature elements of the identified object belonging to the probability distribution of a logged in probability space; at the same time, calculating the fuzzy event probability measure of the feature elements of the identified object belonging to the probability distribution of another logged in probability space; confronting these two fuzzy event probability measures; and finally determining to which logged in probability space the probability distribution the feature elements belong.

[0070] Moreover, the adversarial learning is based on the characteristic elements to which the maximum probability confidence belongs.

[0071] Moreover, the distance based on crossing different spaces refers to: the probability information in the complex field of the probability distribution obtained through self-organizing learning of multiple probability scales; or the distance across different spaces obtained by learning the probability information of multiple probability scales of the actual probability distribution.

[0072] Moreover, the probability distribution refers to: normal distribution; multivariate normal distribution; lognormal distribution; exponential distribution; t distribution; F distribution; X 2 Distribution; Binomial distribution; Negative binomial distribution; Multinomial distribution; Poisson distribution; Erlang distribution; Hypergeometric distribution; Geometric distribution; Traffic distribution; Weibull distribution; Triangular distribution; Beta distribution; Gamma distribution or any probability distribution derived from Bayesian analysis; Gaussian process.

[0073] Moreover, the distances of different spaces refer to any combination of spaces in: Euclidean space; probability space including Manhattan space; Chebyshev space; Minkowski space; Mahalanobis space; and angle cosine space.

[0074] Moreover, the fuzzy event probability measure can be expressed by the following formula:

[0075] Suppose there exists a set of probability spaces w in Euclidean space j ∈W(w1,w2,…,w n ) has an element w j The center point of the probability distribution is w j The first scale of the multi-probability scale of this probability distribution is M1, the second scale is M2, the third scale is M3, and the center point is w j The distance from the first scale M1 is D 1j =M 1j , in this interval belongs to the set w j The probability value of the probability distribution is P 1j (wj) , the distance from the second scale to the first scale is D 2j=M 2j -M 1j , in this interval belongs to the set w j The probability value of the probability distribution is P 2j (wj) , the distance from the third scale to the second scale is D 3j =M 3j -M 2j The probability value of the probability distribution belonging to the set W in this interval is P 3j (wj) , assuming that the crossing belongs to w j The number of scales m passed in the probability space j (wj) =3, and suppose that there is another probability space set v in Euclidean space j ∈V(v1,v2,…,v n ), there is another point in the middle of the two sets, any point r j ∈R(j=1,2,…,n), find r j In w j to v j Between j Belong to w j The probability value of the probability distribution is rp j (wj) , find r j Arrival j Fuzzy event probability measure:

[0076]

[0077] The present invention proposes an ultra-deep adversarial learning method for image recognition, which is characterized in that adversarial learning is performed on the strict fuzzy event probability measurement relationship established between a data and two probability distributions, which can obtain higher classification accuracy and make the accuracy of pattern recognition reach the best accuracy, providing an algorithm with higher accuracy learning effect for unsupervised machine learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 A diagram of multiple probability scales defined for probability distributions

[0079] Figure 2 It is a multi-probability scale self-organizing machine learning flowchart

[0080] Figure 3 This is a diagram of the definition of distance across different spaces including probability space.

[0081] Figure 4 This is a schematic diagram of the pattern recognition model of image recognition using ultra-deep adversarial learning.

[0082] Figure 5 This is a schematic diagram of the best classification model for image recognition using ultra-deep adversarial learning.

[0083] Figure 6 The best classification flowchart for image recognition using ultra-deep adversarial learning

[0084] Figure 7 It is a flowchart of the scale of multiple probability scales to obtain the actual probability distribution through learning.

[0085] Figure 8 This is a schematic diagram of a super deep adversarial learning neural network.

[0086] Fig. 9 This is a flowchart of a super deep adversarial learning neural network.

[0087] Fig.10 This is a flowchart of ultra-deep strong adversarial learning for mobile phone anti-counterfeiting identification

[0088] Fig.11 This is a schematic diagram of the characteristics of the printed image color space and the electronic image color space.

[0089] Fig.12 This is a flowchart for solving regression analysis using ultra-deep adversarial learning

[0090] Fig.13 This is a comparison diagram of the results of the two regression analyses.

[0091] Fig.14 This is a schematic diagram of the algorithm and calculation results of ultra-deep manifold learning.

[0092] Explanation of symbols

[0093] 101 is a probability distribution in probability space

[0094] 102 is the center value of the probability distribution in a certain area

[0095] 103 is the first scale value

[0096] 104 is the second scale value

[0097] 105 is the third scale value

[0098] 106 is the first scale value 103 corresponds to the probability distribution value in the area

[0099] 107 is the second scale value 104 corresponds to the probability distribution value in the area

[0100] 108 is the third scale value 105 corresponds to the probability distribution value in the area

[0101] 301 is a Euclidean space covering the probability space

[0102] 302 is the center point w of a probability distribution in probability space j

[0103] 303 is the first scale M1 of a multi-probability scale of a probability distribution in probability space

[0104] 304 is the second scale M2 of the multi-probability scale of a probability distribution in probability space

[0105] 305 is the third scale M3 of the multi-probability scale of a probability distribution in probability space

[0106] 309 is a point v in Euclidean space j

[0107] 310 is a point v in Euclidean space j ∈V to the center point w of the probability distribution j ∈W any point r j ∈R(j=1,2,…,n)

[0108] 4000 is the identified object vector

[0109] 4001 is the characteristic element sv1 of the identified image

[0110] 4002 is the characteristic element sv2 of the identified image

[0111] 4003 is the characteristic element sv3 of the recognized image

[0112] 400e is the characteristic element sv of the recognized image e

[0113] 4100 is the feature vector data FV1 after self-organization through multiple probability scales

[0114] 4200 is the feature vector data FV2 after self-organization through multiple probability scales

[0115] 4110 is fv 11 The central value of

[0116] 4111 is the feature vector data fv1 11 The first scale

[0117] 4112 is fv 11 The second scale is fv 11 The third scale

[0118] 4120 is fv 12 The center value 4121 is fv 12 The first scale

[0119] 4122 is fv 12 The second scale 4123 is fv 12 The third scale

[0120] 4130 is fv 13 The center value 4131 is fv 13 The first scale

[0121] 4132 is fv 13 The second scale 4133 is fv 13 The third scale

[0122] 41e0 is fv 1e The center value 41e1 is fv 1e The first scale

[0123] 41e2 is fv 1e The second scale 41e3 is fv 1e The third scale

[0124] 4210 is fv 21 The central value of

[0125] 4211 is the feature vector data fv2 21 The first scale

[0126] 4212 is fv 21 The second scale 4213 is fv 21 The third scale

[0127] 4220 is fv 22 The center value 4221 is fv 22 The first scale

[0128] 4222 is fv 22 The second scale 4223 is fv 22 The third scale

[0129] 4230 is fv 23 The center value 4231 is fv 23 The first scale

[0130] 4232 is fv 23 The second scale 4233 is fv 23 The third scale

[0131] 42e0 is fv 2e The center value 42e1 is fv 2e The first scale

[0132] 42e2 is fv 2eThe second scale 42e3 is fv 2e The third scale

[0133] 500 is any point

[0134] 501 is a diagram of the Euclidean space covering the probability space

[0135] 502 is the center value of probability distribution 520

[0136] 520 is the probability distribution

[0137] 503 is the first tick of the probability distribution 520

[0138] 504 is the second tick of the probability distribution 520

[0139] 505 is the third scale of the probability distribution 520

[0140] 506 is the first scale area of ​​probability distribution 520, and the probability value is p 1j (520)

[0141] 507 is the second scale area of ​​the probability distribution 520, and the probability value is p 2j (520)

[0142] 508 is the third scale area of ​​the probability distribution 520, and the probability value is p 3j (520)

[0143] 510 is the center value of probability distribution 530

[0144] 511 is the first tick of the probability distribution 530

[0145] 512 is the second tick of the probability distribution 530

[0146] 513 is the third tick of the probability distribution 530

[0147] 514 is the first scale area of ​​probability distribution 530, and the probability value is p 1j (530)

[0148] 515 is the second scale area of ​​the probability distribution 530, and the probability value is p 2j (530)

[0149] 516 is the third scale area of ​​probability distribution 530, and its probability value is p 3j (530)

[0150] 520 and 530 are probability distributions of two probability spaces in the Euclidean space 501 covering the probability space.

[0151] 500 is the projection point of any point at the junction of the two probability distributions on the straight line connecting the centers 502 and 510 of the two probability distributions.

[0152] 801 is a spatial mapping of the perceived image

[0153] 802 and 803 are two different mapping space images.

[0154] 804 is a local area of ​​image 801

[0155] 805 is a multi-probability scale self-organizing machine learning unit connected between the perceived object and the perception layer

[0156] 806 is the perception layer of the neural network

[0157] 807 is a node in the perception layer of the neural network

[0158] 808 is a machine learning unit connected between the perception layer and the neural layer with adversarial learning capabilities and multi-probability scale self-organization.

[0159] 809 is a database storage unit

[0160] 810 is the neural layer of the neural network

[0161] 811 is the node of the neural layer of the neural network corresponding to each node of the perception layer

[0162] 812 is a multi-probability scale self-organizing machine learning unit connected between the neural layer and the cerebral cortex.

[0163] 813 is a database between the neural layer and the cerebral cortex

[0164] 814 is the cerebral cortex of the neural network

[0165] 815 is the cerebral cortex node

[0166] 1101 is the original image color

[0167] 1102 is the color of the image after scanning

[0168] 1301 is the regression analysis straight line calculated using the traditional formula

[0169] 1302 and 1303 are the data between the given range and the regression analysis line.

[0170] 1304 is the regression analysis line obtained through ultra-deep adversarial learning

[0171] 1305 and 1306 are given and Fig.13 (a) Data in the same range as the regression line obtained by ultra-deep adversarial learning DETAILED DESCRIPTION

[0172] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings, but the embodiments of the present invention are illustrative rather than restrictive.

[0173] Figure 1 is a schematic diagram of multiple probability scales defined for a probability distribution.

[0174] like Figure 1 As shown: Suppose a probability distribution in the probability space is 101, there must be a multi-probability scale that can represent the probability distribution status of the multiple positions of the probability distribution. Here, when the center value of the probability distribution in a certain area is 102, the first scale value can be set to 103, corresponding to the probability distribution value in the area is 106, the second scale value can be set to 104, corresponding to the probability distribution value in the area is 107, and the third scale value can be set to 105, corresponding to the probability distribution value in the area is 108. The benchmark for quantitatively calibrating the probability distribution values ​​of multiple areas is called a multi-probability scale.

