OCR (optical character recognition) and network electronic identity authentication no-certificate comparison information system, device and inspection method

Abstract

The invention provides an OCR (Optical Character Recognition) and network electronic identity authentication no-certificate comparison information system and device and an inspection method, and relates to the technical field of information security. Firstly, the OCR technology is adopted to recognize certificate characters and upload names and identity cards, face key point data are extracted through a deep learning model, deep feature vectors are output, a local end preferentially queries database electronic identity cards through the names and the identity card numbers, similarity is calculated through a cosine similarity formula, offline authentication is supported, and the face recognition is realized. And if not, network electronic identity authentication is realized through a cloud identity authentication system, and an electronic identity card is returned by calculating the similarity. Meanwhile, based on the track, communication and consumption data, risk indexes are calculated through normalization processing and a weighted summation formula, the risk indexes are compared with a preset risk threshold value to evaluate the risk level, when the audited personnel are evaluated to be high in risk, the track, communication and consumption data are synchronously returned, and identity authentication and risk prevention and control in the non-certified scene are achieved.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and specifically to a non-certificate comparison information system, device, and verification method for OCR recognition and network electronic identity authentication. Background Art

[0002] In the prior art, mainstream solutions mostly use OCR recognition to extract certificate information and combine it with a cloud database for face feature comparison. For example, through image preprocessing and deep learning models, certificate text information such as name and ID number is extracted, and the registered face features stored in the system are retrieved through the network. Algorithms such as cosine similarity are used to complete identity verification. Although such technologies have high recognition efficiency in a networked environment, when the network is interrupted, the certificate is missing, or the user only holds an informal certificate such as a temporary certificate or a damaged ID card, the system cannot retrieve cloud data in real time or accurately extract key information, resulting in the interruption of the verification process. In addition, the prior art mostly relies on single biometric comparison and lacks dynamic risk assessment of user behavior data such as whereabouts, communication, and consumption, making it difficult to cope with security threats such as identity theft and abnormal behavior;

[0003] However, the above technologies still have significant deficiencies: First, identity verification cannot be effectively completed in an offline scenario, relying too much on cloud interaction, resulting in system failure when the network is interrupted; Second, the OCR recognition accuracy for informal certificates is insufficient, especially under conditions such as low image quality and uneven illumination, which is prone to character misjudgment or missed detection; Third, the risk prevention and control mechanism is single, only based on static identity information comparison, without integrating multi-dimensional behavior data, making it difficult to accurately assess potential risks. These technical defects limit the reliability and security of the identity authentication system in complex scenarios, and there is an urgent need for a solution that supports offline verification, high-precision informal certificate recognition, and multi-dimensional risk assessment.

[0004] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and therefore it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present invention is to provide a non-certificate comparison information system, device, and verification method for OCR recognition and network electronic identity authentication to solve the problems raised in the above background art.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] A non-certificate comparison information system for OCR recognition and network electronic identity authentication, comprising:

[0008] A feature extraction module, which is used to use OCR technology to perform character recognition on the informal identity document held by the person to be verified to extract information features, and simultaneously use a deep learning model to extract facial key point data and calculate a deep feature vector. The information features include name and ID number;

[0009] A local comparison module, which is used to query a database based on the information features, retrieve the photo of the person to be verified and calculate a registration feature vector, and calculate the facial similarity using the deep feature vector and the registration feature vector at the local end, and confirm whether the identity match is successful through the comparison result of the facial similarity and the comparison threshold;

[0010] A cloud comparison module, which is used to query the network electronic identity card that conforms to the information features from the cloud identity authentication system using the collected information features when the identity match fails at the local end or the information features recognized by OCR are incomplete, calculate a registration feature vector, and calculate the facial similarity using the deep feature vector and the registration feature vector in the cloud, and confirm whether the identity match is successful through the comparison result with the comparison threshold;

[0011] A risk assessment module, which is used to synchronize the network electronic identity and return it to the local end after the identity match is successful, and simultaneously calculate a risk index based on the whereabouts, communication, and consumption data, and judge whether the whereabouts, communication, and consumption data of the person to be verified are synchronously returned to the local end according to the comparison result of the risk index and the risk threshold.

[0012] Further, the method for using OCR technology to perform character recognition on the informal identity document held by the person to be verified to extract information features is as follows:

[0013] Preprocess the image of the informal identity document held by the person to be verified:

[0014] First, convert the color image to a single-channel grayscale image using the weighted average method:

[0015] I ,

[0013] ,

[0017] , ,

[0016] , i , ,

[0015] , gray , , gray ,

[0014] , , I(x, y)=0.2989·R(x, y)+0.5870·G(x, y)+0.1140·B(x, y)

[0016] Wherein, (x, y) represents the pixel coordinates in the character recognition process, I gray (x, y) represents the pixel value of the grayscale image at the coordinate (x, y), R(x, y) represents the pixel value of the red channel at the coordinate (x, y), G(x, y) represents the pixel value of the green channel at the coordinate (x, y), and B(x, y) represents the pixel value of the blue channel at the coordinate (x, y);

[0017] After that, count the pixel frequency p of the gray level [0, 255] i , and construct the between-class variance formula:

[0018] σ 2 σ(t) = ω0(t)·ω1(t)·(μ0(t) - μ1(t)) 2

[0019] In the formula, σ 2 σ(t) represents the between-class variance, and t represents the binarization threshold. Among them, ω1 = 1 - ω0,

[0020] Select the t that maximizes the value of the between-class variance σ 2 (t) as the binarization threshold. Pixels less than or equal to t are classified as the foreground, and pixels greater than t are classified as the background;

[0021] Finally, median filtering is used for denoising:

[0022] I denoised I(x, y) = median{I bin (x + 1, y + 1)}

[0023] In the formula, I denoised (x, y) represents the pixel value of the denoised image at the coordinate (x, y), and median represents taking the median of the neighboring pixels;

[0024] Based on the processed image, use the OCR engine to segment the character region and extract the identity information features, namely the name and ID number.

