Digital e-commerce platform based on cloud computing

Abstract

The invention discloses a digital e-commerce platform based on cloud computing, and relates to the field of data processing, and the platform comprises the steps: collecting a plurality of frames of face images; preprocessing the collected multiple frames of face images; a pre-trained convolutional neural network is adopted to carry out feature extraction on the preprocessed multiple frames of face images, and feature vectors are obtained; performing similarity calculation on the obtained feature vector and a pre-stored feature vector in an authorized officer face feature library to obtain a similarity score; if the calculated similarity score is greater than a threshold value, the identity verification is passed, and the user is allowed to access the platform; if the calculated similarity score is less than or equal to the threshold value, the identity verification fails; when the identity authentication fails, performing authentication according to a preset second identity authentication mode; in view of low security of identity verification of an e-commerce platform in the prior art, a second identity verification mode based on combination of voiceprint recognition and iris recognition is used, so that the security of identity verification is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a digital e-commerce platform based on cloud computing. Background Art

[0002] With the rapid development of Internet technology, e-commerce has become an important part of modern business activities. E-commerce platforms provide consumers with a convenient online shopping experience and also provide merchants with a broad sales channel. However, as e-commerce flourishes, security issues such as identity fraud and account theft have become increasingly prominent, posing a serious threat to user privacy and property security, and becoming a bottleneck restricting the further development of e-commerce.

[0003] Existing e-commerce platforms usually use a simple username and password method for identity authentication, which is easy to crack and has low security. Although some platforms have introduced auxiliary verification methods such as SMS verification codes and hardware keys, it is still difficult to fundamentally solve the problem of identity fraud. In addition, most existing identity authentication methods are single-modal, that is, only a single biometric feature such as face, fingerprint, etc. is used for verification. This single-modal verification method is not reliable enough and is difficult to adapt to complex and changeable fraud scenarios. Therefore, there is an urgent need for a multi-modal, hierarchical and integrated identity authentication solution to protect the account security and transaction security of e-commerce platforms from multiple angles and in all directions. Summary of the invention

[0004] In response to the problem of low identity authentication security in e-commerce platforms existing in the prior art, the present application provides a digital e-commerce platform based on cloud computing. Face recognition is used as the first identity authentication method. When the first identity authentication fails, the second identity authentication method based on the combination of voiceprint recognition and iris recognition is enabled. This can achieve hierarchical identity authentication and effectively improve the security of identity authentication.

[0005] Technical solution, the purpose of this application is achieved through the following technical solution.

[0006] The present specification provides a digital e-commerce platform based on cloud computing, including: collecting multiple frames of face images; preprocessing the collected multiple frames of face images; preferably, the image preprocessing also includes: using an adaptive histogram equalization algorithm to perform illumination equalization processing on the multiple frames of face images to reduce the influence of illumination changes on face feature extraction; using affine transformation to perform posture correction on the multiple frames of face images after illumination equalization processing, adjusting the face to a frontal posture, and obtaining a face image after posture correction; using a liveness detection algorithm based on blinking frequency and facial micro-expression changes to perform liveness analysis on the face image after posture correction to determine whether it is a real face; detecting the blinking frequency of the face area, if the blinking frequency is lower than a set threshold, it is determined to be non-live; analyzing the changing characteristics of facial micro-expressions, extracting the movement trajectory of facial key points such as mouth corners and eyebrows, and constructing a facial micro-expression feature vector; inputting the facial micro-expression feature vector into a support vector machine classifier to identify whether it is a natural expression change, if it is an unnatural expression change, it is determined to be non-live. A pre-trained convolutional neural network is used to extract features from preprocessed multi-frame face images to obtain feature vectors; the convolutional neural network adopts a residual network structure and introduces an attention mechanism to strengthen the network's attention to key feature areas such as eyes, nose, mouth, etc., and improve the feature representation capability; and an image quality assessment module is added before the convolutional neural network to eliminate images that do not meet the standards for lighting, blur, occlusion, etc., and improve the accuracy of subsequent recognition.

[0007] The obtained feature vector is similarly calculated with the feature vector in the pre-stored authorized personnel face feature library to obtain a similarity score; preferably, the cosine similarity of the extracted feature vector and multiple feature vectors in the face feature library is respectively calculated to obtain a group of similarity scores; the group of similarity scores is weighted averaged, and the weight is determined according to the quality score of each feature vector at the time of registration to obtain a final similarity score; if the calculated similarity score is greater than the threshold, the identity authentication is passed and the user is allowed to access the platform; if the calculated similarity score is less than or equal to the threshold, the identity authentication fails; when the identity authentication fails, verification is performed according to the pre-set second identity authentication method; preferably, a distributed storage architecture is adopted to divide the face feature sub-library among multiple service nodes to achieve load balancing and parallel processing; for each sub-library, an algorithm such as local sensitive hashing is used to generate an index to accelerate large-scale face feature retrieval; a master-slave hot standby mechanism is established to ensure high availability and data consistency of the face feature library.

[0008] Among them, the feature vector represents the features of an object or sample as a vector. In pattern recognition and machine learning, data such as images and speech are often converted into feature vectors as input to the algorithm. The feature vector extracts the key information of the data, making similar samples closer in the feature space. A deep neural network is an artificial neural network model that contains multiple hidden layers. DNN can automatically extract high-level semantic representations of data through hierarchical feature learning. In tasks such as speech recognition and image classification, DNN has made breakthrough progress with its powerful modeling capabilities. Common DNN models include convolutional neural networks (CNN) and recurrent neural networks (RNN).