[0175] The calibration of multiple probability scales is based on normal distribution, multivariate normal distribution, lognormal distribution, exponential distribution, t distribution, F distribution, X 2 Distribution, binomial distribution, negative binomial distribution, multinomial distribution, Poisson distribution, Erlang distribution, hypergeometric distribution, geometric distribution, traffic distribution, Weibull distribution, triangular distribution, beta distribution, gamma distribution, different probability values ​​of a given complex area calculated by at least one of the probability distributions having the probability characteristics of the probability distribution.

[0176] The correlation coefficient and correlation distance in probability space can also be used as multi-probability scales.

[0177] For non-probability space, as a multi-probability scale, it can also be extended to the Euclidean space distance (EuclideanDistance) scale, Manhattan distance (Manhattan Distance) scale, Chebyshev distance (ChebyshevDistance) scale, Minkowski distance (Minkowski Distance) scale, Mahalanobis distance (MahalanobisDistance) scale, angle cosine (Cosine) scale, W distance (Wasserstein Distance) scale, KL distance (Kullback-Leibler Distance) scale, PE distance (Pearson Distance) scale.

[0178] Distances in other spaces can also be extended to multiple probability scales, such as the Jaccardsimilaritycoefficient scale, the Hamming distance scale, and the information entropy scale.

[0179] Some probability algorithms such as Bayesian analysis, Gaussian processes, or Gaussian process and Bayesian hybrid algorithms can constitute benchmark algorithms and can also be extended to the definition of multiple probability scales.

[0180] In order to simplify the description of technical features, the specific mathematical formulas of the above-mentioned probability scales are not listed one by one. All scales that can calibrate data can imitate the following multi-probability scale processing method to form a machine learning model, which all fall within the scope of the present invention.

[0181] In summary, the above-mentioned method of constructing a multi-probability scale for machine learning is a measurement method that can divide the probability distribution of data that traverses complex space, such as Euclidean space and probability space, into more than one domain.

[0182] The data that traverses the complex space refers to: at least one or more of the data that can traverse the Euclidean space; probability space; Manhattan space; Chebyshev space; Minkowski space; Mahalanobis space; and angle cosine space.

[0183] The fuzzy event probability measure refers to: establishing a fuzzy relationship between data by crossing the distance in at least one of the above-mentioned spaces; an algorithm that can establish a strict relationship between different data by integrating microscopic fuzzy information and probability information to obtain macroscopic stable information.

[0184] Figure 2 It is a self-organizing machine learning flowchart at multiple probability scales.

[0185] Here, according to the above definition of multi-probability scale self-organization, Figure 2 The processing flow of the multi-probability scale self-organizing machine learning is as follows:

[0186] S1 initialization step: It is the initialization setting step of the multi-probability scale self-organizing machine learning program. In this step, the target function D(x, y) must be input first. This data can be one-dimensional data or two-dimensional data, and any dimension is acceptable. Secondly, the initialized multi-probability scale M must be given in advance. (0) ', and the initial center value of the self-organization, or the initial eigenvalue (x0, y0) (0) There are two ways to carry out multi-probability scale self-organization. The first method is to start from the maximum scale of the multi-probability scale and self-organize to smaller scales one by one. The second method is to start from the smallest scale of the multi-probability scale and self-organize to larger scales one by one. Therefore, the initialization processing should also be divided into two methods to set the initialization parameters respectively.

[0187] The first multi-probability scale self-organization initialization method:

[0188] Initialized multiple probability scales M (0) 'With the probability scale M defined above (0) The difference is M (0) It is sufficient to give an estimate of the scale of the maximum probability range, and M (0) 'Because it is a multi-probability scale, the data of the entire probability distribution needs to be self-organized, so an estimate of the probability scale of the entire probability distribution needs to be given, that is, M (0) '>M (0) , generally about 3 times can be taken.

[0189] The second initialization method of multi-probability scale self-organization:

[0190] Initialized multiple probability scales M (0) 'With the probability scale M defined above (0) Same M (0) '=M (0) That's it.

[0191] The above two methods are (0)The selection method of the value is similar, there is no need to set it strictly, through manual prediction, with (x0, y0) (0) Centered on M (0) ' is the radius, which must include a part of the final learning result data, so the initial M (0) 'The value should be as large as possible. Initialize the probability scale M (0) 'The larger the value, the longer the calculation time. If it is too small, the correct result may not be obtained.

[0192] Regarding other initialization settings, V is used as the convergence value of the self-organization, that is, is there a gap between the result of the last organization and the result of this self-organization, that is, has the processing result of the self-organization met the requirements? If the convergence value V is too large, it may not get the correct result. If the convergence value V is too small, the calculation takes longer. The correct setting method is about 5-10% of the initial probability scale. MN is used as the maximum number of self-organization times, which can generally be set at 5-10 times to prevent the self-organization from being in an infinite dead loop state. The setting of the scale value m of the multi-probability scale, m is set to the number of scales of the multi-probability scale. For example: the multi-probability scale has three scales, which can calibrate the probability values ​​of three probability distribution areas. In this case, m=3 is sufficient. As a reflection of the number of self-organizations, n=0 is used as the current initial number of self-organizations.

[0193] S2 is the multi-probability scale self-organization step: in this step, the nth self-organization processing of the mth scale of the probability scale is performed. For the target function D(x, y), (x0, y0) m (n-1) As the self-organizing center value, the probability scale M of the mth scale m (n-1) As the radius, the radius M m (n-1) Within the objective function data d m (n) (x i ,y j )(i=1,2,…,k;j=1,2,…,l) is used as a probability distribution in the new probability space. For the new probability distribution, a new eigenvalue (x0,y0) will be generated. m (n) , and the probability scale M of the mth scale m (n) , here d m (n) (x i ,y j )∈D(x, y), n=n+1, MN=MN-1. That is, in this step, each time the objective function data d is executed m (n) (x i,y j )(i=1,2,…,k;j=1,2,…,l) migrated one step toward the direction of maximum probability.

[0194] S3 is the judgment step: it is the judgment step whether the probability scale self-organization of the mth scale is completed. In this step, it is necessary to judge whether |M m (n) -M m (n-1) |≤V? Or MN=0? If one of the conditions is met, the probability scale self-organization of the mth scale of the multi-probability scale is completed. If it is completed, it will go to step S4. If it is not completed, it will jump to step S2 and continue the operation of multi-probability scale self-organization.

[0195] S4 data preservation step: After the probability scale of the mth scale is self-organized, the characteristic value (x n ,y n ) m (n) , the mth probability scale M m (n) Save it as a learning result, m = m-1, correct the scale of the probability scale, and correct the set of target functions, that is, the range of self-organized data, to the previous multi-probability scale M m (n) and eigenvalue (x0, y0) m (n) The probability distribution formed is

[0196] D(x,y)←d m (n) (x i ,y j )(i=1,2,…,k;j=1,2,…,l),

[0197] If it is the second multi-probability scale self-organization method, since the multi-probability scale starts from the minimum scale and is set according to the normal distribution, and since the scales of the subsequent probability scales are the same as the range of the first maximum probability scale, there is no need to continue the multi-probability scale self-organization process. The interval of the first maximum probability scale is set as the interval of other scales, and m=0 is set, and then the process proceeds to the following multi-probability scale final completion judgment step S5.

[0198] S5 judgment step: is the final judgment step. In this step, it is determined whether m=0. If "yes", the multi-probability scale self-organization processing is completed and transferred to S6. If "no", it jumps to S2 to continue the self-organization processing of the next scale of the multi-probability scale.

[0199] S6 return step: is the final step of returning to the main program.

[0200] Next, we define the distance between a point in Euclidean space and the center point of the probability distribution in probability space under multiple probability scales.

[0201] Figure 3 It is a schematic diagram of the definition of distance across different spaces including probability space.

[0202] like Figure 3 As shown: 301 is a Euclidean space covering the probability space, and 302 is the center point w of a probability distribution in the probability space. j , 303 is the first scale M1 of the multi-probability scale of a probability distribution in the probability space, 304 is the second scale M2 of the multi-probability scale of a probability distribution in the probability space, 305 is the third scale M3 of the multi-probability scale of a probability distribution in the probability space, and 309 is a point v in the Euclidean space j , find w j to v j distance.

[0203] In actual pattern recognition, each eigenvalue must be learned multiple times under different conditions, and the learning results form a probability distribution. A eigenvector is composed of n eigenvalues, so it is necessary to calculate the distance between vectors. Here, suppose there is a point 302 in the set W as w j ∈W, there is a point 309 in the set V called v j ∈V, in the j=1, 2, ..., nth vector element, let the scale interval between 302 and 303 be D 1j =M 1j , belongs to w j The probability value of the probability distribution is P 1j (wj) , the scale distance between 303 and 304 is D 2j =M 2j -M 1j , belongs to w j The probability value of the probability distribution is P 2j (wj) , the scale distance between 304 and 305 is D 3j =M 3j -M 2j Belong to w j The probability value of the probability distribution is P 3j (wj) ,from Figure 3 It can be seen from v j to w j After belonging to w j The probability distribution of 3 scale areas, so m j(wj) =3, then the probability space distance between 309 and 302 is:

[0204] Formula 6

[0205]

[0206]

[0207] Here the command,

[0208]

[0209] Δ j (wj) For v j to w j In the direction of, when the data enters the probability space from the Euclidean space, the distance error between the distance represented by the Euclidean space in the probability space and the distance in the actual probability space is adjusted by Δ j (wj) This method can unify the distance in Euclidean space and the distance in probability space, and solve the definition of the strict distance relationship between machine learning data in crossing the two spaces.

[0210] In w j to v j Any set R between has a point r j ∈R 310 and w j The ambiguity of 302 is:

[0211] Formula 7

[0212]

[0213] The above formula is a membership function. When any point r j The closer ∈R is to w j ∈W, F j (wj) The closer the result is to "1", the j ∈R is farther from w j ∈W, F j (wj) The closer it is to "0", the distance involved here is a distance across Euclidean space and probability space, as can be seen from Formula 6, so F j (wj) The fuzzy value of is also the fuzzy value of the data that spans the Euclidean space and the probability space.

[0214] The definition method of the membership function is not restricted to any form, and various formulas can be defined according to human intervention. Regardless of the definition method, as long as it can reflect the fuzzy relationship between the two elements of the objective function, it belongs to the scope of the present invention.

[0215] Here, let rp j (wj) Any point r j ∈R belongs to w j ∈W probability distribution. Fuzzy information F j (wj) With probability rp j (wj) The integral of the product of (j=1, 2, ..., n) constitutes the following measure formula for the probability of fuzzy events from set R to set W.