[0025] Furthermore, the method of using a deep learning model to extract the facial key point data and calculate the depth feature vector is as follows:

[0026] Collect N color image samples of human faces. Each sample corresponds to an identity label i, where N is a positive integer greater than or equal to 0, and i ∈ {1, 2,..., N}. Convert the color image samples of human faces into grayscale images:

[0027] I′ gray I′(x′, y′) = 0.2989·R(x′, y′) + 0.5870·G(x′, y′) + 0.1140·B(x′, y′)

[0028] In the formula, (x′, y′) represents the pixel coordinates during the conversion of the color image sample of the human face into a grayscale image, I′ gray (x′, y′) represents the grayscale image pixel value at the coordinate (x′, y′), R(x′, y′) represents the red channel pixel value at the coordinate (x′, y′), G(x′, y′) represents the green channel pixel value at the coordinate (x′, y′), and B(x′, y′) represents the blue channel pixel value at the coordinate (x′, y′);

[0029] Input the grayscale image into Dlib's face detector to output a face rectangle, whose coordinates are represented as (a, b, c, d). Here, a represents the column number of the left boundary of the face, b represents the row number of the upper boundary of the face, c represents the column number of the right boundary of the face, and d represents the row number of the lower boundary of the face. Input the face rectangle and the grayscale image into Dlib's 68-point key point model to output the absolute pixel coordinates (x′, y′) of 68 key points. Extract the center coordinates of the two eyes (x left ′, y left ′), (x right ′, y right ′) from the absolute pixel coordinates (x′, y′) of 68 key points, and calculate the rotation angle:

[0030] Calculate the horizontal difference and vertical difference between the centers of the two eyes:

[0031] Δx′ = x right ′ - x left ′

[0032] Δy′ = y right ′ - y left ′

[0033] Here, Δx′ represents the horizontal difference between the centers of the two eyes, and Δy′ represents the vertical difference between the centers of the two eyes;

[0034] Calculate the slope of the two eyes:

[0035]

[0036] Here, slope represents the slope of the two eyes;

[0037] Calculate the rotation angle through the arctangent:

[0038] θ = arctan(slope)

[0039] Here, θ represents the rotation angle;

[0040] Use affine transformation to rotate the face until the two eyes are horizontally aligned. When Δy′ > 0, rotate the face counterclockwise by θ; when Δy′ < 0, rotate the face clockwise by θ.

[0041] Crop the aligned face image to a fixed size of 160×160 pixels, and then calculate the normalized value for each pixel value:

[0042]

[0043] Here, P represents the pixel value, P′ represents the normalized pixel value, and the identity label corresponding to each normalized pixel value of the face image is M;

[0044] Using the deep feature extraction model FaceNet, traverse the face images normalized based on the pixel values of the RGB color channels, randomly select the anchor point g, the positive sample h, and the negative sample n to construct a triplet (g, h, n), and ensure that the anchor point g and the positive sample h are face images normalized with the same pixel values, and the negative sample n satisfies the face images normalized with different pixel values, and construct the triplet loss function:

[0045]

[0046] In the formula, f(g i ), f(h i ) and f(n i ) represent the functions that map the face images normalized with pixel values to feature vectors, that is, the feature mapping functions, and α represents a positive boundary value. Among them, the definition of the feature mapping function is:

[0047] f(l) = FC(ReLU)(Conv3(ReLU(Conv2(ReLU(Conv1(l))))))

[0048] In the formula, Conv1 represents 64 7×7 convolutional kernels with a stride of 2; Conv2 represents 192 5×5 convolutional kernels with a stride of 2; Conv3 represents 384 3×3 convolutional kernels with a stride of 1; FC represents a 128-dimensional fully connected layer, and the output result is V local , that is, the deep feature vector;

[0049] Construct the complete loss function:

[0050]

[0051] Use 70% of the color image samples of faces as the training set, and then update the network parameter θ′ for the complete loss function through gradient descent. The gradient is: Adopt the Adam optimizer, set the learning rate η = 0.001, the decay rate β1 = 0.9, β2 = 0.999, and terminate the training when the loss value no longer decreases for 50 consecutive epochs and the accuracy of the validation set reaches more than 99.5%.

[0052] Furthermore, the method for retrieving the photo of the person to be verified and calculating the registered feature vector is:

[0053] Convert the photo of the person to be verified from the database into a grayscale image:

[0054] I′ gray (x′, y′) = 0.2989·R(x′, y′) + 0.5870·G(x′, y′) + 0.1140·B(x′, y′)

[0055] Input the grayscale image into Dlib’s face detector to obtain the coordinates of the face rectangle (a, b, c, d). Then input the face rectangle and grayscale image into Dlib’s 68-point key point model and output the absolute pixel coordinates (x′, y′) of the 68 key points.

[0056] Extract the center coordinates of the two eyes (x left ′,y left ′),(x right ′,y right ′), calculate the horizontal and vertical differences between the centers of the two eyes:

[0057] Δx′=x right ′-x left '

[0058] Δy′=y right ′-y left '

[0059] Calculate the slope of the two eyes:

[0060]

[0061] Calculate the rotation angle using the inverse tangent:

[0062] θ=arctan(slope)

[0063] Use affine transformation to rotate the face so that the eyes are horizontally aligned, crop the aligned face image to a fixed size of 160×160 pixels, and normalize each pixel value:

[0064]

[0065] Use the trained deep feature extraction model FaceNet to process the normalized pixel value photos and map them into feature vectors:

[0066] f(l)=FC(ReLU)(Conv3(ReLU(Conv2(ReLU(Conv1(l))))))

[0067] Output registered feature vector V reg .

[0068] Furthermore, the face similarity is calculated locally using the deep feature vector and the registered feature vector. The method for confirming whether the identity match is successful is as follows:

[0069] Calculate the depth feature vector V for the photos collected by the verified person local , use cosine similarity to calculate the similarity between the deep feature vector and the registered feature vector:

[0070]

[0071] Wherein, S represents the face similarity, and V local represents the depth feature vector, and V reg represents the registered feature vector;

[0072] Preset a comparison threshold T yz , and T yz ∈[0.85, 0.98]. When S≥T yz , it is determined that the matching is successful; when S<T yz , it is determined that the matching fails.

[0073] Furthermore, when using the collected information features to query the network electronic ID card that conforms to the information features in the cloud identity authentication system, query through the name and ID number in the cloud identity authentication system. The cloud identity authentication system retrieves the matching network electronic ID card, calculates the registered feature vector, calculates the face similarity between the depth feature vector and the registered feature vector in the cloud, and returns the network electronic ID card with successful identity matching through the comparison result with the comparison threshold; when the ID number is missing, query through the name in the cloud identity authentication system. The cloud identity authentication system retrieves the completely matching network electronic ID card, calculates the registered feature vector, calculates the face similarity between the depth feature vector and the registered feature vector in the cloud, and returns the network electronic ID card with successful identity matching through the comparison result with the comparison threshold; when there is no certificate and the person being audited does not provide information features either, calculate the face similarity between the depth feature vector and the registered feature vector of the cloud identity authentication system, and return the network electronic ID card with successful identity matching through the comparison result with the comparison threshold.