[0009] Furthermore, the second identity authentication method includes: extracting user voice data, using a deep neural network model to perform voiceprint recognition, and obtaining voice recognition similarity; collecting user iris images, extracting iris features, calculating the similarity with pre-stored iris features, and obtaining iris recognition similarity; inputting the voice recognition similarity and iris recognition similarity into a Bayesian network fusion algorithm to obtain a user identity authenticity verification result; preferably, constructing a dynamic Bayesian network model, on the basis of a static Bayesian network, introducing a time series factor, using the voice recognition similarity and iris recognition similarity verified for multiple consecutive times as input, and modeling the dynamic change process of identity authentication; each time slice of the dynamic Bayesian network includes a voice recognition similarity node, an iris recognition similarity node, a voice credibility node, an iris credibility node, and an identity authentication result node, and the time slices are connected by state transition probabilities; based on the user's historical verification data, an expectation maximization (EM) algorithm is used to learn the parameters of the dynamic Bayesian network, including the conditional probability distribution and state transition probability distribution of each node; wherein, a forward-backward algorithm is used to calculate the expected steps in the EM algorithm, and a Junction Tree algorithm is used for inference calculation; during identity authentication, the similarity of voice recognition and iris recognition of users' continuous interactions is obtained to construct an observation sequence; then, the Viterbi algorithm is used to perform inference on the dynamic Bayesian network to find the most likely hidden state sequence, that is, the dynamic change process of identity authenticity; according to the probability of identity authenticity at the final moment, combined with the global threshold and decision rules, it is determined whether the user's identity has passed the verification; the decision rule comprehensively considers factors such as the global threshold, historical verification results, and user suspicious behavior, and adopts a weighted voting mechanism to adaptively adjust the standard of identity passing; if the identity authentication passes, the verification data of this time is added to the user's historical data set, and the parameters of the dynamic Bayesian network are updated using the incremental learning algorithm to achieve online adaptive optimization of the model; otherwise, the user's access is denied, and the early warning mechanism is activated to record suspicious behaviors for subsequent security analysis. According to the results of user identity authenticity verification, the user's access to the platform is controlled.

[0010] The iris is a colorful ring-shaped tissue in the human eye, and its texture contains rich details, such as spots and grooves. Iris features refer to a series of numerical indicators extracted from iris images, which are used to characterize the visual properties of iris texture. Commonly used iris features include Gabor wavelet features and local binary pattern (LBP) features. Iris features are unique and stable and are widely used in iris recognition. Bayesian network is an inference framework based on probabilistic graphical models. The dependency relationship between variables is represented by a directed acyclic graph, the nodes represent random variables, and the edges represent conditional dependencies. The Bayesian network fusion algorithm uses Bayesian networks to fuse multi-source information, and by modeling the causal relationship between features, comprehensively considers the degree of support for the hypothesis by each piece of evidence, and makes the final decision.

[0011] Furthermore, the speech recognition similarity is obtained, including: collecting user speech data; using a speech denoising algorithm based on a deep neural network to denoise the collected user speech data; performing endpoint detection and silence removal on the denoised speech data to obtain preprocessed speech data; using a linear prediction analysis method to extract linear prediction cepstral coefficient features of the preprocessed speech data; using a resonance peak estimation algorithm to extract resonance peak frequency features of the preprocessed speech data; performing feature splicing on the extracted linear prediction cepstral system features and resonance peak frequency features to obtain a voiceprint feature vector; using the obtained voiceprint feature vector as input, and using a pre-trained deep neural network model to obtain speech similarity. The voiceprint feature vector is input into a pre-trained deep neural network model, and the similarity with the pre-stored voiceprint is calculated through the forward propagation of the network to obtain the similarity as the speech recognition similarity; the deep neural network model adopts the attention mechanism and the connection time classification (CTC) loss function to improve the accuracy and robustness of voiceprint recognition; at the same time, an adversarial learning strategy is introduced in the model training, and a highly simulated false voiceprint sample is synthesized by generating an adversarial network (GAN) to enhance the anti-deception ability of the model; the input of the deep neural network model is the voiceprint feature vector, and the output is the first verification similarity; the network structure of the deep neural network model includes multiple convolutional layers, pooling layers and fully connected layers, as well as a Softmax classification layer; the deep neural network model is obtained through pre-training, and the pre-training process includes: using a large amount of user voice data as training samples, optimizing the network parameters using the BP algorithm and the stochastic gradient descent algorithm, minimizing the recognition loss function, and obtaining the optimal model parameters. Preferably, when calculating the speech recognition similarity, not only the static similarity of the voiceprint features is considered, but also the similarity measurement of the dynamic features of the voiceprint is introduced; a sequence matching algorithm based on dynamic time warping (DTW) is used to calculate the optimal alignment distance of two voiceprint feature sequences to characterize the dynamic change characteristics of the voiceprint in the time dimension; the static similarity and dynamic similarity are weighted and fused to obtain the final speech recognition similarity.

[0012] Among them, speech denoising refers to the process of removing background noise from noisy speech signals and restoring pure speech. Speech denoising algorithms adaptively adjust filter parameters to suppress noise components by modeling the statistical characteristics of speech and noise. Common speech denoising algorithms include spectral subtraction, Wiener filtering, and statistical models such as MMSE. Deep learning methods such as convolutional denoising autoencoders are also used for speech denoising tasks. Endpoint detection refers to automatically determining the start and end positions of valid speech segments in speech signals. Silence removal is to remove silence segments from speech signals and only retain segments containing speech content. Both are common steps in speech preprocessing, the purpose of which is to extract effective information from speech and reduce data redundancy. Features such as short-time energy and zero-crossing rate are often used for endpoint detection. Methods based on energy thresholds or statistical models can be used for silence removal.

[0013] Furthermore, a linear prediction analysis method is used to extract linear prediction cepstral coefficient features of preprocessed speech data, including: performing frame processing on the preprocessed speech data to obtain a speech frame sequence; performing windowing processing on the speech frame sequence using a Hamming window function to obtain a windowed speech frame; performing autocorrelation function calculation on the windowed speech frame to obtain an autocorrelation function coefficient; according to the autocorrelation function coefficient, using a Levinson-Durbin recursive algorithm to calculate a linear prediction coefficient; performing a fast Fourier transform on the linear prediction coefficient to obtain a frequency domain prediction coefficient; taking the logarithm and inverse Fourier transform of the frequency domain prediction coefficient to obtain a cepstral coefficient; and extracting multidimensional cepstral coefficients as linear prediction cepstral coefficient features.