[0216] Formula 8

[0217]

[0218] or

[0219]

[0220] The measure F of the probability of the fuzzy event defined above (w) The role of is to make full use of the microscopic trace fuzzy information and the microscopic trace probability information. After integration, the macroscopic information can produce considerable stable information, which can be used to calibrate the similarity relationship between any sets R and W. This is the best judgment basis for maximally reflecting the relationship between the two sets in information theory. In the application of pattern recognition, the set R can be regarded as the identified feature vector, and W can be regarded as the feature vector of a logged pattern. The fuzzy event probability measure values ​​of the two sets R and W can reflect the strict judgment basis for whether the identified feature vector belongs to the logged feature vector.

[0221] Summarize Figure 3 A proposed method for obtaining the distance across the Euclidean space and the probability space is characterized in that there is at least one probability space in the Euclidean space, and when crossing an area of ​​the probability space, the probability distance of the interval is related to the probability value of the area passed through.

[0222] The above-mentioned Euclidean space can be extended to: one of the following: Manhattan Space; Chebyshev Space; Minkowski Space; Mahalanobis Space; and Cosine Space.

[0223] Furthermore, the distance scale of the probability space is related to the probability distribution value passed through when passing through the probability space. It must be directional and does not meet the symmetry conditions of the ordinary distance scale. j to w j When the distance is v, the distance in the probability space is j The location reaches w j The process of changing the probability of the probability distribution during the process of moving to the position is related to the change process of the probability of the probability distribution, which is the w of the final position to be reached. j The probability distribution of the probability change process, and the starting position v j It doesn't matter, even if v j is also a probability distribution, also related to v j The probability distribution of is irrelevant.

[0224] Figure 4 This is a schematic diagram of the pattern recognition model of image recognition based on ultra-deep adversarial learning.

[0225] Problems with image recognition using ultra-deep adversarial learning: Figure 4 As shown: Given 4100 and 4200 are two feature vector data after multi-probability scale self-organization, fv 1j ∈FV1 and fv 1j ∈FV2, given a feature vector SV of 4000 identified objects, the feature element sv j ∈SV(j=1,2,…,e), the problem to be solved is to which feature vector does the identified object vector SV belong? Here, let 4001 be sv1, 4002 be sv2, 4003 be sv3,…, 400e be sv e .

[0226] Feature vector data fv of FV1 11 , the first scale is 4111, fv 11 The second scale is 4112, fv 11 The third scale is 4113, fv 11 The center value of fv is 4110. 12 The first scale is 4121, fv 12 The second scale is 4122, fv 12The third scale is 4123, fv 12 The center value of fv is 4120. 13 The first scale is 4131, fv 13 The second scale is 4132, fv 13 The third scale is 4133, fv 13 The center value of fv is 4130. 1e The first scale is 41e1, fv 1e The second scale is 41e2, fv 1e The third scale is 41e3, fv 1e The center value of is 41e0.

[0227] feature vector data fv2 21 The first scale is 4211, fv 21 The second scale is 4212, fv 21 The third scale is 4213, fv 21 The center value of fv is 4210. 22 The first scale is 4221, fv 22 The second scale is 4222, fv 22 The third scale is 4223, fv 22 The center value of fv is 4220. 23 The first scale is 4231, fv 23 The second scale is 4232, fv 23 The third scale is 4233, fv 23 The center value of fv is 4230. 2e The first scale is 42e1, fv 2e The second scale is 42e2, fv 2e The third scale is 42e3, fv 2e The center value of is 42e0.

[0228] Let the feature element sv in the feature vector SV of the identified object be j ∈SV is the feature element fv in the registered feature vector data FV1 1j The probability value in the probability distribution is sp j (fv1j) ∈SP (FV1) , in the feature vector data FV2, the feature element fv 2j The probability value in the probability distribution is sp j (fv2j) ∈SP (FV2) (j=1, 2,…, e).

[0229] Suppose sv jto fv 2j The number of multi-probability scales that the probability distribution center passes through mv 2j , and then assume sv j to fv 2j The number of probability areas that the center of the probability distribution passes through is m j (fv2j) =mv 2j +1, fv 2j The multi-probability scale interval of the probability distribution is D ij , at a spacing of D ij The domain belongs to fv 2j The probability value of the probability distribution is P ij (fv2j) (i=1,2,…,m j (fv2j) ).

[0230] According to formula 7 and formula 8, the fuzzy event probability measure of the identified object vector SV and the feature vector data FV2 can be calculated:

[0231]

[0232] Similarly, let sv j ∈SV to fv 1j The number of multi-probability scales that the probability distribution center passes through mv 1j , m j (fv1j) =mv 1j +1. The fuzzy event probability measure of the identified object vector SV and the feature vector data FV1 can be calculated:

[0233]

[0234]

[0235] The identified object vector SV belongs to the feature vector data FV2 and the adversarial formula is as follows:

[0236] Formula 9

[0237] F=F (FV2) / F (FV1)

[0238] When F>1, the identified object vector SV belongs to the feature vector data FV2, and conversely when F<1, the identified object vector SV belongs to the feature vector data FV1.

[0239] Figure 5 It is a schematic diagram of the best classification model for image recognition using ultra-deep adversarial learning.

[0240] like Figure 5As shown: 501 is a schematic diagram of the Euclidean space covering the probability space. In the Euclidean space of 501, there are probability distributions 520 and 530 of two probability spaces, 502 is the center value of the probability distribution 520, 503 is the first scale of the probability distribution 520, 504 is the second scale of the probability distribution 520, 505 is the third scale of the probability distribution 520, and 506 is the first scale area of ​​the probability distribution 520. Suppose the probability value is p 1j (520) , 507 is the second scale area of ​​the probability distribution 520, and the probability value is p 2j (520) , 508 is the third scale area of ​​the probability distribution 520, and the probability value is p 3j (520) .

[0241] 510 is the center value of probability distribution 530, 511 is the first scale of probability distribution 530, 512 is the second scale of probability distribution 530, 513 is the third scale of probability distribution 530, 514 is the first scale area of ​​probability distribution 530, and the probability value is p 1j (530) , 515 is the second scale area of ​​the probability distribution 530, and the probability value is p 2j (530) , 516 is the third scale area of ​​probability distribution 530, and its probability value is p 3j (530) .

[0242] A model of ultra-deep strong adversarial learning is: let the Euclidean space be 501. In the Euclidean space of 501, there are two probability distributions 520 and 530, where the centers of the two probability distributions can be an element in the set W, denoted by w j ∈W, and an element in the set V is represented by v j ∈V, the projection point 500 of any point at the junction of the two probability distributions on the straight line connecting the centers 502 and 510 of the two probability distributions can be expressed by r j ∈R to express, find any point 500, that is, r j ∈R belongs to which probability distribution.

[0243] First, let m j (wj) For r j Arrival j The number of regions of probability involved in the multiple probability scales passed between the centers of the probability distribution, e.g. Figure 5 is m j (wj) =2, and then let p ij (520) =pij (wj) , (i=1,2,…,m j (wj) ), find any point 500r in the set R j ∈R and 520 probability distribution center is a point w in the set W j The fuzzy relationship between

[0244]

[0245]

[0246] Here we set m j (v j ) For r j Arrival j The number of regions of probability involved in the multiple probability scales passed between the centers of the probability distribution, e.g. Figure 5 is m j (vj) =3, and then let p ij (530) =p ij (vj) , (i=1,2,…,m j (vj) ), then find any point 500r in the set R j ∈R and 530 probability distribution center is a point v in the set V j The fuzzy relationship between

[0247]

[0248] Here we consider r j ∈R belongs to the measure of 520 probability distribution, its probability measure value should be "r j ∈R has a probability value P in the 520 probability distribution j (Wj) Similarly, when considering r j When ∈R belongs to the measure of 530 probability distribution, its probability measure value should be r j ∈R has a probability value P in the 530 probability distribution j (Vj) .

[0249] By referring to Formula 7, Formula 8 and Formula 9, we can get the formula to which set any point set R belongs:

[0250] Formula 9'

[0251]

[0252] Here, Formula 9' is different from Formula 9 above in that adversarial learning is achieved through microscopic fuzzy information and probability information between feature elements, which is more effective than macroscopic adversarial learning. The results of microscopic adversarial learning are integrated to form the fuzzy event probability measure after adversarial learning, which reflects the relationship between the two macroscopic sets more accurately. Therefore, if F ≥ 100, any point set R belongs to set W, otherwise it belongs to set V.

[0253] Fix the center values ​​of the two probability distributions to w and v, and then set the above formula w j =w,v j =v, any point r in the set R of all points at the junction of two probability distributions j ∈R, (j=1, 2, ..., n) are all calculated by the above method, and F j (w) and F j (v) The confrontation result F of (j=1, 2, ..., n) j =(F j (w) / F j (v) )×100, the two data with probability distribution can be optimally classified.

[0254] and Figure 4 The difference of the introduced ultra-deep adversarial learning is: Figure 4 The pattern recognition model of image recognition based on ultra-deep adversarial learning is to solve the optimal pattern recognition of which feature learning vector data a set of feature recognition vectors belongs to. Figure 5 It is to solve the problem of how to optimize the classification of points in the connection area of ​​two probability distributions.

[0255] From the above, we can know that the probability measurement value of a fuzzy event crossing different spaces is related to the distance of crossing two or more different spaces, and is composed of microscopic fuzzy information and microscopic probability information. The closer the distance, the greater the fuzzy relationship; conversely, the greater the distance, the more distant the fuzzy relationship. The above microscopic probability information refers to: the probability value of one data in the probability distribution of the other data in two data crossing different spaces.

[0256] Here, we only take two probability distributions as an example to introduce the model construction method of image recognition using ultra-deep adversarial learning. In actual applications, there may be 3 probability distributions, or 4, 5...n probability distributions that are optimally classified with each other through ultra-deep adversarial learning.

[0257] Figure 6 It is the optimal classification flowchart for image recognition with ultra-deep adversarial learning.

[0258] like Figure 6Shown: Combined Figure 5 Ultra-deep adversarial learning can be achieved by the following steps.

[0259] S1 is the initialization step. In this step, you can refer to Figure 2 The initialization content of multi-probability scale self-organization is first input into g sets W h The data w in jh ∈W h (h = 1, 2, ... g), and the set V h Each data v in jh ∈V h , set the number of scales of the multi-scale probability, for example m = 3, set the number count value PN of the probability distribution = 1, set the data storage space, and other contents that need to be processed in the initialization step.

[0260] S2 is a multi-probability scale self-organization step, refer to Figure 2 The second step S2 of the multi-probability scale self-organizing machine learning flowchart is to h The data w in jh ∈W h , and the set V h Each data v in jh ∈V h , to conduct a probabilistic scale self-organizing machine learning.

[0261] S3 is a judgment step, whether the self-organization of one probability scale in S2 is completed? If "no", it jumps back to the self-organization step of multiple probability scales in S2, and if "yes", it goes to the data storage step S4.