[0074] Furthermore, the method for calculating the risk index based on whereabouts, communication, and consumption data is as follows:

[0075] Statistical number of times crossing cities T within 7 days k , number of calls to strange numbers C m and consumption amount O r , and perform normalization processing on the data:

[0076]

[0077] Wherein, T kmin represents the historical minimum value of the number of times crossing cities, T kmax represents the historical maximum value of the number of times crossing cities, C mmin represents the minimum value of the number of calls to strange numbers, C mmax represents the maximum value of the number of calls to strange numbers, O rmin represents the historical minimum value of the consumption amount, O rmaxLet the maximum historical consumption amount be denoted as, the normalized value of the number of cross - city trips be denoted as T, the normalized value of the number of calls to unfamiliar numbers be denoted as C, and the normalized value of the consumption amount be denoted as O;

[0078] The risk index is calculated by the method of weighted summation:

[0079] R = 0.5T+0.3C + 0.2O

[0080] In the formula, R represents the risk index.

[0081] Furthermore, the method for judging whether the whereabouts, communication and consumption data of the verified person are synchronously returned to the local end according to the comparison result between the risk index and the risk threshold is as follows:

[0082] The risk threshold δ is preset to be 0.7. When the risk index R≥0.7, the verified person is judged to be a high - risk person, and the network electronic identity card, the whereabouts, communication and consumption data of the verified person are returned to the local end; when the risk index R < 0.7, only the network electronic identity card is returned to the local end.

[0083] The present invention also provides a method for verifying non - document comparison information of OCR recognition and network electronic identity authentication. The detection method is obtained by the execution of a non - document comparison information system for OCR recognition and network electronic identity authentication as described above, and includes:

[0084] Step 1: Use OCR technology to perform character recognition on the informal identity document held by the verified person to extract information features, and simultaneously use a deep learning model to extract face key point data and calculate the deep feature vector. The information features include name and ID number;

[0085] Step 2: Query the database based on the information features, retrieve the photo of the verified person and calculate the registered feature vector, and calculate the face similarity using the deep feature vector and the registered feature vector at the local end. Confirm whether the identity match is successful through the comparison result between the face similarity and the comparison threshold;

[0086] Step 3: When the identity match fails at the local end or the information features recognized by OCR are incomplete, use the collected information features to query the network electronic identity card that meets the information features from the cloud identity authentication system, calculate the registered feature vector, and calculate the face similarity using the deep feature vector and the registered feature vector in the cloud, and confirm whether the identity match is successful through the comparison result with the comparison threshold;

[0087] Step 4: After the identity match is successful, the network electronic identity is synchronized and returned to the local end. At the same time, calculate the risk index based on the whereabouts, communication and consumption data, and judge whether the whereabouts, communication and consumption data of the verified person are synchronously returned to the local end according to the comparison result between the risk index and the risk threshold.

[0088] The present invention further provides a license-free comparison information inspection device for OCR recognition and network electronic identity authentication, and the device includes a processor and a memory;

[0089] The memory is used for storing computer programs or instructions;

[0090] The processor is used for executing the computer programs or instructions in the memory to enable the device to execute the above-mentioned license-free comparison information inspection method for OCR recognition and network electronic identity authentication.

[0091] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0092] By integrating OCR technology and deep learning face key point extraction technology, the present invention realizes accurate recognition of the name and ID number information on the informal identity documents of the person to be verified, and at the same time calculates and extracts the deep feature vector of the face to realize double verification of information features and face features, significantly improving the accuracy and reliability of identity recognition. The present invention also confirms the identity through local database query and feature comparison, and uses the cloud comparison method as a supplement to improve the recognition rate for the situation of local recognition failure or incomplete information. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 It is a schematic diagram of the overall method flow of the present invention;

[0094] Figure 2 It is a schematic diagram of the overall system module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0095] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the following further describes the present invention in detail with reference to specific embodiments.

[0096] It should be noted that unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or items appearing before this word cover the elements or items listed after this word and their equivalents, without excluding other elements or items. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to indicate relative position relationships, and when the absolute position of the object to be described changes, the relative position relationships may also change accordingly.

[0097] Embodiment:

[0098] Please refer to Figure 1 , the present invention provides a technical solution:

[0099] An identity comparison information system without documents for OCR recognition and network electronic identity authentication, including a feature extraction module, a local comparison module, a cloud comparison module, and a risk assessment module, wherein:

[0100] The feature extraction module uses OCR technology to perform character recognition on the informal identity document held by the person to be verified to extract information features, and simultaneously uses a deep learning model to extract face key point data and calculate deep feature vectors. The information features include name and ID number.

[0101] Take a photo of the informal identity document held by the person to be verified, such as a driver's license, temporary ID card, household register certificate, or screenshot of an unofficial electronic document, etc., while avoiding backlighting or direct strong light to prevent the document from reflecting light or shadows from obscuring the text, and then preprocess the image:

[0102] First, convert the color image to a single-channel grayscale image using the weighted average method:

[0103] I gray (x, y) = 0.2989·R(x, y) + 0.5870·G(x, y) + 0.1140·B(x, y)

[0104] In the formula, (x, y) represents the pixel coordinates in the character recognition process, I gray (x, y) represents the pixel value of the grayscale image at the coordinate (x, y), R(x, y) represents the pixel value of the red channel at the coordinate (x, y), G(x, y) represents the pixel value of the green channel at the coordinate (x, y), and B(x, y) represents the pixel value of the blue channel at the coordinate (x, y), eliminating color interference and simplifying the computational complexity;

[0105] After that, count the pixel frequency p of the gray level [0, 255] i , and construct the between-class variance formula:

[0106] σ 2 (t) = ω0(t)·ω1(t)·(μ0(t) - μ1(t)) 2

[0107] In the formula, σ 2 (t) represents the between-class variance, t represents the binarization threshold, where ω1 = 1 - ω0,

[0108]

[0109] Select the t that maximizes the between-class variance σ 2 as the binarization threshold. Pixels less than or equal to t are classified as foreground, and pixels greater than t are classified as background. By threshold segmentation, the image is converted into a black-and-white binary image, which can highlight the contrast between the foreground and the background and facilitate subsequent OCR recognition;