[0014] Preferably, the linear prediction cepstral coefficient feature is normalized to obtain a normalized linear prediction cepstral coefficient feature; wherein the normalization adopts a Zero Mean and Unit Variance (ZMUV) algorithm; the mean and standard deviation of the linear prediction cepstral coefficient feature are calculated; the mean is subtracted from the linear prediction cepstral coefficient feature and then divided by the standard deviation to obtain a normalized feature; the normalized linear prediction cepstral coefficient feature is compensated for the channel length using a vocal tract length normalization algorithm to obtain a linear prediction cepstral coefficient feature normalized by the channel length; the vocal tract length parameter of the normalized linear prediction cepstral coefficient feature is estimated; according to the vocal tract length parameter, the normalized linear prediction cepstral coefficient feature is frequency scaled to compensate for the difference in vocal tract length; the linear prediction cepstral coefficient feature normalized by the channel length is subjected to discrete cosine transform (DCT) to obtain the DCT coefficient of the cepstral feature; and the low-order coefficients in the DCT coefficients of the cepstral feature are extracted as the final linear prediction cepstral coefficient feature.

[0015] Among them, linear prediction analysis is a speech signal processing technology. Through short-term analysis of speech signals, a linear combination relationship between the current speech sample and several historical samples is established to achieve linear prediction of speech signals. The linear prediction coefficient characterizes the resonance peak characteristics of speech signals and has important applications in speech coding, speech synthesis and other fields. The autocorrelation function describes the correlation between the sample values ​​of the signal at different times. For a discrete signal x(n), its autocorrelation function R(k) is defined as: R(k) = ∑x(n)x(nk), which represents the product sum of the signal and its delayed k samples. The autocorrelation function has properties such as symmetry and the maximum value at the origin. The short-term autocorrelation function of speech is often used to extract features such as pitch and resonance peaks. The Levinson-Durbin algorithm is an efficient recursive algorithm for solving linear prediction coefficients. Given the autocorrelation function of the speech signal, the algorithm avoids the complexity of solving the matrix equation by recursively calculating the reflection coefficient and the linear prediction coefficient. The Levinson-Durbin algorithm ensures the stability of the prediction coefficient so that it meets the requirements of the minimum phase system.

[0016] Furthermore, a resonance peak estimation algorithm is used to extract the resonance peak frequency characteristics of the preprocessed speech data, including: calculating a linear prediction spectrum according to the obtained linear prediction coefficient; smoothing the linear prediction spectrum using a triangular window function to obtain a smoothed linear prediction spectrum; obtaining the local maximum point of the smoothed linear prediction spectrum as the initial resonance peak frequency; performing harmonic detection on the initial resonance peak frequency, eliminating the harmonic peaks, and obtaining the optimized resonance peak frequency; obtaining the top N resonance peak frequencies with the highest energy in the optimized resonance peak frequency as the resonance peak frequency characteristics; performing discrete Fourier transform on the resonance peak frequency characteristics to obtain frequency domain resonance peak characteristics; performing triangular Mel filter group filtering on the frequency domain resonance peak characteristics to obtain Mel frequency domain resonance peak characteristics; performing logarithmic transform and discrete cosine transform on the Mel frequency domain resonance peak characteristics to obtain MFCC characteristics of the resonance peak; splicing the MFCC characteristics of the resonance peak with the corresponding first-order and second-order difference parameters to obtain the final resonance peak frequency characteristics.

[0017] Among them, the Fourier transform converts the time domain signal into the frequency domain signal, and the inverse Fourier transform restores the frequency domain signal to the time domain signal. For the frequency domain signal X(f), its inverse Fourier transform is: x(t) = ∫X(f)e j2πftdf. Inverse Fourier transform is often used in frequency domain filtering, spectrum analysis and other fields. In speech processing, inverse Fourier transform is used to convert prediction coefficients in the frequency domain into linear prediction coefficients in the time domain. Cepstrum analysis obtains cepstrum coefficients by taking the logarithm of the spectrum of the speech signal and then performing an inverse Fourier transform. Multidimensional cepstrum coefficients refer to cepstrum coefficients extracted from multiple dimensions (such as time, frequency, etc.), which comprehensively characterize the spectral envelope characteristics of the speech signal. Commonly used multidimensional cepstrum coefficients include Mel frequency cepstrum coefficients (MFCC), perceptual linear prediction coefficients (PLP), etc. The triangular window is a time domain windowing function with a linearly rising and falling triangular shape. Its mathematical definition is: Where N is the window length. The frequency domain characteristics of the triangular window are low side lobes and wide main lobes. In speech processing, the triangular window is often used to smooth the speech spectrum and suppress rapid changes in the spectrum. The local maximum point refers to the point where the function achieves the maximum value in the local range. For a discrete sequence x(n), if x(n)>x(n-1) and x(n)>x(n+1), then x(n) is a local maximum point. In speech processing, local maximum points are often used to detect the formant, fundamental period and other characteristics of the speech signal. Harmonics refer to sine waves whose frequencies are integer multiples of the fundamental frequency. The speech signal consists of fundamental and harmonic waves, and its spectrum has spectral peaks with concentrated energy at the fundamental and harmonic frequencies, which are called harmonic peaks. Harmonic peaks reflect the periodic structure of speech, but they interfere with the estimation of formant peaks. The influence of harmonic peaks needs to be eliminated when extracting formant peaks. Discrete Fourier transform is a Fourier transform for discrete signals, which converts time domain signals into frequency domain signals. For a discrete signal x(n) of length N, its DFT is: k=0, 1,..., N-1.