[0262] S4 is the data preservation step, which converts S2 into a probability scale self-organized w j ∈W and v j ∈V, the maximum probability center value of each element, the scale position of multiple probability scales, the probability value between each scale, etc. After the self-organization of a probability scale is completed, each data r in the dot position set R of the adjacent two probability distributions under this scale j ∈R(j=1,2,…,n) is found and saved.

[0263] S5 is the probability distribution judgment step. Is the multi-probability scale self-organizing machine learning completed? If "no", the number of probability distributions PN = PN + 1, jump to S2 multi-probability scale self-organizing step to continue calculating the new probability distribution; if "yes", go to the following S6 ultra-deep adversarial learning step.

[0264] S6 is the ultra-deep adversarial learning step. In this step, all the points in the connection area of ​​each distribution are first obtained, that is, each data r in the set R j∈R(j=1,2,…,n), connect the center position “w” and the value of “v” between the two connected probability distributions taken from the S4 database into a straight line, and then connect any point r in the connection area of ​​the two probability distributions j ∈R is projected onto this line, and r is calculated according to Formula 7 j ∈R and the fuzzy relationship F between the central values ​​of the two probability distributions "w" and "v" respectively i (w) With F j (v) , and then from formula 8, formula 9, and assuming P j (w) For r j ∈R probability value on the probability distribution 520, let P j (v) For r j ∈R is the probability value in the probability distribution 530, then r j ∈R The measure of the probability of microscopic blurry events for image recognition using ultra-deep adversarial learning between two probability distributions is:

[0265] F j ={(F j (w) ×P j (w) ) / (F j (v) P j (v) )}100

[0266] F j ≥100 then r j ∈R belongs to probability distribution 520, otherwise it belongs to another probability distribution 530.

[0267] To simply and clearly demonstrate the effectiveness of the ultra-deep adversarial learning model with specific numbers, a set of data is given below: Let the Euclidean distance between w and v be 70mm, the scale interval of the 520 probability distribution be 20mm, the scale interval of the 530 probability distribution be 14mm, and r j ∈R is 39mm from w and 31mm from v, P j (w) is 27.2, P j (v) is 4.2, the results of the four algorithms are shown in Table 1.

[0268] Table 1

[0269] Learning Algorithms Same w distance Same v distance Same w relationship Same v relationship Phase Difference Judgment results Euclidean distance 39mm 31mm 1.25 times Belong to 530 Probability distance 21mm 13.7mm 1.53 times Belong to 530 Fuzzy Relationship 58 76 1.31 times Belong to 530 Measure Relationship 39 3 13 times Belong to 520

[0270] From the results in Table 1, it can be seen that the effectiveness of introducing the fuzzy event probability measurement theory can be proven.

[0271] The above-mentioned measurement theory of the probability of fuzzy events in ultra-deep adversarial learning is implemented based on the multi-probability scale defined by the probability distribution of the probability space. In practical applications, the probability scale can also be obtained by repeatedly learning the feature vector data of the same identified image under different conditions and deriving the scale of the multi-probability scale according to the actual probability distribution of the feature elements.

[0272] Figure 7 It is a flowchart of the scale of multiple probability scales for obtaining the actual probability distribution through learning.

[0273] like Figure 7 As shown: the processing of obtaining the scale of multiple probability scales of the actual probability distribution through learning is divided into four steps, including initialization step S1, recording the data of each characteristic element of the characteristic vector step S2, calculating the distribution range and probability of each characteristic element step S3, and returning to the main program step S4.

[0274] In the initialization step S1, the pointer of the feature vector data, the data scale, the number of iterations, and the number of scales of the multi-probability scale are mainly set. Because the probability distribution calculation is for actual data, in order to calculate the multi-probability scale between the center points of the two probability distributions with higher accuracy, as many scales as possible should be set.

[0275] In step S2 of recording the data of each characteristic element of the characteristic vector, a large number of characteristic vectors generated by multiple learning of the identified object are read. The data can be saved first and processed offline, or directly processed online.

[0276] In step S3 of calculating the distribution range and probability of each feature element, the central value of the maximum probability of the data is first calculated through the self-organizing algorithm of the probability scale, and then the Euclidean distribution distance of each feature element in all feature vectors is calculated with this central value of the maximum probability as the center. According to the number of scales of the multi-probability scale set by initialization and the probability value within the scale range, the scale range in which each feature element of the feature vector is distributed is calculated respectively, and the ratio of the number of each feature element within a certain scale range to the total number of learning times is used as the probability value within the scale range to obtain the probability value of each feature element within the scale range. It is also possible to use a given probability value to find the range of the ratio of the number of feature elements corresponding to this probability.

[0277] In the return step S4, the scale range in which each characteristic element of the characteristic vector is distributed and the probability value within the scale range are saved so as to be used for calculating the measure of the probability of the fuzzy event, and then the main program is returned.

[0278] The schematic diagram of multiple probability scales for obtaining the actual probability distribution through learning can be referred to Figure 1 , but the scale can be increased a lot, which can further improve the accuracy of fuzzy event probability measurement.

[0279] Summarize Figure 7 A proposed method for obtaining a multi-probability scale calibration by learning the actual probability distribution is to repeatedly learn the feature vector to obtain the distribution position of each feature element data belonging to the multiple feature vectors, and then calculate the data proportion in a given field according to the actual distribution of each feature element, or calculate the distribution area of ​​the given data distribution proportion, to obtain the scale position of the multi-probability scale, or the probability value at the scale position.

[0280] That is, based on the ratio of the number of characteristic element data distributed in a given field to the total number of characteristic element data, the area that needs to be occupied is calculated under the condition of the ratio of the number of given characteristic elements to the total number of data.

[0281] Figure 8 It is a schematic diagram of a very deep adversarial learning neural network.

[0282] Taking image recognition as an example, we will introduce the construction principle of image recognition using ultra-deep adversarial learning. Figure 8 As shown: 801 is a spatial mapping of the perceived image. Super deep learning focuses on the perceived image. Through various image transformations, such as frequency space images, color space images, energy space images, edge space images, etc., 802 and 803 are two different mapping space images. 804 is a local area of ​​image 801. Here, when considering mobile phone recognition, it can be correctly recognized at any angle of 360 degrees. The set area segmentation method is based on the ring. The size of this area should be determined according to the needs of the application. If the range is too large, the calculation speed will be fast, but it will affect the recognition accuracy. On the contrary, if the range is too small, the recognition accuracy will be improved, but the calculation speed will be slow. It can be considered that the number of pixels or area of ​​each area is consistent, so that the amount of input information can be more balanced.

[0283] 805 is a self-organizing learning unit connected between the perceived object and the perception layer, and is a multi-probability scale self-organizing machine learning unit. This learning unit is mainly responsible for the learning and analysis of the perceived target information. 806 The perception layer of the neural network mainly plays the role of inputting the target function information. Since the perception layer is connected to a large number of machine learning units, it has an analysis function for the target function information, and can extract the characteristic value of the target function information under the condition of maximum probability. It also has a deep mining function, which can mine the deep information of the target function of the complex system problem through the method of spatial mapping. It also has a human intervention function, which can process the form of the target function through human intervention according to the needs of the application. 807 is a node of the perception layer of the neural network. The deepening of image recognition of ultra-deep adversarial learning is achieved by increasing the amount of information of the target function, that is, increasing the number of spatial transformations, and increasing the number of multi-probability scale self-organizing machine learning units, as well as the number of nodes in the perception layer.

[0284] Since the multi-probability scale self-organizing machine learning connected between the perceived target and the perception layer can autonomously migrate in the direction of maximum probability and automatically acquire the target, in order to solve the problem of serious image position offset when performing online recognition of video images, a large amount of learning is required using a fixed position method. It is proposed that according to the structural features of the recognized image, each multi-probability scale self-organizing machine learning unit automatically captures the features of a part of the recognized image, and can automatically track the features of this part of the image even if this part of the feature moves a lot. For example, if the recognized object is a face, different multi-probability scale self-organizing machine learning units can be used to track the contour position of the person, the position of the person's eyes, the position of the person's mouth, the position of the person's nose, etc. In this way, no matter where the face is on the image, face recognition can be performed correctly, and the object to be recognized can be found from a large number of video programs. This is an important indicator that can reflect the level of machine learning.

[0285] 808 is a multi-probability scale self-organizing machine learning unit with adversarial learning capability connected between the perception layer and the neural layer. It mainly plays the role of deep mining of information in the perception layer through adversarial learning. The learning data of the machine learning unit 808 depends on historical storage data, and its learning results must also be preserved as historical data. Therefore, a database storage unit 809 is set.

[0286] 810 is a neural layer of the neural network, which is the initial decision-making layer that provides important reference opinions for the final judgment of the cerebral cortex 814 based on the data of the adversarial results generated after the machine learning unit 808 learns. 811 is a node of the neural layer of the neural network corresponding to each node of the perception layer.

[0287] 812 is a multi-probability scale self-organizing machine learning unit connected between the neural layer 810 and the cerebral cortex 814, which mainly provides the final learning data for the final decision of the cerebral cortex 814. Its important function is to use machine learning to determine the value of a large amount of characteristic elements input by the perception layer to achieve the objective function, so that the cerebral cortex can make the final correct judgment. The maximum probability scale generated by the machine learning results of the probability scale self-organization of 812, or the proportional value of the adversarial results, are all results with the properties of the threshold of the neural layer as a whole, which serves as the final judgment signal of the cerebral cortex 814. 813 is a database storage unit for the machine learning results of 812 between the neural layer and the cerebral cortex, which is responsible for providing and storing the data required for machine learning.

[0288] Summarize Figure 8 A method for constructing a new type of artificial intelligence neural network is composed of a perception layer, a neural layer and a cerebral cortex, wherein the perception layer is connected to the perceived object through a plurality of unsupervised machine learning units with direct input of small data; the input perception object is learned with maximum probability information. The neural layer and the perception layer are connected with a plurality of unsupervised machine learning units with direct input of small data; adversarial learning of fuzzy event probability measurement is performed. The cerebral cortex and the neural layer are connected with unsupervised machine learning units with direct input of small data; adversarial decision-making with maximum trust is learned.

[0289] The perceived objects here are those that are spatially mapped, including images, facial information, sound, emotional information, text information, industrial, agricultural or commercial site information, Internet of Things information, financial and economic forecast information, and any other digital information that requires artificial intelligence processing.

[0290] The perceived objects of the complex system are mapped to multiple spaces that can carry the information of the perceived objects, and the purpose of obtaining the deep information of the complex system is achieved by perceiving the mapped spatial information.

[0291] The above maximum probability information is obtained through unsupervised learning with direct input of small data; it goes beyond the information of the perceived object obtained by traditional statistics, that is, the numerical information of the perceived object that is closest to the mother body.