[0110] Finally, median filtering is used for denoising:

[0111] I denoised (x, y) = median{I bin (x + 1, y + 1)}

[0112] In the formula, I denoised (x, y) represents the pixel value of the denoised image at the coordinate (x, y), and median represents taking the median of the neighborhood pixels, which improves the OCR recognition accuracy and avoids character misjudgment caused by noise;

[0113] Based on the processed image, use the OCR engine to segment the character regions:

[0114] Traverse each row of the image, count the number of foreground pixels, and obtain the sum of foreground pixels for each row. After accumulating the sum of foreground pixels for all rows and dividing by the number of rows of the image, the row average projection value is obtained. Divide the row average projection value by 2 to get the row segmentation line threshold. Next, traverse each row, count the sum of foreground pixels, and the row where the sum of foreground pixels is first less than the row segmentation line threshold will be marked as the starting position of row segmentation. Continue traversing until a row greater than the row segmentation line threshold is encountered, and this row is the ending position of row segmentation. The region between the starting position and the ending position of row segmentation is defined as the row segmentation line. This process will be repeated until all row segmentation lines are recognized. Define the interval between the starting position and the ending position of row segmentation as a single-line binary text image;

[0115] Count the number of foreground pixels for each column of the single-line binary text image, accumulate the sum of foreground pixels for all columns, divide by the number of columns to get the column average projection value, and then divide the column average projection value by 2 as the column segmentation line threshold. Next, traverse each column, count the sum of foreground pixels for the column, and the column where the sum of foreground pixels is first less than the column segmentation line threshold will be marked as the starting position of column segmentation. Continue traversing until a column greater than the column segmentation line threshold is encountered, and this column is the ending position of column segmentation. The region between the starting position and the ending position of column segmentation is defined as the column segmentation line. This process will be repeated until all column segmentation lines are recognized, and the interval between the starting position and the ending position of column segmentation is defined as the column range of a single character;

[0116] According to the obtained row division areas and column division areas above, crop the sub-images of each character from the binary image and input them into an OCR engine, such as Baidu OCR or Tencent OCR, for single-character recognition. In the same row, splice the characters in sequence from left to right, and for the text in different rows, arrange them from top to bottom. Locate the two keywords "name" and "ID card". Regardless of whether there are spaces, extract the character sequences following "name" and "ID card", and discard delimiters such as colons and quotation marks.

[0117] Collect color image samples of N human faces. For example, N = 100, and each sample corresponds to an identity label i, where N is a positive integer ≥ 0, and i ∈ {1, 2, …, N}. Convert the color image samples of human faces into grayscale images:

[0118] I′ gray (x′, y′) = 0.2989·R(x′, y′) + 0.5870·G(x′, y′) + 0.1140·B(x′, y′)

[0119] In the formula, (x′, y′) represents the pixel coordinates during the conversion of the color image sample of the human face into a grayscale image, and I′ gray (x′, y′) represents the grayscale image pixel value at the coordinate (x′, y′), R(x′, y′) represents the red channel pixel value at the coordinate (x′, y′), G(x′, y′) represents the green channel pixel value at the coordinate (x′, y′), and B(x′, y′) represents the blue channel pixel value at the coordinate (x′, y′);

[0120] Input the grayscale image into Dlib's face detector. Dlib's face detector is a face detection algorithm based on HOG features and cascade regression trees, which has stable output even in low-quality images and relatively accurate positioning. Obtain the face rectangular box, and its coordinates are represented as (a, b, c, d). In the formula, a represents the column number of the left boundary of the face, b represents the row number of the upper boundary of the face, c represents the column number of the right boundary of the face, and d represents the row number of the lower boundary of the face. Input the face rectangular box and the grayscale image into Dlib's 68-point key-point model, which can mark key positions such as the face contour, eyes, nose, and mouth, and output the absolute pixel coordinates (x′, y′) of 68 key points. Extract the coordinates (x left ′, y left ′), (x right ′, y right ′) of the centers of the two eyes, and calculate the rotation angle:

[0121] Calculate the horizontal difference and vertical difference between the centers of the two eyes:

[0122] Δx′ = xright ′ - x left ′

[0123] Δy′ = y right ′ - y left ′

[0124] Wherein, Δx′ represents the horizontal difference between the centers of the two eyes, and Δy′ represents the vertical difference between the centers of the two eyes;

[0125] Calculate the slope of the two eyes:

[0126]

[0127] Wherein, slope represents the slope of the two eyes;

[0128] Calculate the rotation angle through the arctangent:

[0129] θ = arctan(slope)

[0130] Wherein, θ represents the rotation angle;

[0131] Use affine transformation to rotate the face until the two eyes are horizontally aligned. When Δy′ > 0, the face is rotated counterclockwise by θ; when Δy′ < 0, the face is rotated clockwise by θ. In this way, the pose difference is eliminated, and face images at different angles are unified to the frontal face view, ensuring the consistency of subsequent feature extraction, such as pupil distance and facial feature ratio;

[0132] Crop the aligned face image to a fixed size of 160×160 pixels (the default size of FaceNet) to improve the model inference speed, and then calculate the normalized value for each pixel value:

[0133]

[0134] Wherein, P represents the pixel value, and P′ represents the normalized pixel value, which can enhance comparability. Each pixel value-normalized face image corresponds to an identity label M;

[0135] Adopt the deep feature extraction model FaceNet, which has high accuracy and is generally recognized as suitable for large-scale face recognition scenarios. Traverse the face images normalized based on the pixel values of the RGB color channels, randomly select the anchor point g, the positive sample h, and the negative sample n, and construct the triplet (g, h, n), which has the effect of optimizing the feature space, making the distance between similar samples closer and that between dissimilar samples farther. When constructing the triplet (g, h, n), it is necessary to ensure that the anchor point g and the positive sample h are face images normalized with the same pixel value, and the negative sample n satisfies being a face image normalized with a different pixel value, and construct the triplet loss function:

[0136]

[0137] Wherein, f(g i ), f(h i ), and f(n i ) represent functions that map a face image normalized to pixel values into feature vectors, i.e., feature mapping functions, and α represents a positive boundary value. Among them, the definition of the feature mapping function is:

[0138] f(l) = FC(ReLU)(Conv3(ReLU(Conv2(ReLU(Conv1(l))))))