[0018] DFT decomposes the signal into sine waves of different frequencies to obtain the spectrum representation of the signal. Fast Fourier Transform (FFT) is an efficient implementation algorithm of DFT. Mel filter bank is a group of triangular bandpass filters whose center frequencies are arranged according to the Mel frequency scale. Mel frequency takes into account the nonlinear characteristics of human hearing, with high resolution in the low-frequency part and low resolution in the high-frequency part. The triangular Mel filter bank extracts the frequency components related to perception by weighted averaging the energy of the speech spectrum to obtain the Mel spectrum. Logarithmic transformation converts the amplitude spectrum of the Mel spectrum into a logarithmic spectrum, compressing the dynamic range of the amplitude to make it closer to the response characteristics of the human ear. Discrete cosine transform is an orthogonal transform in the real number domain, which converts the logarithmic Mel spectrum into the cepstrum domain to obtain the Mel frequency cepstrum coefficient (MFCC). DCT has excellent properties such as energy concentration and decorrelation, making MFCC a common feature of speech recognition. The differential parameter describes the dynamic change information of speech features and is often used as a supplement to static features. Assume that the static feature vector is x(n), the first-order difference Δx(n)=x(n+1)-x(n), which reflects the instantaneous rate of change of the feature; the second-order difference ΔΔx(n)=Δx(n+1)-Δx(n), which reflects the trend of the rate of change of the feature. The differential parameters are generally estimated by linear regression and other methods, and then used together with the original features.

[0019] Furthermore, the iris recognition similarity is obtained, including: using an iris detection algorithm based on a convolutional neural network to locate the iris of an iris image to obtain an iris area; using an image segmentation algorithm to segment the iris area to obtain a segmented iris image; wherein the image segmentation algorithm constructs an energy function of the iris area to obtain a global optimal segmentation curve to perform iris segmentation; using a Gabor wavelet transform method to extract iris texture features of the image after iris segmentation; using a method based on Zernike moments to extract iris shape features of the image after iris segmentation; using a method based on Markov random fields to extract iris texture statistical features of the image after iris segmentation; performing feature fusion on the extracted iris texture features, iris shape features and iris texture statistical features to obtain an iris feature vector; using the iris feature vector as input and using the Hamming distance metric to calculate the iris recognition similarity.

[0020] Among them, iris detection refers to the process of automatically locating the iris area from the eye image. The iris detection algorithm based on convolutional neural network uses deep learning methods to automatically extract the significant features of the iris area by training the convolutional neural network model to achieve accurate iris positioning. This algorithm overcomes the shortcomings of traditional methods that are sensitive to image quality and illumination changes through end-to-end feature learning. Image segmentation refers to the process of dividing an image into several regions with specific meanings and extracting targets of interest. The image segmentation algorithm achieves semantic image segmentation by defining the criteria for region division and optimizing the region boundaries. Common image segmentation algorithms include threshold segmentation, region growing, edge detection, graph segmentation, etc. In iris recognition, image segmentation is used to extract iris texture images with clear texture from the iris area. Gabor wavelet is a local frequency analysis tool that can simultaneously obtain the local frequency and phase information of the signal in the time-frequency domain. The two-dimensional Gabor wavelet is obtained by modulating the Gaussian function and has directional selectivity and multi-scale characteristics. Gabor wavelet transform extracts multi-directional and multi-scale texture features of images by convolving images with a set of Gabor wavelets of different scales and directions, and has important applications in iris texture feature extraction. Zernike moment is an orthogonal moment function defined by the product of Zernike polynomials and radial polynomials. Zernike moment has the advantages of rotational invariance, noise resistance and low information redundancy, and is often used for image shape description and recognition. The method based on Zernike moment calculates the Zernike moment of iris image to extract global features related to iris shape, which is used for feature representation in iris recognition. Markov random field is a probabilistic graphical model that describes spatial contextual dependencies. It characterizes the texture characteristics of images by defining the local dependency structure between pixels. The method based on Markov random field models the iris image as a Markov random field, and extracts the statistical features of iris texture, such as gray-level co-occurrence matrix, local binary pattern, etc., by defining the interaction between pixels and cluster potential function, for iris feature representation.

[0021] Furthermore, an image segmentation algorithm is used to perform iris segmentation on the iris area to obtain a segmented iris image; wherein the image segmentation algorithm constructs an energy function of the iris area to obtain a global optimal segmentation curve to perform iris segmentation, including: representing the iris area as a weighted undirected graph; wherein the nodes of the graph are pixels of the iris area, and the weights of the edges are the similarities between adjacent pixels; constructing an energy function of the weighted undirected graph, the energy function includes a region term and a boundary term; wherein the region term represents the probability that a pixel belongs to the iris area or the non-iris area, and the boundary term represents the probability that adjacent pixel labels are consistent; the maximum flow minimum cut algorithm is used to minimize the energy function to obtain a global optimal segmentation curve; the iris area is segmented according to the global optimal segmentation curve to obtain a segmented iris image; wherein the inside of the segmentation curve is the iris area, and the outside is the non-iris area.

[0022] Among them, the weighted undirected graph is a data structure in graph theory, which consists of nodes and weighted edges connecting nodes. In image segmentation, the image pixels are regarded as nodes of the graph, and the similarity between pixels is used as the weight of the edge to construct a weighted undirected graph representation of the image. By defining the energy function on the graph, the image segmentation problem can be transformed into the optimal partitioning problem of the graph for solution. The energy function is the objective function that describes the image segmentation problem and is used to measure the quality of the segmentation result. The energy function of image segmentation usually includes two terms: the region term and the boundary term. The region term measures the similarity of pixels inside the segmented region and encourages the consistency of pixel features in the region; the boundary term measures the smoothness of the boundary of the segmented region and encourages the boundary to coincide with the edge of the image. By minimizing the energy function, the global optimal image segmentation result can be obtained. The maximum flow minimum cut algorithm is a classic algorithm for solving the minimum cut problem of the graph. In graph theory, the maximum flow problem refers to the problem of finding the maximum flow between the source and the sink; the minimum cut problem refers to the problem of finding the minimum cut set that divides the graph into two parts. The maximum flow minimum cut theorem proves the equivalence of the two problems. In image segmentation, the image is divided into foreground and background areas by constructing a flow network of the image and using the maximum flow minimum cut algorithm to solve the global optimal segmentation.