[0292] The adversarial learning of the above-mentioned fuzzy event probability measure means: for a perception data, it is necessary to learn the relationship between the fuzzy event probability measure belonging to one probability distribution data among the two closest probability distribution data; at the same time, the adversarial learning of the relationship between the fuzzy event probability measure that does not belong to this probability distribution data is carried out, and the learning method belonging to that probability distribution data is obtained.

[0293] The above-mentioned adversarial learning is based on the characteristic elements belonging to the maximum probability confidence.

[0294] Adding a multi-probability scale self-organizing machine learning between the brain nerve layer and the cerebral cortex is to improve the accuracy of image recognition of ultra-deep adversarial learning. It is to learn the actual recognition success probability of each feature element of the feature vector and log it as an important information for recognition. Through multi-probability scale self-organizing machine learning, the feature element information with the greatest reliability can be obtained, and the data of feature elements with high reliability can be used. Such ultra-deep adversarial learning can achieve the best recognition effect.

[0295] Fig. 9 It is a flowchart of a very deep adversarial learning neural network.

[0296] Here is the reference Figure 4 and Figure 8 The example of image recognition specifically introduces the process of ultra-deep adversarial learning neural network. Fig. 9 As shown: The flowchart of the ultra-deep adversarial learning neural network is implemented by the following 16 steps:

[0297] Step S1 of inputting the perceived image: In this step, the image to be recognized is input, but not only image information, but also sound information, predicted data information, fast Fourier transform data information and other data that require artificial intelligence perception.

[0298] Spatial mapping step S2: The identified image is transformed into an image that can carry different information of the original image through image processing, such as a color space image, b color space image, edge space image, energy space image, frequency space image, Hough space image, etc. These transformed images are in the form of image or data that can reflect some implicit information features of the identified image, and can also be used as specific information carriers of the identified image and added to the identification object. Other types of data can be spatially mapped in the same way.

[0299] Image region segmentation step S3: In this step, the image can be directly segmented into several rectangular regions according to the traditional method. In order to consider that the mobile phone can realize correct recognition at any angle of 360 degrees, the recognized image can also be segmented into several rectangular regions according to the traditional method. Figure 8In the annular segmentation method shown, each area is an annular area, so there is no need to consider the direction of the captured image. As for the number of segmented areas and the number of spatially mapped images, since the multi-probability scale self-organization has the characteristic of extracting the eigenvalues ​​with the maximum probability for a larger range of images, the present invention recommends increasing the number of spatial mappings as much as possible, for example, spatial mapping can be performed within a range of 4 to 10, and reducing the number of image segmentation areas as much as possible, for example, image segmentation can be performed within 9 to 100 areas, so that better recognition effects can be obtained with the same number of nodes in the perception layer and the same processing time.

[0300] Read an image of a region Step S4: Refer to Figure 8 The neural network diagram of image recognition of ultra-deep adversarial learning. The following steps can be in the form of vertical processing and horizontal processing. The vertical processing form is: starting from an image area, after processing the content of a node in the perception layer, and then processing a node in the neural layer, until all image areas are completed. The horizontal processing form is: after completing the content processing between all image areas and all nodes in the perception layer, then processing the content between all perception layer nodes and all neural layer nodes. Here, taking the vertical form processing as an example, the model processing flow of the ultra-deep adversarial learning neural network is introduced. In this step, select the processing of an image area first.

[0301] Multi-probability scale self-organization step S5: This is a step of directly performing machine learning on the input image information. For an image region, the eigenvalue with the highest probability can be obtained by eliminating the false and retaining the true. After all the eigenvalues ​​are obtained, a eigenvector that can carry the overall information of the identified image can be formed by all the eigenvalues. Its purpose is to use the combination of the structural information of the image to express the overall information of the image. Since only one eigenvalue with the highest probability is solved, the ordinary probability scale self-organization algorithm can be used in this step.

[0302] Inputting a perception layer node step S6: inputting the maximum probability feature value obtained in the above multi-probability scale self-organization step S5 into a node of the perception layer.

[0303] Read historical data step S7: Ultra-deep adversarial learning adopts the form of online learning, that is, image learning and image recognition are completed in one system. Figure 8When the data in the 809 database is less than the set value z, it is in the learning state. Learning is performed each time an image is input, and the recognition result data is stored in the database until the data reaches z. During normal recognition, when the data of the recognized image is of "value" to the improvement of recognition accuracy in the future, this online learning data will also be used to improve the feature vector data logged in the database. In this way, for example, when recognizing a person, the person can adaptively recognize the image of the person due to age or climate change, body changes, such as wearing glasses, etc.

[0304] For the convenience of the following description, Figure 4 The image recognition example involves the registered Figure 8 The feature vector in the 809 database can be expressed by the following formula:

[0305] Formula 10

[0306]

[0307] The scale of the multi-probability scale that can represent the probability distribution information of each feature element in the feature vector of formula 10 can also be expressed by the following formula:

[0308] Formula 11

[0309] M=(M 1kj ,M 2kj ,…,M mkj )k=1,2,…,z; j=1,2,…,e

[0310] The Euclidean distance between the scales of the multiple probability scales that can represent the probability distribution information of each feature element in the feature vector of Formula 10 can also be expressed by the following formula:

[0311] Formula 12

[0312] D=(D 1kj , D 2kj ,…,D mkj )k=1,2,…,z; j=1,2,…,e

[0313] The probability value corresponding to the spacing of the Euclidean distance between the scales of the multi-probability scales that can represent the probability distribution information of each feature element in the feature vector of Formula 10 can also be expressed by the following formula:

[0314] Formula 13

[0315] P=(P 1kj , P 2kj ,…,P mkj )k=1,2,…,z; j=1,2,…,e

[0316] Identified object vector:

[0317] Formula 14

[0318] SV=(sv1,sv2,...,sv e )

[0319] The data of the above formulas 10, 11, 12, 13 and 14 are stored in Figure 8 809 in the database.

[0320] Multi-probability scale self-organization step S8: This is the most important machine learning process between the perception layer and the neural layer, and it runs through the core theory of image recognition of ultra-deep adversarial learning. Traditional image recognition is based only on the distance scale of Euclidean space. By calculating the Euclidean distance between the feature vector of the recognized image in formula 14 and the feature vectors of each login in formula 10, the image corresponding to the feature vector with the closest Euclidean distance is used as the recognition result. However, in the image recognition process, since the image data captured by the camera is greatly affected by environmental factors, the obtained feature vector data is randomly distributed. Therefore, only using the distance method of Euclidean space, the improvement of recognition accuracy is subject to certain limitations.

[0321] Ultra-deep adversarial learning uses multi-probability scale self-organization to obtain the probability distribution of the characteristic values ​​of each characteristic element of the characteristic vector of multiple recognition learning results of the same image, and obtains the exact values ​​of different probability values ​​in different scale ranges shown in formulas 11 to 13. At the same time, through the strict definition of probability space distance, the characteristic value sv of each characteristic element of the characteristic vector of the recognition object image in formula 14 is obtained. j (j=1, 2, ..., e), is a determined value of the distance between the eigenvalues ​​of the eigenvalues ​​of the eigenvectors logged in formula 10, crossing the Euclidean space and the probability space.

[0322] Fuzzy event probability measure calculation step S9: In this step, by using the strict definition of the distance across the Euclidean space and the probability space, the fuzzy relationship between the eigenvalues ​​of the object image identified by formula 14 and the eigenvalues ​​of each eigenvector logged in formula 10 can be obtained:

[0323] Formula 15

[0324]

[0325] here

[0326]

[0327] Here V maxIt is the eigenvalues ​​sv of the eigenvector of the recognition target image in the set formula 14 j The maximum distance between the eigenvalues ​​of each eigenvector in Formula 10 through the Euclidean space and the probability space. m is the jth eigenvalue sv of the eigenvector of the recognition object image in Formula 14. j There are several multi-probability scales between the jth eigenvalue of a eigenvector in formula 10, D ikj With P ikj It is sv j The intervals of each multi-probability scale traversed between the j-th eigenvalue of a certain eigenvector and the probability value on this interval.

[0328] According to the definition of the measure of fuzzy event probability, each eigenvalue sv of the image to be identified j The probability value sp in the interval of the nth scale in the probability distribution of the characteristic element of each characteristic vector in formula 10 j , and the product of the fuzzy relationship between the feature elements of each registered feature vector, we can get a micro-measurement value of the feature element of a certain registered feature vector of the identified image:

[0329] Formula 16

[0330]

[0331] here:

[0332] sm kj (fvkj) =sf kj (fvkj) ·sp j k=1,2,…,z; j=1,2,…,e

[0333] or

[0334]

[0335] In fact, all the feature elements in the feature vector sv of the image to be recognized j(j=1, 2, ..., e) and the measure between the feature elements in each feature vector in the FV determinant of formula 10 are integrated, and the measure value of the fuzzy event probability obtained can be used as a measure of the strict relationship between the feature vector of a certain recognized image and the registered feature vectors. By selecting the image corresponding to the registered feature vector with the closest relationship between the fuzzy event probability measure between the feature vectors in the FV determinant, the result of the recognized image can be obtained with high precision. This method not only takes into account the probability distribution information of each feature element in all feature vectors, as well as the strict distance relationship across different spaces including the probability space, but also takes into account the measure of the fuzzy event probability, integrates various fuzzy information and probability information at the micro level to obtain stable and valuable information at the macro level, and finally obtains the recognition result after confrontation with two different registered feature vectors. This is an optimal image recognition machine learning model based on the theory of fuzzy event probability measure. Therefore, the image corresponding to the feature vector with the closest relationship between the feature vector of the recognition object and the registered fuzzy event probability is provided to the cerebral cortex as the pre-selected result between the perception layer and the neural layer of the neural network for the final decision.

[0336] Combined with the biological brain neuron theory, because the brain is a probabilistic model, there needs to be a threshold for neural triggering. Only by using the probability scale of the maximum probability as the threshold for neural triggering can it conform to the brain's mechanism. The probability measure of fuzzy events is a phase space and cannot be used as the final judgment result. In this step, the present invention proposes a new neural network that uses the pre-selected results between the perception layer and the neural layer to make the final decision.

[0337] The specific approach is: the feature element sv of the initially selected recognition object j The eigenvalue fv of the characteristic element obtained between the eigenvector FVk that is most closely related to the registered fuzzy event probability kj , the corresponding microscopic fuzzy event probability measure value sm kj (fvkj) , serving as the final decision-making basis for multi-probability scale self-organizing machine learning between the neural layer and the cerebral cortex.

[0338] exist Figure 8 The multi-probability scale self-organizing unit of 808 also undertakes the feature element sv of the previous recognition object image through multi-probability scale self-organizing machine learning. j ', the registered feature vector fv corresponding to the recognized result image kj ', re-perform machine learning, find the new maximum probability center value, the scale value of the multi-probability scale, and re-log these results.