[0139] Wherein, Conv1 represents 64 7×7 convolutional kernels with a stride of 2; Conv2 represents 192 5×5 convolutional kernels with a stride of 2; Conv3 represents 384 3×3 convolutional kernels with a stride of 1; FC represents a 128-dimensional fully connected layer, and the output result is V local , that is, the deep feature vector;

[0140] Construct the complete loss function:

[0141]

[0142] Take 70% of the color image samples of faces as the training set, which is a common choice to balance the data volume and the model generalization ability. Then, update the network parameter θ' of the complete loss function through gradient descent, and the gradient is: That is, batch gradient descent. Calculate the gradient mean of all color image samples in each iteration, which converges more stably, reduces the interference of noise on the gradient, and is suitable for the deep feature learning of face recognition;

[0143] Adopt the Adam optimizer, set the learning rate η = 0.001. The adaptive learning rate mechanism will dynamically adjust according to the gradient history. 0.001 is a commonly used initial value in practice, which balances the convergence speed and stability. The decay rates are β1 = 0.9 and β2 = 0.999. When the loss value does not decrease for 50 consecutive epochs and the verification set accuracy rate reaches more than 99.5%, terminate the training, which can not only prevent overfitting but also avoid misjudgment due to short-term fluctuations, ensuring that the model meets the practical standard in the real scenario.

[0144] The local comparison module queries the database based on the information features, retrieves the photo of the person to be verified and calculates the registered feature vector, and calculates the face similarity using the deep feature vector and the registered feature vector at the local end, and confirms whether the identity match is successful through the comparison result of the face similarity and the comparison threshold.

[0145] Store the names, ID numbers, and corresponding deep feature vectors of all registered users in the database. Even offline, an electronic ID card can be obtained locally, which has the advantages of short time consumption and reduced cloud requests. Convert the photo of the person to be verified into a grayscale image:

[0146] I′ gray (x′, y′) = 0.2989·R(x′, y′) + 0.5870·G(x′, y′) + 0.1140·B(x′, y′)

[0147] Input the grayscale image into Dlib's face detector to obtain the coordinates (a, b, c, d) of the face rectangular box. Then, input the face rectangular box and the grayscale image into Dlib's 68-point key point model to output the absolute pixel coordinates (x′, y′) of 68 key points;

[0148] Extract the center coordinates (x left ′, y left ′) of the two eyes from the 68 key points, (x right ′, y right ′), and calculate the horizontal difference and vertical difference between the centers of the two eyes:

[0149] Δx′ = x right ′ - x left ′

[0150] Δy′ = y right ′ - y left ′

[0151] Calculate the slope of the two eyes:

[0152]

[0153] Calculate the rotation angle through the arctangent:

[0154] θ = arctan(slope)

[0155] Use affine transformation to rotate the face until the two eyes are horizontally aligned. Crop the aligned face image to a fixed size of 160×160 pixels and normalize each pixel value:

[0156]

[0157] Use the trained deep feature extraction model FaceNet to process the photo with normalized pixel values. In the neural network, a large numerical range may lead to problems such as gradient explosion or gradient disappearance, affecting the training and convergence of the model. By normalizing the pixel values to the range of -1 to 1, the values can be made more stable, which helps the training and optimization of the model, and map it to a feature vector:

[0158] f(l) = FC(ReLU)(Conv3(ReLU(Conv2(ReLU(Conv1(l))))))

[0159] Output the registered feature vector V reg 。

[0160] Cosine similarity is a commonly used vector space metric method. It can measure the cosine value of the angle between two vectors, reflecting the directional similarity between the two vectors. In face recognition, using cosine similarity can well measure the similarity degree between two face feature vectors.

[0161] Calculate the depth feature vector V for the photo collected from the person to be verified local , and use cosine similarity to calculate the similarity between the depth feature vector and the registered feature vector:

[0162]

[0163] In the formula, S represents the face similarity, V local represents the depth feature vector, and V reg represents the registered feature vector;

[0164] Preset the comparison threshold T yz , and T yz ∈[0.85, 0.98]. When S ≥ T yz , it is determined that the match is successful; when S < T yz , it is determined that the match fails. Generally speaking, for scenarios with a lower requirement for the false recognition rate, T yz can be set to 0.95 to minimize the risk of false matching.

[0165] When the local end fails to pass the identity match or the information features recognized by OCR are incomplete, the cloud comparison module uses the collected information features to query the network electronic ID that meets the information features from the cloud identity authentication system, calculates the registered feature vector, and calculates the face similarity between the depth feature vector and the registered feature vector in the cloud, and confirms whether the identity match is successful according to the comparison result with the comparison threshold.

[0166] When the local end fails to pass the identity match or the information features recognized by OCR are incomplete, the local end can either identify the information features (name, ID number) of the certificate through OCR technology, or let the person to be reviewed orally state the name and ID number, and use speech recognition technology to convert it into text information, and send the collected information features and the depth feature vector to the cloud comparison module. This is because the local end, due to resource and data limitations, cannot complete accurate identity matching in some cases, so the task needs to be handed over to the cloud to ensure the accuracy of the identity verification process;

[0167] When querying the network electronic ID that matches the information features from the cloud identity authentication system using the collected information features, query the cloud identity authentication system by name and ID number. The cloud identity authentication system retrieves the matching network electronic ID, calculates the registration feature vector, calculates the face similarity using the deep feature vector and the registration feature vector in the cloud, and returns the network electronic ID with successful identity matching based on the comparison result with the comparison threshold;

[0168] In actual application scenarios, situations such as damaged or unclear documents may occur, resulting in the inability to accurately obtain the ID number, or the user forgets to bring the ID card but can provide information such as the name. That is, when the ID number is missing and querying the cloud identity authentication system by name can narrow the query scope to a certain extent and provide clues for identity verification. The cloud identity authentication system retrieves the completely matching network electronic ID, calculates the registration feature vector, calculates the face similarity using the deep feature vector and the registration feature vector in the cloud, and returns the network electronic ID with successful identity matching based on the comparison result with the comparison threshold;

[0169] In some urgent or special situations, the user may not be able to provide complete document information. When there is no document and the audited person does not provide information features either, calculate the face similarity using the deep feature vector and the registration feature vector of the cloud identity authentication system. The face is a biometric feature with uniqueness and relative stability. Each person's facial features are unique and difficult to forge within a certain period. Based on the comparison result with the comparison threshold, return the network electronic ID with successful identity matching. Without the user providing any documents or information features, only by collecting the face image can the identity comparison be carried out, greatly simplifying the identity authentication process and saving time and effort. Especially in emergency situations, it can quickly give the identity matching result and improve work efficiency.