[0023] Furthermore, a Gabor wavelet transform method is used to extract iris texture features of the image after iris segmentation, including: constructing a Gabor wavelet basis function, wherein the parameters of the Gabor wavelet include scale, direction, center frequency and bandwidth; performing radial normalization processing on the segmented iris image to convert the annular iris image into a rectangular image; performing a convolution operation on the obtained rectangular image using the constructed Gabor wavelet basis function to obtain Gabor wavelet transform coefficients; taking the modulus of the Gabor wavelet transform coefficients to obtain a Gabor wavelet energy image; dividing the Gabor wavelet energy image into blocks, calculating the mean and variance of each block, and obtaining block Gabor features; and combining the block Gabor features according to scale and direction to obtain an iris texture feature vector.

[0024] Among them, radial normalization refers to the process of expanding the annular iris image into a rectangular image. Since the iris image is affected by factors such as pupil dilation and camera distance, the iris radius may change. Radial normalization interpolates or resamples the iris image in the radial direction and maps it to a rectangular image of fixed size to achieve a normalized representation of the iris size and shape, which is convenient for subsequent feature extraction. Convolution operation is a basic operation in signal processing and is used to extract local features of the signal. For a two-dimensional image signal, the convolution operation performs weighted summation of the image through a sliding convolution kernel to obtain a convolution response map. In Gabor wavelet transform, convolution operation is used to calculate the similarity between the image and the Gabor wavelet basis function at different scales and directions, and to extract the multi-scale and multi-directional texture features of the image. The Gabor wavelet energy image is a modulus value representation of the Gabor wavelet transform coefficients, reflecting the energy distribution of the image at different scales and directions. Since the Gabor wavelet transform coefficients are complex numbers, the modulus operation converts them into real numbers to obtain the Gabor wavelet energy image. Gabor wavelet energy image describes the intensity and directionality of local texture of an image and is an effective method for representing texture features.

[0025] Furthermore, a method based on Zernike moments is used to extract iris shape features of the image after iris segmentation, including: normalizing the coordinates of the segmented iris image, converting the Cartesian coordinate system into a polar coordinate system, and obtaining a normalized iris image; in the polar coordinate system, representing the normalized iris image as a radial distance function; wherein the radial distance function represents a function of the distance from the iris boundary to the iris center varying with the angle; using Zernike polynomial basis functions to expand the radial distance function to obtain Zernike moment coefficients; wherein the Zernike polynomial basis functions are a set of orthogonal basis functions used to describe the radial distance function; selecting preset low-order Zernike moment coefficients as iris shape feature vectors; wherein the preset low-order coefficients correspond to the global features of the iris shape.

[0026] Among them, the radial distance function is a mathematical representation method for describing the shape of the iris boundary. In the polar coordinate system, the radial distance function represents the function of the distance from the iris boundary to the iris center as a function of the angle. By transforming the coordinates of the iris image and converting the Cartesian coordinate system into the polar coordinate system, the radial distance function of the iris boundary can be obtained. The radial distance function describes the shape characteristics of the iris boundary and is the basis for iris shape modeling. The Zernike moment is an image moment representation method based on orthogonal Zernike polynomial basis functions. Through the Zernike polynomial expansion, the radial distance function can be represented as a series of Zernike moment coefficients. Low-order Zernike moments correspond to the global shape features of the image, such as translation, scale, rotation, etc., and high-order moments correspond to local shape details. Presetting low-order Zernike moment coefficients refers to selecting a set of fixed-order (such as the first 10 orders) Zernike moments as iris shape features, which can effectively represent the iris shape and reduce the feature dimension.

[0027] Furthermore, a method based on Markov random field is used to extract iris texture statistical features of the image after iris segmentation, including: dividing the segmented iris image into multiple overlapping local image blocks; establishing a Markov random field model for each local image block; wherein the pixels in the image block are field points, and the correlation between pixels is the interaction between field points; using Gibbs distribution to define the energy function of the Markov random field model; wherein the energy function is composed of pixel grayscale distribution and correlation between pixels; using maximum likelihood estimation method to estimate the parameters of the energy function to obtain Markov random field model parameters; performing statistics on the Markov random field model parameters to obtain texture statistical features of the local image block; wherein the texture statistical features include mean, variance and entropy; cascading the texture statistical features of all local image blocks to obtain an iris texture statistical feature vector.

[0028] Among them, Gibbs distribution is the probability distribution that describes the equilibrium state of a particle system in statistical mechanics, and is also often used to define the energy function of a Markov random field. In the Markov random field model, the Gibbs distribution defines the joint probability distribution of field point configurations, and its probability is exponentially related to the system energy. For a field point configuration x, its Gibbs probability distribution is: Where E(x) is the energy function, T is the temperature parameter, and Z is the partition function. Gibbs distribution has a local Markov property, that is, the state of a field point is only related to the state of its neighboring field points. By defining a suitable energy function, Gibbs distribution can characterize the correlation and constraint relationship between field points and is used for image texture modeling. Maximum likelihood estimation is a commonly used parameter estimation method used to estimate the parameters of a probability model based on observed data. Assuming that the observed data obeys a certain probability distribution, maximum likelihood estimation solves the estimated value of the probability distribution parameters by maximizing the likelihood function of the observed data. For the Markov random field model, maximum likelihood estimation is used to estimate the energy function parameters of the Gibbs distribution. By constructing the likelihood function of the local image block, the energy function parameters are solved using optimization algorithms such as gradient descent and iterative reweighted least squares to obtain the parameter estimate of the Markov random field model. Maximum likelihood estimation has the excellent properties of asymptotic consistency and asymptotic efficiency. Feature concatenation refers to the process of connecting multiple feature vectors into a high-dimensional feature vector in a certain order. In the iris texture statistical feature extraction, the iris image is divided into multiple local image blocks, and the texture statistical features of each image block, such as mean, variance, entropy, etc., are extracted respectively. In order to obtain the texture feature representation of the entire iris image, it is necessary to cascade the feature vectors of all local image blocks to form a feature vector of length n*d, where n is the number of image blocks and d is the feature dimension of each image block. Feature cascading can integrate local texture features into global features and improve the robustness and discriminability of feature representation. The cascaded high-dimensional feature vector can be used for subsequent feature selection, dimensionality reduction, and classifier design.