[0339] Enter the database step S 10:This refers to Figure 8 The step of logging in to the 809 database includes three types of data logging and processing. The first one is to log in through the results of online probability scale self-organization machine learning, such as formulas 10, 11, 12 and 13.

[0340] The second data is the feature element sv of the identified image obtained in the previous step j The eigenvalues ​​fv of each eigenvalue obtained between the eigenvectors that are most closely related to the probability of the logged fuzzy event kj , the corresponding microscopic fuzzy event probability measure value is sm in formula 16 kj (fvkj) , the total macroscopic measurement value:

[0341] Formula 17

[0342]

[0343] and the corresponding image to be logged in.

[0344] Step S to determine whether the entire area is completed 12 : If the entire area has not been completed, modify the area pointer, modify the perception layer pointer, and the neural layer pointer, and then jump to step S4 of reading the image of an area. If the entire area has been completed, proceed to the next step.

[0345] Multi-probability scale self-organization step S 13 :This is a machine learning unit between the neural layer and the cerebral cortex, which is mainly responsible for the following machine learning functions:

[0346] 1. Evaluation of the reliability of the measured characteristic values ​​of characteristic elements.

[0347] The measured eigenvalue sm of the characteristic elements of each eigenvector of the FV of the discriminant neural layer input formula 10 kj The judgment of whether it is reliable can enable the cerebral cortex to make judgments on reliable data, thereby reducing the risk of incorrect judgments due to the use of unreliable data.

[0348] The specific method for calculating the probability value of the reliability is: based on the usual machine learning performed many times, that is, assuming that the success rate of the jth feature element in the feature vector of the kth image that has been registered corresponding to the kth image to be recognized is CD kj , and the error recognition rate when it is not the image recognition is ED kj , the trust value can be expressed by the following formula 18:

[0349] Formula 18

[0350]

[0351] 2. Obtaining the measured characteristic values ​​of the characteristic elements with maximum reliability

[0352] Through the multi-probability scale self-organized machine learning in this step, the confidence vector R of the k-th feature vector is kj By self-organizing learning, we can find the maximum probability trust vector R of the kth eigenvector in the scale of the maximum probability scale. kl (l = 1, 2, ..., a), according to the characteristic element sm of the measure corresponding to k and i kl (fvki) (l=1, 2, ..., a) is the characteristic element of the maximum probability reliability measure.

[0353] The maximum probability confidence level can also be determined based on whether the probability distribution of the j-th feature element in the feature vector of the k-th image that has been logged in is concentrated or dispersed. The range values ​​of the first scale of all feature elements can be self-organized on multiple probability scales to obtain the first scale range value with the maximum probability. The feature element values ​​within the maximum first scale range can be used as the feature elements with the highest reliability measurement.

[0354] 3. Composition of adversarial vectors

[0355] Among the k measured feature vectors, find the two closest feature vectors FVk in the fuzzy event probability measure value between the identified vector SV and each feature vector FVk according to formula 17. max1 and FV max2 The fuzzy event probability measures corresponding to these two vectors and SV are SM max1 and SM max2 , using their various elements sm max1 ∈SM max1 and sm max2l ∈SM max2 (l=1, 2,…, a) is used as the final decision basis.

[0356] Step S for storing or reading data 14 :In this step, we first calculate the trust value R of the measured characteristic value of the characteristic element kj Login, and then the vector R of the maximum probability confidence kl (l = 1, 2, ..., a), and sm max1 ∈SM max1 and sm max2l ∈SM max2 (l=1, 2, ..., a) Login.

[0357] Input cortical decision step S 15:In this step, the characteristic element sm of the maximum probability reliability measure of the recognized image obtained by the neural layer max1 ∈SM max1 and sm max2l ∈SM max2 (l=1, 2, ..., a), in the cortical step, decisions can be made according to the following formula for image recognition using ultra-deep adversarial learning.

[0358] Formula 19

[0359]

[0360] Here, the threshold T>1 is set. According to formula 19, when AL>T, the neural layer excites the cerebral cortex to judge that the recognized image belongs to SM. max1 The image corresponding to the probability measure value of the fuzzy event is rejected, otherwise the recognition result is rejected.

[0361] Return to step S 16 :The recognition process ends here and returns to the main program.

[0362] So far, ultra-deep adversarial learning is a machine learning model for each feature element of the feature vector. The following provides a model that can perform ultra-deep strong adversarial learning on feature vectors.

[0363] Here we only use the imaging of optical logos in space recognized by mobile phones, and transform the spatial images into codes ITC (Image To Code) through multi-probability scale self-organized ultra-deep adversarial learning. Since the light source of the scanner is directional, images at a certain angle in space can be copied. Therefore, it is necessary to solve how to separate the feature vector set of images that can be copied by the scanner and the feature vector set of images that cannot be copied from various images of optical logos in space through ultra-deep strong adversarial learning. In the feature vector set of images that cannot be copied, find a set of feature vector values ​​with the smallest fuzzy event probability measure relationship with the feature vector set of copiable images as the anti-counterfeiting authentication code for public recognition of mobile phones, so as to achieve the application purpose of public authenticity recognition through mobile phones.

[0364] Fig.10 It is a flowchart of ultra-deep strong adversarial learning for mobile phone anti-counterfeiting identification.

[0365] like Fig.10 As shown: in order to obtain the feature vector set of the image that can be reproduced by the scanner, it is necessary to include the following four steps: step S1 of reading the scanned image, step S2 of ultra-deep adversarial learning, step S3 of saving the feature vector data of the reproducible image, and step S4 of judging whether the scanned image is completed:

[0366] In step S1 of reading the scanned image, the optical mark is scanned in different directions by a scanner, and a spatial image of the optical mark in one direction is read to proceed to the next step.

[0367] Ultra-deep adversarial learning step S2, in this step, mainly for the image read in the previous step, by referring to Figure 4 , Figure 8 ,or Fig. 9 Ultra-deep adversarial learning is performed on the content to obtain a feature vector with maximum probability for the scanned image and various scale values ​​of multiple probability scales for the distribution of each characteristic element of the feature vector.

[0368] Step S3 of saving the reproducible image feature vector data is to log the result of the previous step into the database.

[0369] Determine whether the scanned image has completed step S4, and determine whether it is necessary to input the scanned image again. Here, it is based on the fact that the previously scanned image will not generate a new feature vector. If it is determined that the scanned image needs to be input, then go to step S1, change the direction of the optical mark to continue scanning to obtain a new scanned image. If it is determined that it is not necessary to continue scanning the image, then go to the next step.

[0370] In terms of obtaining the feature vector set of the mobile phone recognition image, it is necessary to include step S5 of reading the mobile phone recognition image, step S6 of ultra-deep adversarial learning, step S7 of saving the feature vector data of the mobile phone recognition image, and step S8 of judging whether the mobile phone recognition image is completed.

[0371] Read the mobile phone recognition image step S5, read the arbitrary image of the optical mark in space through the mobile phone, and go to the next step.

[0372] Ultra-deep adversarial learning step S6, in this step, mainly for the image read by the mobile phone in the previous step, by referring to Figure 4 , Figure 8 ,or Fig. 9 Ultra-deep adversarial learning is performed on the content to obtain a feature vector with the maximum probability of the image recognized by the mobile phone, as well as the scale values ​​of multiple probability scales for the distribution of the characteristic elements of the feature vector.

[0373] In step S7 of saving the feature vector data of the mobile phone recognized image, a feature vector of the maximum probability of the image recognized by the mobile phone in the previous step and each scale value of the multi-probability scale for the distribution of the feature elements of the feature vector are logged into the database.

[0374] Determine whether the mobile phone recognition image has completed step S8, and determine whether it is necessary to perform mobile phone recognition image again. Here, it is based on the fact that the past mobile phone recognition image will not generate a new feature vector. If it is determined that the mobile phone recognition image needs to be input, then go to step S5, and the mobile phone recognition image operation obtains a new mobile phone recognition image. If it is determined that there is no need to continue scanning the image, then go to the next step.

[0375] In obtaining the anti-counterfeiting authentication code for mass mobile phone identification, it is necessary to obtain the anti-counterfeiting verification code step S9 by ultra-deep strong adversarial learning. 10 , and return to step S 11 , consisting of three steps:

[0376] Ultra-deep strong adversarial learning step S9, in this step refer to Figure 4 , Figure 8 ,or Fig. 9 The method of image recognition based on ultra-deep adversarial learning determines whether the feature vector stored in the mobile phone recognition image belongs to the one stored in the reproducible image feature vector set through fuzzy event probability measurement relationship. When it is determined that it belongs to any one in the reproducible image feature vector set, this feature vector is removed from the feature vector set stored in the mobile phone recognition image until any feature vector in the reproducible image feature vector set does not exist.

[0377] Get the anti-counterfeiting verification code step S 10 In this step, the mobile phone is used to identify the optical mark, referring to Figure 4 , Figure 8 ,or Fig. 9 Then, through ultra-deep adversarial learning, the feature value of the image recognized by the mobile phone is compared with the feature vector stored in the image recognized by the mobile phone, and the feature vector of a spatial image with the smallest tilt angle of the mobile phone is found as the mobile phone recognition anti-counterfeiting code.

[0378] Return to step S 11 The above steps obtain the anti-counterfeiting code that can be recognized by the public mobile phone and return to the main program.

[0379] Fig.11 It is a schematic diagram of the characteristics of printed image color space and electronic image color space.

[0380] like Fig.11 As shown in the figure: Printing color is a color space composed of CMYK, and electronic images are a color space composed of RGB. The color spaces of electronic images and printed images are mostly overlapped, but there are also some parts that do not overlap. Due to such characteristics, Fig.11 As shown: 1101 is the original image color, which becomes 1102 after scanning. Therefore, the printed image copied by the scanner will inevitably be different from the original image.

[0381] Furthermore, tiny dots, such as those below 0.04 mm, will inevitably lose a portion of the dots after being scanned by a scanner. That is to say, the grayscale of each color of the printed image will be lost to varying degrees after being scanned by a scanner.

[0382] However, after some compensation, the printed image copied after scanning by the scanner can be closer to the original printed image. In particular, it is difficult for the naked eye to distinguish which is the original image and which is the copied image. This is a common problem that cannot be solved in the current society in terms of product anti-counterfeiting.

[0383] To address this problem, we can introduce ultra-deep strong adversarial learning. By separately training the original image and the copied image, we can distinguish the original image from the copied image. We can also obtain a high-precision anti-counterfeiting identification code for popular mobile phones that can distinguish between the original image and the copied image, thereby solving the problem of anti-counterfeiting identification for popular mobile phones.