[0170] After successful identity matching, the risk assessment module synchronizes the network electronic identity and returns it to the local end. At the same time, calculate the risk index based on the whereabouts, communication, and consumption data, and judge whether the whereabouts, communication, and consumption data of the audited person are synchronously returned to the local end according to the comparison result between the risk index and the risk threshold.

[0171] When the local end detects that the user does not carry the ID card, it automatically enters the risk assessment process to prevent identity theft or fraud caused by the lack of documents. The number of times crossing cities, the number of calls to strange numbers, and the consumption amount reflect the behavior patterns of the audited person from different aspects. Frequent crossing of cities may imply special activities; frequent calls to strange numbers may pose potential risks; abnormal changes in the consumption amount may also be manifestations of risky behaviors. By comprehensively considering these factors, the risk status of the audited person can be evaluated more comprehensively. With the help of travel records, obtain the call records of the audited person from the communication operator and through bank transaction records, and count the number of times T that the audited person crosses cities within 7 daysk and the number of calls C from unfamiliar numbers m and the consumption amount O r , normalize the data as follows:

[0172]

[0173]

[0174] In the formula, T kmin represents the historical minimum value of the number of cross - city trips, T kmax represents the historical maximum value of the number of cross - city trips, C mmin represents the minimum value of the number of calls from unfamiliar numbers, C mmax represents the maximum value of the number of calls from unfamiliar numbers, O rmin represents the historical minimum value of the consumption amount, O rmax represents the historical maximum value of the consumption amount. T represents the normalized value of the number of cross - city trips, C represents the normalized value of the number of calls from unfamiliar numbers, and O represents the normalized value of the consumption amount, thus eliminating the dimension difference;

[0175] Calculate the risk index using the weighted summation method:

[0176] R = 0.5T+0.3C + 0.2O

[0177] In the formula, R represents the risk index. Among them, cross - city data is usually directly provided by the database. Experience shows that it is the core index for risk assessment and can directly reflect the abnormality degree of the user's whereabouts. Therefore, the weight of the number of cross - city trips is set to 0.5. Making calls from unfamiliar numbers (such as frequently dialing virtual numbers, overseas numbers) is a typical feature of telecom fraud and illegal transactions, which can better reflect the current risk status of the user. Therefore, the weight of the number of calls from unfamiliar numbers is set to 0.3. The consumption amount itself is not directly equivalent to risk, but combined with scenarios such as a large number of small - value payments and consumption in sensitive places, it may increase the risk probability. Therefore, the weight of the consumption amount is set to 0.2. Calculate the risk index using the weighted summation method to avoid misjudgment of a single behavior.

[0178] Preset the risk threshold δ to 0.7. This risk threshold is verified by historical data and achieves a balance between the false alarm rate and the missed detection rate. At the same time, the threshold can be adjusted according to regions, time periods, and policies. For example, it is increased to 0.8 during major events. When the risk index R≥0.7, it is judged that the person to be verified is at high risk, and the network e - ID, the whereabouts, communication, and consumption data of the person to be verified are returned to the local end; when the risk index R < 0.7, only the network e - ID is returned to the local end.

[0179] For example, when a user checks into a hotel without an ID card, has traveled across cities twice in the past 7 days, made 6 strange calls, and spent 1,500 yuan, after normalization, T = 0.5, C = 0.6, and O = 0.3;

[0180] Calculate the risk index:

[0181] R = 0.5×0.5 + 0.3×0.6 + 0.2×0.3 = 0.49

[0182] The risk index R = 0.49 < 0.7, only return the electronic ID card

[0183] When the user travels across cities 4 times, makes 8 strange calls, and spends 5,000 yuan, after normalization, T = 0.9, C = 0.8, and O = 0.7;

[0184] Calculate the risk index:

[0185] R = 0.5×0.9 + 0.3×0.8 + 0.2×0.7 = 0.83

[0186] The risk index R = 0.83 > 0.7, the local end only returns the electronic ID card of the person to be verified, and the whereabouts, communication, and consumption data need to be focused on.

[0187] Please refer to Figure 2 , the present invention also provides a method for verifying unlicensed comparison information for OCR recognition and network electronic identity authentication. The detection method is obtained by executing the above-mentioned system for verifying unlicensed comparison information for OCR recognition and network electronic identity authentication, and includes:

[0188] Step 1: Use OCR technology to perform character recognition on the informal identity document held by the person to be verified to extract information features, and simultaneously use a deep learning model to extract face key point data and calculate the deep feature vector. The information features include name and ID number;

[0189] Step 2: Query the database based on the information features, retrieve the photo of the person to be verified and calculate the registration feature vector. Calculate the face similarity using the deep feature vector and the registration feature vector at the local end, and confirm whether the identity match is successful through the comparison result of the face similarity and the comparison threshold;

[0190] Step 3: When the local end fails to match the identity or there is no information feature for OCR recognition, use the collected information features to query the network electronic ID card that meets the information features from the cloud identity authentication system, calculate the registration feature vector, and calculate the face similarity using the deep feature vector and the registration feature vector in the cloud, and confirm whether the identity match is successful through the comparison result with the comparison threshold;

[0191] Step 4: After successful identity matching, the network electronic identity is synchronized and returned to the local end. Meanwhile, a risk index is calculated based on whereabouts, communication, and consumption data, and it is determined whether the whereabouts, communication, and consumption data of the person being verified are synchronized and returned to the local end according to the comparison result between the risk index and the risk threshold.

[0192] The present invention further provides an OCR recognition and network electronic identity authentication device for verifying unlicensed comparison information. The device includes a processor and a memory.

[0193] The memory is used to store computer programs or instructions.

[0194] The processor is used to execute the computer programs or instructions in the memory, so that the device executes the above-mentioned method for verifying unlicensed comparison information of OCR recognition and network electronic identity authentication.

[0195] The above formulas are all calculated by taking the numerical values without dimensions. The formula is obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. The preset parameters in the formula are set by those skilled in the art according to the actual situation.

[0196] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0197] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units. They can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0198] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application, and all should be covered by the protection scope of the present application.