[0029] Compared with the prior art, the advantages of this application are:

[0030] The use of facial feature extraction and similarity calculation based on convolutional neural networks can overcome the influence of changes in facial posture, lighting, etc., and achieve robust and efficient facial identity verification. When facial verification fails, the introduction of voiceprint and iris dual biometric verification forms multimodal biometric authentication, which greatly improves the security of identity verification.

[0031] In voiceprint recognition, deep neural networks are used for speech denoising, which can effectively remove environmental noise; extracting cepstral coefficient features through linear prediction analysis and obtaining vocal tract frequency features using resonance peak estimation algorithms can accurately characterize individual differences in speech; fusing cepstral coefficients and resonance peak frequency features can obtain a more comprehensive and fine-grained voiceprint feature representation.

[0032] In iris recognition, convolutional neural networks are used to locate the iris, which can accurately obtain the iris area; iris segmentation is achieved through the graph cut algorithm, which can overcome the local optimal problem of traditional methods and obtain the global optimal segmentation result; Gabor wavelets, Zernike moments and Markov random fields are used to extract iris texture, shape and statistical features, which can characterize the uniqueness of the iris at multiple levels and angles to obtain a highly recognizable iris feature representation.

[0033] Using Bayesian network to make fusion decisions on the similarity of voiceprint recognition and iris recognition can make full use of the complementarity of the two biometric features, overcome the defects of a single feature, and improve the reliability and accuracy of identity authentication. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0035] Figure 1 is an exemplary flow chart of a digital e-commerce platform based on cloud computing according to some embodiments of this specification;

[0036] Figure 2 is an exemplary flow chart of the second identity authentication according to some embodiments of this specification;

[0037] Figure 3 is an exemplary flow chart of obtaining speech similarity according to some embodiments of this specification;

[0038] Figure 4 is an exemplary flow chart of obtaining linear prediction cepstral coefficient features according to some embodiments of this specification;

[0039] Figure 5 is an exemplary flow chart of obtaining resonance peak frequency characteristics according to some embodiments of this specification;

[0040] Figure 6 is an exemplary flow chart of obtaining iris recognition similarity according to some embodiments of this specification;

[0041] Figure 7 is an exemplary flow chart of obtaining a segmented iris image according to some embodiments of the present specification;

[0042] Figure 8 is an exemplary flow chart of obtaining an iris texture feature vector according to some embodiments of this specification;

[0043] Fig. 9is an exemplary flow chart of obtaining an iris shape feature vector according to some embodiments of this specification;

[0044] Fig.10 This is an exemplary flow chart of obtaining texture statistical features according to some embodiments of this specification. DETAILED DESCRIPTION

[0045] The method and system provided in the embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0046] Figure 1 This is an exemplary flow chart of a digital e-commerce platform based on cloud computing as shown in some embodiments of the present specification. The digital e-commerce platform based on cloud computing includes: collecting multiple frames of facial images; preprocessing the collected multiple frames of facial images; using a pre-trained convolutional neural network to extract features from the preprocessed multiple frames of facial images to obtain feature vectors; calculating the similarity between the obtained feature vectors and feature vectors in a pre-stored authorized personnel facial feature library to obtain a similarity score; if the calculated similarity score is greater than a threshold, the identity authentication is passed and the user is allowed to access the platform; if the calculated similarity score is less than or equal to the threshold, the identity authentication fails; when the identity authentication fails, verification is performed according to a pre-set second identity authentication method.

[0047] When a user accesses the e-commerce platform, the system collects multiple frames of the user's facial images and transmits the collected images to the cloud server. When the user accesses the e-commerce platform through a terminal such as a PC or mobile device, the system automatically calls the camera of the front-end device to collect multiple frames of continuous facial images of the user. To ensure the quality of the collected facial images, the system can guide the user to adjust the head posture so that the face faces the camera and control the lighting conditions to meet the preset requirements. The collection process can last for 2 to 3 seconds, and the camera continuously collects 60 to 90 frames of facial images at a speed of 30 frames per second. The system encodes and compresses the collected original image frames, and securely transmits them to the cloud authentication server through network protocols such as HTTP and TCP. At the same time, the local terminal deletes the original collected images to protect user privacy.

[0048] The cloud server preprocesses the received multi-frame face images and uses the pre-trained convolutional neural network model to extract the face feature vector. After receiving the multi-frame face images, the cloud server first evaluates the image quality and removes the image frames with poor quality such as insufficient lighting, skewness, and blur. Then, the screened qualified image frames are normalized to a uniform size and scaled to 256×256 pixels. Next, the image is normalized using the adaptive histogram equalization algorithm to reduce the impact of lighting changes on feature extraction. Finally, the pre-processed image is input into the pre-trained deep convolutional neural network for feature extraction. The neural network adopts the ResNet-18 architecture, which contains 18 convolutional layers and fully connected layers, and can extract 4096-dimensional face feature vectors. The network model parameters are obtained through pre-training of a large-scale face dataset and have strong feature representation capabilities. For each frame of the image, the network outputs a 4096-dimensional feature vector. The feature vectors of multiple frames are fused into a final 4096-dimensional face feature vector through mean pooling, which serves as the feature representation of the user's identity authentication.