[0384] For specific methods, please refer to Fig.10 In steps S1-S4, the printed images copied by different methods are scanned to obtain the feature vector of the copied printed images and the set of scales of multiple probability scales. In steps S5-S8, the original printed images are photographed by the mobile phone in different environments to obtain the feature vector of the original images photographed by the mobile phone and the set of scales of multiple probability scales. Then, in steps S9-S 11 In the step of ultra-deep strong adversarial learning, the original printed image is photographed by a mobile phone in different environments, and the feature vector of the original image photographed by the mobile phone is obtained, and the feature vector of the printed image that does not belong to the copied image and the feature vector in the set of multi-probability scale scales are used as anti-counterfeiting codes, and the feature vector of the copied printed image and the feature vector in the set of multi-probability scale scales are logged as copy codes. The above method is used for product identification processing, which can achieve the effect of mass mobile phone recognition of product authenticity.

[0385] Similarly, a small anti-counterfeiting identification area can be set on the product logo, and an optical mark can be set on the anti-counterfeiting identification area, or Fig.11 The anti-counterfeiting printing area shown is used to realize the authenticity identification of the mass-market mobile phone according to the above method.

[0386] Modeled after Fig.10The principle of optical identification of mass mobile phones can also be extended to the application of OVI color-changing film. Color-changing film and color-changing ink are made by vacuum coating process, and the interference effect of a certain spectrum is achieved by controlling the thickness of the film. This film or ink can produce different color changes when the angle is changed by 30-60 degrees, that is, it can produce images of different colors in space, which is similar to optical identification. Fig.10 The method realizes the anti-counterfeiting structure of mobile phones for public recognition of the authenticity of product logos.

[0387] use Fig.10 The method can also be introduced into the micro-lens array film to achieve the effect of mass mobile phone authenticity identification. Specifically, the micro-lens array can project the representation image to a certain angle. This angle can be determined by the design of the diffraction angle of the micro-lens array. At the same time, the image projected into space can also be determined by the printed image corresponding to the micro-lens array. Therefore, compared with optical labels and color-changing films, it has controllability and stability. Therefore, it has unique characteristics for the design of identification structure to realize the authenticity identification of mass mobile phones. Fig.10 The processing method can achieve the effect of mobile phone public identification of product authenticity.

[0388] Whether it is possible to process a given set of discrete data and implement regression analysis through the above-mentioned ultra-deep adversarial learning is an important means to test the performance of machine learning. The test of actual data has proved that ultra-deep adversarial learning can obtain results that exceed actual calculations. The following specifically introduces a method of using ultra-deep adversarial learning to solve regression analysis.

[0389] Fig.12 It is a flowchart for solving regression analysis using ultra-deep adversarial learning.

[0390] like Fig.12 As shown: Using ultra-deep adversarial learning to solve regression analysis is achieved through the following steps:

[0391] Initialization step S1: In this step, we first need to give the initial regression analysis line, which can be the point position (x i ,y i )∈RD(i=1,2,…,n), the point with the minimum position (x min ,y min ), and the point at the maximum position (x max ,y max) are connected into a straight line as the initial regression analysis line. Since the probability scale self-organization can automatically find the place where the data is most densely distributed, the above method will not affect the final result because the initial regression analysis line is far away from the data distribution. In this step, the initial value of the probability scale must be given in advance. You can refer to the initialization requirements of the above multi-probability scale self-organization, and the maximum number of regression analysis cycles is given.

[0392] Calculate the distance from point to line Step S2: Substitute (x i ,y i )∈RD(i=1,2,…,n), calculate the distance from each point to the regression analysis line according to the following formula:

[0393] Formula 20

[0394]

[0395] Here, let the regression analysis line be: Ax+By+C=0

[0396] Probability scale self-organization step S3: follow the above Figure 2 Flow chart of multi-probability scale self-organization, for the regression analysis data set RD obtained in the previous step (x i ,y i )∈RD(i=1,2,…,n) and the distance d between each data and the regression analysis line i (i=1, 2, ..., n) performs self-organization of probability scale, retains data within the probability scale, and removes data outside the probability scale.

[0397] Obtain the regression analysis result Step S4: Initialize the retained dot matrix data in the maximum x data (x max ) to find a y center value near (x0, y0) max , the smallest x data (x min ) also finds a y center value near (x0, y0) min , use these two data to form a new regression analysis line, or use the center corresponding to the probability scale from point to line as a new regression analysis line, repeat this self-organization process, and you can get the maximum probability regression analysis line and the maximum probability range from point to line. The difference here from the result calculated by the traditional regression analysis formula is that because it is machine learning through probability scale self-organization, it can filter out noise and migrate to high probability data across the obstacles of low probability data.

[0398] Determine whether to continue learning step S5: According to the method for determining whether the multi-probability scale self-organization is completed, determine whether the result of the probability scale self-organization has stabilized, or whether it has exceeded the maximum number of self-organization iterations. If "yes", it ends and proceeds to the next step. If "no", it jumps to step S2 to calculate the distance from the point to the straight line and continue the machine learning of regression analysis.

[0399] Return to step S6: After machine learning is completed, return to the main program.

[0400] In order to verify the effect of regression analysis obtained through ultra-deep adversarial learning, we give a practical example of the statistics of cold coffee sales performance related to outdoor temperature in Table 2.

[0401] Table 2

[0402] Outdoor temperature 22 23 23 24 24 25 25 26 26 Sales quantity 300 310 320 330 320 330 310 320 310 Outdoor temperature 27 27 28 29 32 28 24 31 31 Sales quantity 340 360 350 360 400 370 310 360 390 Outdoor temperature 33 34 34 35 35 20 20 21 Sales quantity 400 450 460 440 480 350 300 380

[0403] Fig.13 It is a comparative diagram of the results of two regression analyses.

[0404] like Fig.13 As shown: Fig.13 (a) is the regression analysis result calculated by the traditional regression analysis formula for the data in Table 2, 1301 is the regression analysis straight line calculated by the traditional formula, 1302 and 1303 are the data sandwiched between the given range and the regression analysis straight line, and there are only 17 data within this range.

[0405] Fig.13 (b) is the regression analysis result achieved by ultra-deep adversarial learning, 1304 is the straight line of regression analysis achieved by ultra-deep adversarial learning, 1305 and 1306 are given and Fig.13 (a) The data in the same range as the regression analysis line obtained by ultra-deep adversarial learning, there are 21 data in this range.

[0406] By comparison, we can see that since the regression analysis result achieved through ultra-deep adversarial learning has the result of the regression analysis with the highest probability, the noise can be eliminated, so the area with the highest data distribution density can be found in the same area, so it can be called the regression analysis result with the highest probability.

[0407] Fig.14 It is a schematic diagram of the algorithm and calculation results of ultra-deep manifold learning.

[0408] like Fig.14 As shown: For the given manifold learning data, the ultra-deep manifold learning algorithm can be obtained through 47 steps Fig.14 The results are described below:

[0409] The first step S1: Randomly start from any position of manifold learning, first set the step size, the minimum step size is twice the width of the maximum probability scale of the probability scale self-organization. In this case, it can adapt to any data and achieve the highest precision calculation results, but the calculation time is relatively long. In order to assess the ability of ultra-deep manifold learning, here, a relatively long interval is deliberately used for processing, but according to the sampling theorem, the interval cannot be less than half of the distance between data points with no direct relationship, otherwise there will be errors on the path.

[0410] In step S1, the multi-probability scale self-organizing machine learning algorithm is first used at both ends to obtain the position of the central value of the maximum probability scale at both ends, and the two positions are connected into a straight line. The regression analysis algorithm is used to obtain a straight line that is closest to the data of the two distribution segments, and this straight line is used as the result of approximating the manifold learning curve between the two segments.

[0411] The second step S2: You can choose any end of the result of step S1 to continue the calculation. Here, we choose to continue processing from the left. Make an extension line represented by a dotted line with the same width as the interval to the left of the result of step S1. Here it just falls on the manifold learning data, so directly process this point according to the above process to obtain a central value of the maximum probability scale. As in the previous step, use this central value to connect into a straight line, continue regression analysis, and finally obtain the best approximation result of manifold learning within this step interval. The regression analysis result obtained here by ultra-deep adversarial learning can eliminate noise and obtain the regression analysis result closest to the matrix. Therefore, the approximation effect of manifold learning will also achieve the removal of noise and the approximation result with the maximum probability closest to the matrix.

[0412] The third step S3: In this step, the extension line is obviously out of the data of manifold learning, but it is covered by the processing range of multi-probability scale self-organization. Therefore, through multi-probability scale self-organization, it can autonomously migrate in the direction of the highest probability, and finally find the central value of the highest probability.

[0413] Step 4 S4: In this step, the extension line is obviously out of the manifold learning data, but Fig.14 As shown: there is a data that still falls within the processing range of the multi-probability scale self-organization. Through the guidance of this data, the multi-probability scale self-organization can autonomously migrate towards the high probability, and can pass through this low-probability data and directly migrate to the location of the data with higher probability density.

[0414] The fifth step S5: In this step, the extension line is seriously deviated from the manifold learning data. The given multi-probability scale self-organization range includes two scattered data on the left and the relatively high density data on the right. No matter which data is the result of the multi-probability scale self-organization, regression analysis within this range will definitely approach the high-density data on the right and eliminate the low-density data on the left.

[0415] Step 6 S6: A special case occurs in step S6 because the extension line approaches the manifold learning data that has no direct relationship. If this situation is followed, it will deviate from the original trajectory of the manifold learning. The reason for this problem is that in order to satisfy the sampling theorem emphasized in the first step, the spacing of the manifold learning processing cannot be greater than half of the span of the data that has no direct relationship. Therefore, it can be assumed here that the length of the extension line satisfies the above sampling theorem to ensure that there will be no problem of deviation from the manifold learning trajectory.

[0416] The seventh step S7: Here, when there is no data within the given multi-probability scale range, the initial multi-probability scale range can be increased.

[0417] Steps 8 to 12 S8-S 12 : It has been going smoothly, and the end point of manifold learning can be reached directly by extending the line.

[0418] Step 13S 13 : Go back to the first step S1, start from the right end through the extension line, find the manifold learning data through multi-probability scale self-organized migration, and perform regression analysis to obtain the straight line with the highest probability and the closest to the manifold learning data.

[0419] Step 14S 14 :The main feature of this step is that when the end point of the extension line deviates seriously from the manifold data, the initial size of the multi-probability scale can be increased until the data of manifold learning is covered.

[0420] Steps 15 to 16 15 -S 16 :S 15 No special handling requirements, S 16 It is necessary to increase the size of the multi-probability scale twice until the data for manifold learning is covered.

[0421] Steps 17 to 19 17 -S 19 :S 17 With S 18 No special handling requirements, S 19 Similar to S6, the extension cord is too long, so the extension cord should be shortened according to the sampling theorem.