Claims

1. An identity comparison information system without certificates for OCR recognition and network electronic identity authentication, characterized in that, Including: A feature extraction module, which is used to use OCR technology to perform character recognition on the informal identity document held by the person to be verified, extract information features, and simultaneously use a deep learning model to extract face key point data and calculate a deep feature vector. The information features include name and ID number; A local comparison module, which is used to query the database based on the information features, retrieve the photo of the person to be verified and calculate the registered feature vector, calculate the face similarity using the deep feature vector and the registered feature vector at the local end, and confirm whether the identity match is successful through the comparison result of the face similarity and the comparison threshold; A cloud comparison module, which is used to query the network electronic ID that conforms to the information features from the cloud identity authentication system using the collected information features when the identity match fails at the local end or the information features recognized by OCR are incomplete, calculate the registered feature vector, calculate the face similarity using the deep feature vector and the registered feature vector in the cloud, and confirm whether the identity match is successful through the comparison result with the comparison threshold; A risk assessment module, which is used to synchronize the network electronic identity to the local end after the identity match is successful, and at the same time calculate the risk index based on the whereabouts, communication, and consumption data, and judge whether the whereabouts, communication, and consumption data of the person to be verified are synchronized and returned to the local end according to the comparison result of the risk index and the risk threshold.

2. The non-certificate comparison information system for OCR recognition and network electronic identity authentication according to claim 1, characterized in that: The method for using OCR technology to perform character recognition on the informal identity document held by the person to be verified and extract information features is as follows: Convert the color image to a single-channel grayscale image using the weighted average method: I gray (x, y) = 0.2989·R(x, y) + 0.5870·G(x, y) + 0.1140·B(x, y) Wherein, (x, y) represents the pixel coordinates in the character recognition process, I gray (x, y) represents the gray-scale image pixel value at the coordinates (x, y), R(x, y) represents the red-channel pixel value at the coordinates (x, y), G(x, y) represents the green-channel pixel value at the coordinates (x, y), and B(x, y) represents the blue-channel pixel value at the coordinates (x, y); Count the number of pixels at each gray level, and then calculate its pixel frequency to construct the between-class variance formula: σ 2 (t) = ω0(t)·ω1(t)·(μ0(t) - μ1(t)) 2 Where, σ 2 (t) represents the between-class variance, t represents the binarization threshold, and p i represents the proportion of the number of pixels with a gray value of i in the single-channel gray image to the total number of pixels, that is, the pixel frequency, and i represents the gray value index. Among them, ω1(t) = 1 - ω0(t), Select the t that maximizes the between-class variance σ 2 (t) as the binarization threshold, classify the pixels less than or equal to the binarization threshold as the foreground, and classify the pixels greater than the binarization threshold as the background; Use median filtering to denoise: I denoised (x, y) = median{I bin (x + 1, y + 1)} where I denoised (x, y) represents the pixel value of the denoised image at the coordinate (x, y), and median represents taking the median of the neighboring pixels; Based on the processed image, use the OCR engine to segment the character area and extract the identity information features, namely name and ID number.

3. An identityless comparison information system for OCR recognition and network electronic identity authentication according to claim 1, characterized in that: The method for using a deep learning model to extract face key point data and calculate a deep feature vector is as follows: Collect color image samples of N faces, each sample corresponding to an identity label i, where N is a positive integer ≥ 0, and i ∈ {1, 2,..., N}, and convert the color image samples of the faces into grayscale images: I′ gray (x′, y′) = 0.2989·R(x′, y′) + 0.5870·G(x′, y′) + 0.1140·B(x′, y′) Wherein, (x′, y′) represents the pixel coordinates in the process of converting a color image sample of a human face into a grayscale image, and I′ gray (x′, y′) represents the grayscale image pixel value at the coordinate (x′, y′), R(x′, y′) represents the red channel pixel value at the coordinate (x′, y′), G(x′, y′) represents the green channel pixel value at the coordinate (x′, y′), and B(x′, y′) represents the blue channel pixel value at the coordinate (x′, y′); Input the grayscale image into Dlib's face detector to output a face rectangle, whose coordinates are represented as (a, b, c, d). Here, a represents the column number of the left boundary of the face, b represents the row number of the upper boundary of the face, c represents the column number of the right boundary of the face, and d represents the row number of the lower boundary of the face. Input the face rectangle and the grayscale image into Dlib's 68-point key point model to output the absolute pixel coordinates (x′, y′) of 68 key points. Extract the center coordinates of the two eyes (x left ′, y left ′), (x right ′, y right ′) from the absolute pixel coordinates (x′, y′) of 68 key points, and calculate the rotation angle: Calculate the horizontal difference and vertical difference between the centers of the two eyes: Δx′ = x right ′ - x left ′ Δy′ = y right ′ - y left ′ In the formula, Δx′ represents the horizontal difference between the centers of the two eyes, and Δy′ represents the vertical difference between the centers of the two eyes; Calculate the slope of the two eyes: In the formula, slope represents the slope of the two eyes; Calculate the rotation angle through the arctangent: θ = arctan(slope) In the formula, θ represents the rotation angle; Use affine transformation to rotate the face until the two eyes are horizontally aligned. When Δy′ > 0, rotate the face counterclockwise by θ, and when Δy′ < 0, rotate the face clockwise by θ. Crop the aligned face image to a fixed size of 160 × 160 pixels, and then calculate the normalized value for each pixel value: In the formula, P represents the pixel value, P′ represents the normalized pixel value, and the identity label corresponding to each normalized pixel value of the face image is M; Using the deep feature extraction model FaceNet, traverse the face images normalized based on the pixel values of the RGB color channels, randomly select the anchor point g, the positive sample h, and the negative sample n to construct a triplet (g, h, n), and ensure that the anchor point g and the positive sample h are face images normalized with the same pixel values, and the negative sample n is a face image normalized with non-identical pixel values, and construct a triplet loss function: where f(g i ), f(h i ), and f(n i ) represent functions that map a face image after pixel value normalization to a feature vector, i.e., feature mapping functions, and α represents a positive boundary value. Among them, the definition of the feature mapping function is as follows: f(l) = FC(ReLU)(Conv3(ReLU(Conv2(ReLU(Conv1(l)))))) Wherein, Conv1 represents 64 7×7 convolutional kernels with a stride of 2; Conv2 represents 192 5×5 convolutional kernels with a stride of 2; Conv3 represents 384 3×3 convolutional kernels with a stride of 1; FC represents a fully connected layer with 128 dimensions, and the output result is V local , that is, the deep feature vector; Construct a complete loss function: Use 70% of the color image samples of human faces as the training set, and then update the network parameters θ′ for the complete loss function through gradient descent. The gradient is: Adopt the Adam optimizer, set the learning rate η = 0.001, the decay rate β1 = 0.9, β2 = 0.