[0049] The cloud server calculates the similarity between the extracted facial feature vector and the pre-stored authorized user facial feature library to obtain a similarity score. The cloud server calculates the similarity between the extracted 4096-dimensional facial feature vector and the pre-stored authorized user facial feature library. The authorized user facial feature library stores the facial feature vectors of all registered users, and each user corresponds to a 4096-dimensional feature vector. The feature library uses relational databases such as MySQL for efficient and secure storage and management. The similarity calculation uses the cosine similarity metric, that is, the cosine similarity between the facial feature vector to be verified and the feature vector of each user in the feature library is calculated. The larger the cosine similarity, the closer the two feature vectors are, and the more likely the corresponding facial identity is to match. Suppose the facial feature vector to be verified is A, and the feature vector of a user in the feature library is B, then the cosine similarity calculation formula for the two vectors is: Among them, A⊙B represents the inner product of vectors A and B, and ||A|| and ||B|| represent the L2 norm of vectors A and B respectively. In order to improve the efficiency of feature retrieval and reduce the computational complexity under the scale of massive users, the system adopts a feature dimension reduction method based on PCA (Principal Component Analysis) to reduce the dimension of the 4096-dimensional feature vector to 128 dimensions to accelerate the subsequent similarity calculation. After calculating the cosine similarity between the facial features to be verified and the features of all registered users, the maximum similarity value is taken as the identity similarity score of the current face, and the corresponding user ID is recorded at the same time. Let the maximum similarity score be max_score and the corresponding user ID be max_user_id, then the current face identity authentication result can be expressed as a binary form of (max_user_id, max_score). The system returns the identity authentication result to the e-commerce platform for subsequent identity confirmation and authorization judgment. In order to balance security and user experience, the system presets a similarity score threshold. If max_score>threshold, the identity authentication is considered to be successful. Otherwise, when max_score≤threshold, the identity authentication is considered to have failed, and the secondary verification process of voiceprint + iris needs to be started. The threshold can be flexibly set according to the actual business security requirements. The higher the threshold, the higher the accuracy of identity authentication, but the user pass rate may be reduced; the lower the threshold, the higher the user pass rate, but there is a certain security risk.

[0050] Figure 2This is an exemplary flow chart of the second identity authentication shown in some embodiments of this specification. If the similarity score exceeds the preset threshold, the identity authentication is passed and the user can access the platform normally; otherwise, the identity authentication fails and the second identity authentication process is triggered. The cloud server obtains the similarity score max_score and the preset threshold threshold of the user identity authentication. The server determines the identity authentication result according to the size relationship between the two: if max_score>threshold, the user identity authentication is considered to be passed. The cloud server signs and encrypts the verification result (including user ID, timestamp and other information) and returns it to the e-commerce platform. After receiving the verification result, the e-commerce platform allows the user to access platform resources and services normally, including but not limited to browsing products, adding shopping carts, placing orders, and paying. At the same time, the e-commerce platform records the user's login log for subsequent audits and security analysis. If max_score≤threshold, the user identity authentication is considered to have failed. This means that the facial features submitted by the user have a low similarity with all user features in the registered feature library, and the user's identity cannot be confirmed. At this time, the cloud server returns the verification failure result to the e-commerce platform and triggers the second identity authentication process of voiceprint + iris. After receiving the verification failure result, the e-commerce platform rejects the user's access request and pops up a secondary identity verification prompt interface to guide the user to complete the subsequent voiceprint and iris collection.

[0051] Figure 3This is an exemplary flow chart for obtaining voice similarity according to some embodiments of this specification. In the second identity authentication process, the system prompts the user to enter voice and iris. The user's voice data and iris image are collected and uploaded to the cloud. When the user's identity authentication fails, the e-commerce platform automatically pops up the identity secondary authentication page, prompting the user to perform voiceprint and iris dual authentication. For voice data collection, the system generates a short text of 8-10 Chinese characters through a random algorithm, and displays the text on the verification page, requiring the user to read the text aloud. At the same time, the system calls the device microphone to record the user's voice at a sampling rate of 44.1kHz and a 16-bit quantization bit, and the voice recording time is controlled within 5 seconds. After the recording is completed, the system encodes and compresses the voice data, uploads it to the cloud through the HTTPS protocol, and the local terminal deletes the original recording data. For iris image acquisition, the system guides the user to aim his eyes at the camera through the verification page to take a high-definition iris image. To ensure the quality of acquisition, the camera is equipped with an infrared fill light, which illuminates the eyes with 850nm near-infrared light to highlight the details of the iris texture. The resolution of the original iris image collected shall not be less than 1000×1000 pixels. After shooting, the iris image is compressed and encoded and uploaded via the HTTPS protocol, and the local terminal deletes the original image. The entire collection process should be completed within 5 seconds to avoid discomfort caused by users staring at the infrared light source for a long time. After the voice data and iris image are collected, the e-commerce platform will send a secondary verification request to the cloud server, attaching metadata such as user ID and timestamp, and requesting to start the voiceprint + iris identity verification process.

[0052] After the cloud server receives the voice data uploaded by the user, it first uses a voice noise reduction algorithm based on a deep neural network to denoise the voice. The algorithm uses a pre-trained convolutional neural network model to perform convolution operations on the logarithmic amplitude spectrum of the voice signal in the frequency domain to identify and suppress noise components, thereby obtaining a pure voice signal. After denoising, the system performs dual-threshold energy endpoint detection on the voice data, automatically identifies the starting and ending points of the valid voice segment, removes the silent segments before and after the voice, and obtains the pre-processed voice data.

[0053] Figure 4This is an exemplary flow chart for obtaining linear prediction cepstral coefficient features according to some embodiments of this specification. The system uses a linear prediction analysis method to extract the linear prediction cepstral coefficient (LPCC) features of the pre-processed speech data. The specific steps are as follows: The pre-processed speech data is framed, the frame length is set to 25ms, and the frame shift is set to 10ms to obtain a speech frame sequence. A Hamming window is added to each frame of speech data to reduce spectrum leakage. The windowed speech frame is sw(n)=s(n)×w(n), where w(n) is the Hamming window function. The autocorrelation function is calculated for the windowed speech frame to obtain P+1 autocorrelation function coefficients R(k), k=0, 1,..., P, P is the linear prediction order, usually 12. According to the autocorrelation function coefficients, the Levinson-Durbin recursive algorithm is used to calculate the P-order linear prediction coefficients a_i, i=1, 2,..., P. The linear prediction coefficients a_i are subjected to a fast Fourier transform (FFT) to obtain the prediction coefficients A(ω) in the frequency domain. The frequency domain prediction coefficients are logarithmized and inverse Fourier transformed to obtain the cepstral coefficients c(n), i.e., LPCC features. The 12-dimensional LPCC features of each frame of speech are extracted and concatenated with the first-order and second-order differential parameters to obtain a 36-dimensional LPCC feature vector.