[0422] Steps 20 to 23S 20 -S 23 :S 20 With S 21 No special handling requirements, S 22 With S 23 It is necessary to learn the data through the manifold that has been approximated, so the approximated manifold learning data needs to be labeled. This data cannot participate in the self-organizing learning of the probability scale again, but the nodes of the approximated straight line can be used for a second time.

[0423] Steps 28 to 29 28 -S 29 :S 28 No special handling requirements, S 29 Depending on the direction of the extension line, it needs to be connected to one end of S7.

[0424] Steps 30 to 38S 30-38 :S 30 , S 31 , S 32 , S 33 , S 34 and S 35 , no special processing requirements, from S 36 At the beginning, in a very narrow space, manifold learning needs to make a broken line. Similarly, for the data of this manifold learning, according to the sampling theorem, the processing interval cannot be too large, otherwise it cannot solve S 36 , S 37 and S 38 The problem of a 180-degree turn is that three steps are used here to make a 180-degree turn.

[0425] Steps 39 to 44 39 -S 44 :S 39 , S 40 , S 41 , S 42 and S 43 There are no special processing requirements. Following the guidance of the extension line, the multi-probability self-organization has the ability to migrate to the direction of high probability, and the regression analysis of image recognition of ultra-deep adversarial learning has the ability to remove noise and achieve the maximum probability approximation, so the best approximation of manifold learning data can be smoothly achieved. 44 The steps are guided by the extension line and require the use of a node of the processed manifold learning data.

[0426] Steps 45 to 47 45 -S 47 :S 45 , S 46 and S47 There is no special processing requirement and it can directly approach the starting point of manifold learning.

[0427] After ultra-deep manifold learning, a list of the best approximations of the manifold learning data is obtained. Spline functions are used to connect the straight lines into smooth continuous functions, which can also be differentiated.

[0428] The above-mentioned ultra-deep manifold learning uses the extension of the straight line that has already approximated the manifold learning data, and uses the characteristics of multi-probability scale self-organization that can migrate to the high-probability direction. Ultra-deep regression analysis can eliminate noise and achieve the best approximation characteristics, and the initial diameter of the multi-purpose probability scale can be increased as needed to achieve the best approximation of the manifold learning data. The processing interval of ultra-deep manifold learning can range from twice the diameter to several times the diameter of the maximum multi-probability scale after multi-probability scale self-organization.

[0429] Finally, it should be noted that ultra-deep manifold learning is suitable for processing manifold learning big data with a certain probability distribution. Its results can be fitted into a continuous function through a spline function, and even the derivative can be continuous.

[0430] Features of the present invention:

[0431] Summarizing the contents of the above description, the main points and features of the present invention are as follows:

[0432] 1. A method for defining multiple probability scales through the characteristics of probability distribution is proposed.

[0433] This method can calibrate the area of ​​probability distribution and adapt to the needs of machine learning to obtain the probability distribution information of the target function. By utilizing the characteristics of probability distribution, there is no need to study each data separately, and there is no need to study each probability distribution area separately, so an application effect with very high learning efficiency can be obtained. Because it is the scale of maximum probability, it adapts to the scale requirements of self-organizing machine learning.

[0434] 2. A method for defining multiple probability scales by learning the actual probability distribution is also proposed.

[0435] This method uses a very simple machine learning method to achieve learning results of multiple probability distributions for probability distribution data. It has the characteristics of simple method introduction, low hardware overhead, high computational efficiency, and is suitable for pattern recognition applications.

[0436] 3. A multi-probability scale self-organizing unsupervised machine learning algorithm is proposed, which can achieve autonomous migration towards high probability. It can overcome the obstacles of noise data in the process of probability distribution and migrate autonomously towards high probability until the area with maximum probability distribution is obtained, as well as the scale information of each area of ​​the entire probability distribution of the probability space.

[0437] 4. A strict definition of the distance that can be traversed between Euclidean space and probability space is proposed, which provides a theoretical basis for strictly obtaining the distance relationship between different probability distributions. It reveals that the distance in probability space is related to the probability value of the interval of probability distribution, and reveals that probability space exists in Euclidean space, and there can be infinite probability spaces in a Euclidean space, and that the distance in probability space is asymmetric and directional, etc. It provides a theoretical basis for improving the accuracy of pattern recognition, the data accuracy of machine learning results, and improving the efficiency of machine learning.

[0438] 5. A definition method for the probability measure of fuzzy events of data that can traverse different spaces including probability space is proposed. The microscopic fuzzy information and probability information between the position of any point in the space and the position of a certain probability distribution can be fully utilized. After integration, considerable and stable fuzzy information and probability information can be obtained. In addition, a more rigorous definition of the closer or farther relationship between two data in traversing Euclidean space and probability space is given, which provides a theoretical basis for further improving the accuracy and efficiency of machine learning.

[0439] 6. Several adversarial learning methods are proposed. The adversarial learning conducted on the strict fuzzy event probability measurement relationship established between a data and two probability distributions at any position in different spaces including probability space can obtain higher classification accuracy and can make the accuracy of pattern recognition reach the best level, providing a higher accuracy and higher learning effect algorithm for unsupervised machine learning. Strong adversarial learning can separate different feature vectors of data that cannot be classified traditionally through ultra-deep strong adversarial algorithms, which can solve the problem of mass anti-counterfeiting identification through mobile phones that cannot be solved internationally at present.

[0440] 7. A neural network model that best fits the brain mechanism is proposed. Compared with traditional neural networks, it avoids carrying the information of the target function through massive parameters, and avoids the problem of only being able to obtain local optimal solutions when performing data screening in the combination space of cosmic numbers among massive data in deep learning. It solves the black box problem of traditional neural networks, the problem that traditional neural networks cannot achieve unsupervised learning, and the problem that traditional neural networks require huge hardware support and cannot be applied on single-chip chips and wireless terminals. The ultra-deep adversarial learning neural network proposed in the present invention has the characteristics of high learning accuracy and high learning efficiency. It also has the ability to learn in parallel and the configuration of standard modules. It allows users to build the neural network architecture by themselves according to application needs. It is also suitable for the construction of hardware machine learning modules through chips.

[0441] 8. It can best approximate linear and nonlinear probability distribution data, obtain the regression analysis results with the maximum probability, and the best approximation effect of manifold learning data, remove noise, and cross the obstacles of noise data, autonomously approximate the area with the maximum probability, and the ultra-deep regression analysis learning results have obvious approximation effects with the results calculated by the actual formula. It proves that ultra-deep adversarial learning can compete with traditional deep learning algorithms and has advantages over traditional algorithms.

Claims

1. A method of image recognition based on ultra-deep adversarial learning, Its features are as follows: (1) When evaluating the characteristic elements of the probability distribution of the registered probability space to which the characteristic elements of the recognition object belong, the final result is obtained through adversarial learning established by the characteristic elements of the probability distribution of the two closest probability spaces; (2) The above adversarial learning is evaluated based on the fuzzy event probability measure constructed based on the distance across different spaces; specifically, given two feature vector data after self-organization through multiple probability scales, fv 1j ∈FV1 and fv 1j ∈FV2, given a feature element in the feature vector SV of an identified object; let the feature element svj∈SV in the feature vector SV of the identified object be the feature element fvj in the logged feature vector data FV1 1j The probability value in the probability distribution is sp j (fv1j) ∈SP (FV1) , in the feature vector data FV2, the feature element fv 2j The probability value in the probability distribution is sp j (fv2j) ∈SP (FV2) , j = 1, 2, ..., e; let mv2j be the number of multi-probability scales that the probability distribution center from svj to fv2j passes through, and let mj be the number of probability domains that the probability distribution center from svj to fv2j passes through fv2j =mv2j+1,fv 2j The multi-probability scale interval of the probability distribution is D ij , at a spacing of D ij The domain belongs to fv 2j The probability value of the probability distribution is Pij fv2j , i = 1, 2, ..., mj fv2j ; Calculate the fuzzy event probability measure of the identified object vector SV and the feature vector data FV2: Similarly, let sv j ∈SV to fv 1j The number of multi-probability scales that the probability distribution center passes through mv 1j , m j (fv1j) =mv 1j +1; the fuzzy event probability measure of the identified object vector SV and the feature vector data FV1 can be calculated: The identified object vector SV belongs to the feature vector data FV2 and the adversarial formula is as follows: Formula 9 F=F (FV2) / F (FV1) When F>1, the identified object vector SV belongs to the feature vector data FV2, and vice versa when F<1, the identified object vector SV belongs to the feature vector data FV1; Among them, the benchmark for quantitatively calibrating the probability distribution values ​​of complex regions is called multi-probability scale.

2. The method for image recognition using ultra-deep adversarial learning according to claim 1, characterized in that: The probability distribution refers to: normal distribution; multivariate normal distribution; lognormal distribution; exponential distribution; t distribution; F distribution; X 2 Distribution; Binomial distribution; Negative binomial distribution; Multinomial distribution; Poisson distribution; Erlang distribution: Hypergeometric distribution; Geometric distribution; Traffic distribution; Weibull distribution; Triangular distribution; Beta distribution; Any of the Gamma distributions or any probability distributions derived from Bayesian analysis; Any of the probability distributions in Gaussian processes.

3. The method for image recognition using ultra-deep adversarial learning according to claim 1, characterized in that: The distances of different spaces refer to: Euclidean space; probability space including Manhattan space; Chebyshev space; Minkowski space; Mahalanobis space; and any combination of spaces in angle cosine space.

4. The method for image recognition using ultra-deep adversarial learning according to claim 1, characterized in that: The fuzzy event probability measure can be expressed by the following formula: Suppose there exists a set of probability spaces w in Euclidean space j ∈W(w1,w2,…,w n ) has an element w j The center point of the probability distribution is w j The first scale of the multi-probability scale of this probability distribution is M1, the second scale is M2, the third scale is M3, and the center point is w j The distance from the first scale M1 is D 1j =M 1j , in this interval belongs to the set w j The probability value of the probability distribution is P 1j (wj) , the distance from the second scale to the first scale is D 2j =M 2j -M 1j , in this interval belongs to the set w j The probability value of the probability distribution is P 2j (wj) , the distance from the third scale to the second scale is D 3j =M 3j -M 2j The probability value of the probability distribution belonging to the set W in this interval is P 3j (wj) , assuming that the crossing belongs to w j The number of scales m passed in the probability space j (wj) =3, and suppose that there is another probability space set v in Euclidean space j ∈V(v1,v2,…,v n ), there is another point in the middle of the two sets, any point r j ∈R(j=1,2,…,n), find r j In w j to v j Between j Belong to w j The probability value of the probability distribution is rp j (wj) , find r j Arrival j Fuzzy event probability measure:

Images