999. When the loss value no longer decreases for 50 consecutive epochs and the accuracy of the validation set reaches over 99.5%, terminate the training.

4. The license-free comparison information system for OCR recognition and network electronic identity authentication according to claim 3, characterized in that: The method for retrieving the photo of the person to be verified and calculating the registered feature vector is as follows: Convert the photo of the person to be verified from the database into a grayscale image: I′ gray (x′, y′) = 0.2989·R(x′, y′) + 0.5870·G(x′, y′) + 0.1140·B(x′, y′) Input the grayscale image into the face detector of Dlib to obtain the face rectangle coordinates (a, b, c, d), and then input the face rectangle and the grayscale image into the 68-point key point model of Dlib to output the absolute pixel coordinates (x′, y′) of 68 key points; Extract the center coordinates of the two eyes (x left ′, y left ′), (x right ′, y right ′) from 68 key points, and calculate the horizontal difference and vertical difference between the centers of the two eyes: Δx′ = x right ′ - x left ′ Δy′ = y right ′ - y left ′ Calculate the slope of the two eyes: Calculate the rotation angle through the arctangent: θ = arctan(slope) Use affine transformation to rotate the face to make the two eyes horizontally aligned, crop the aligned face image into a fixed size of 160×160 pixels, and normalize each pixel value: Use the trained deep feature extraction model FaceNet to process the photo with normalized pixel values and map it into a feature vector: f(l) = FC(ReLU)(Conv3(ReLU(Conv2(ReLU(Conv1(l)))))) Output the registered feature vector V reg .

5. An identity-free comparison information system for OCR recognition and network electronic identity authentication according to claim 4, characterized in that: The method for calculating the face similarity using the deep feature vector and the registered feature vector at the local end and confirming whether the identity match is successful through the comparison result of the face similarity and the comparison threshold is as follows: Calculate the depth feature vector V for the photo collected from the person to be verified local , and use cosine similarity to calculate the similarity between the depth feature vector and the registered feature vector: Wherein, S represents the face similarity, and V local represents the depth feature vector, and V reg represents the registered feature vector; Preset the comparison threshold T yz , and T yz ∈[0.85, 0.98]. When S ≥ T yz , it is determined that the match is successful; when S < T yz , it is determined that the match fails.

6. The non-certificate comparison information system for OCR recognition and network electronic identity authentication according to claim 1, wherein: When querying the network electronic ID that conforms to the information features from the cloud identity authentication system using the collected information features, query through the name and ID number. The cloud identity authentication system retrieves the matching network electronic ID, calculates the registered feature vector, calculates the face similarity using the deep feature vector and the registered feature vector in the cloud, and returns the network electronic ID with a successful identity match through the comparison result with the comparison threshold; When the ID number is missing, query through the name. The cloud identity authentication system retrieves the completely matching network electronic ID, calculates the registered feature vector, calculates the face similarity using the deep feature vector and the registered feature vector in the cloud, and returns the network electronic ID with a successful identity match through the comparison result with the comparison threshold; When there are no documents and the person being reviewed does not provide information features, calculate the face similarity using the deep feature vector and the registered feature vector of the cloud identity authentication system, and return the network electronic ID with a successful identity match through the comparison result with the comparison threshold.

7. An identityless comparison information system for OCR recognition and network electronic identity authentication according to claim 1, characterized in that: The method for calculating the risk index based on whereabouts, communication, and consumption data is as follows: Statistically count the number of times T of crossing cities within 7 days k and the number of calls C to unfamiliar numbers m and the consumption amount O r , and perform normalization processing on the data: Where, T kmin represents the historical minimum value of the number of times of crossing cities, T kmax represents the historical maximum value of the number of times of crossing cities, C mmin represents the minimum value of the number of calls from unfamiliar numbers, C mmax represents the maximum value of the number of calls from unfamiliar numbers, O rmin represents the historical minimum value of the consumption amount, O rmax represents the historical maximum value of the consumption amount, T represents the normalized value of the number of times of crossing cities, C represents the normalized value of the number of calls from unfamiliar numbers, and O represents the normalized value of the consumption amount; Adopt the method of weighted summation to calculate the risk index: R = 0.5T + 0.3C + 0.2O In the formula, R represents the risk index.

8. An identityless comparison information system for OCR recognition and network electronic identity authentication according to claim 7, characterized in that: The method for judging whether the whereabouts, communication and consumption data of the person to be verified are synchronously returned to the local end according to the comparison result between the risk index and the risk threshold is as follows: The risk threshold δ is preset to 0.

7. When the risk index R≥0.7, the person to be verified is judged as a high-risk person, and the network electronic identity card. The whereabouts, communication and consumption data of the person to be verified are returned to the local end; when the risk index R<0.7, only the network electronic identity card is returned to the local end.

9. A method for verifying unlicensed comparison information in OCR recognition and network electronic identity authentication, characterized in that, The inspection method is obtained by an identityless comparison information system for OCR recognition and network electronic identity authentication described in any one of claims 1-8, and includes: Step 1: Use OCR technology to perform character recognition on the informal identity document held by the person to be verified to extract information features, and at the same time use a deep learning model to extract face key point data and calculate the deep feature vector. The information features include name and ID number; Step 2: Query the database based on the information features, retrieve the photo of the person to be verified and calculate the registered feature vector. Calculate the face similarity using the deep feature vector and the registered feature vector at the local end, and confirm whether the identity match is successful through the comparison result between the face similarity and the comparison threshold; Step 3: When the identity match fails at the local end or the information features recognized by OCR are incomplete, use the collected information features to query the network electronic identity card that meets the information features from the cloud identity authentication system, calculate the registered feature vector, and calculate the face similarity using the deep feature vector and the registered feature vector in the cloud, and confirm whether the identity match is successful through the comparison result with the comparison threshold; Step 4: After the identity match is successful, the network electronic identity is synchronized and returned to the local end. At the same time, calculate the risk index based on the whereabouts, communication and consumption data, and judge whether the whereabouts, communication and consumption data of the person to be verified are synchronously returned to the local end according to the comparison result between the risk index and the risk threshold.

10. An information inspection device for license-free comparison of OCR recognition and network electronic identity authentication, characterized in that, The device includes a processor and a memory; The memory is used to store computer programs or instructions; The processor is used to execute the computer programs or instructions in the memory, so that the device executes an identityless comparison information inspection method for OCR recognition and network electronic identity authentication as described in claim 9.

Images