[0054] Figure 5 This is an exemplary flow chart for obtaining the resonance peak frequency feature according to some embodiments of this specification. The system uses the resonance peak estimation algorithm to extract the resonance peak frequency feature of the speech data. The specific steps are as follows: Using the linear prediction coefficient a_i, calculate the linear prediction spectrum H(ω). Use the triangular window function to smooth the linear prediction spectrum, remove the detail fluctuations in the spectrum peak, and obtain the smoothed spectrum H_s(ω). Obtain the local maximum point in H_s(ω) as the initial resonance peak frequency F_i, i=1, 2,..., L, L is the number of initial resonance peaks. Perform harmonic detection on F_i to determine whether it is an integer multiple of the fundamental frequency, and remove the harmonic peaks therein. Obtain the optimized resonance peak frequency F_oi. Select the first 4 resonance peak frequencies with the highest energy in F_oi as the resonance peak frequency feature. Perform 512-point FFT on the resonance peak frequency feature to obtain the frequency domain resonance peak feature R_FFT. Use 24 triangular Mel filters to filter R_FFT to obtain the 24-dimensional Mel frequency domain resonance peak feature R_Mel. R_Mel is logarithmically transformed and 12-point discrete cosine transform (DCT) is performed to obtain 12-dimensional formant Mel frequency cepstral coefficient (MFCC) features. The 12-dimensional formant MFCC features are concatenated with their first-order and second-order difference parameters to obtain a 36-dimensional formant frequency feature vector.

[0055] The obtained 36-dimensional LPCC features and the obtained 36-dimensional formant frequency features are concatenated to form a 72-dimensional voiceprint feature vector. The voiceprint feature vector is input into a pre-trained deep neural network model, and the network outputs the similarity score between the current voiceprint and the registered user's voiceprint. The network adopts a deep embedding learning architecture, which can map heterogeneous features of different dimensions into a common feature subspace for similarity measurement to improve the accuracy of voiceprint verification. The network training uses a large-scale multilingual speech dataset and is optimized using a contrastive loss function to minimize the distance between the voiceprint features of the same speaker embedded in the feature space and maximize the feature embedding distance of different speakers. Finally, the output layer uses a Sigmoid function to obtain a voiceprint similarity score between 0 and 1 as the credibility basis for voiceprint verification.

[0056] Figure 6 This is an exemplary flow chart of obtaining iris recognition similarity according to some embodiments of this specification. After the cloud server receives the iris image uploaded by the user, it first uses a pre-trained convolutional neural network model to locate the iris of the image. The network model is based on the U-Net architecture, extracts multi-scale features of the image through the encoder, and restores spatial details through the decoder to achieve end-to-end iris positioning. Figure 7 This is an exemplary flow chart for obtaining a segmented iris image according to some embodiments of the present specification. The positioning result is a compact rectangular area that fully covers the iris texture area. Within the iris area, the system uses an image segmentation algorithm to extract an accurate iris area. The specific method is: the iris area is represented as a weighted undirected graph G = (V, E), where the node set V is a pixel set, and the edge set E is the connection relationship between adjacent pixels. The edge weight w(i, j) represents the similarity between pixels i and j, and is calculated using a Gaussian function: w(i, j) = exp(-||I(i)-I(j)||^2 / (2σ^2)), where I(i) is the grayscale value of pixel i, and σ is a scale parameter. Construct the energy function E(f) of the graph, where f is the pixel label (foreground / background) configuration. The energy function consists of the regional term R(f) and the boundary term B(f): E(f) = λR(f) + B(f), where λ is a balancing factor, R(f) considers the posterior probability that the pixel belongs to the iris / non-iris region, and B(f) considers the consistency of the labels of adjacent pixels. The maximum flow minimum cut algorithm is used to solve the energy function globally to obtain the segmentation curve C. So that: minE(f) = min{λR(f) + B(f)}, the iris region is segmented according to the optimal segmentation curve C, the inside of C is iris pixels, and the outside is non-iris pixels. The segmented binary iris image I_seg is obtained.

[0057] Figure 8This is an exemplary flow chart for obtaining iris texture feature vectors according to some embodiments of this specification. Iris texture feature extraction, the system uses Gabor wavelet transform to extract iris texture features. Gabor wavelet is a type of bandpass filter that simulates the human visual system and can effectively capture multi-directional and multi-scale texture information. Construct a Gabor wavelet basis function ψ(x, y; λ, θ, ψ, σ, γ), where λ is a scale parameter, θ is a direction parameter, ψ is a phase offset, σ is a Gaussian kernel function standard deviation, and γ is an aspect ratio. The segmented iris image I_seg is radially normalized, and the annular iris is expanded into a rectangular image I_norm. Convolution operation is performed on I_norm using the Gabor wavelet basis function: G(x, y; λ, θ) = I_norm(x, y)*ψ(x, y; λ, θ, ψ, σ, γ), and the Gabor wavelet transform coefficient G is obtained, and the Gabor wavelet energy image E_mag is obtained by modulo. E_mag is processed in blocks, and the mean μ and standard deviation σ of each block are calculated to form block Gabor features (μ, σ). The block Gabor features at different scales and directions are cascaded to obtain the iris texture feature vector f_texture. Taking 8 directions and 5 scales, the feature dimension is 40.

[0058] Fig. 9 This is an exemplary flow chart for obtaining iris shape feature vectors according to some embodiments of this specification. For iris shape feature extraction, the system uses a method based on Zernike moments to extract iris shape features. Zernike moments are a shape descriptor based on orthogonal basis functions, which remain invariant to rotation and scale changes. The segmented iris image I_seg is normalized, and the Cartesian coordinates (x, y) are mapped to polar coordinates (ρ, θ) to obtain a normalized iris image I_polar. In polar coordinates, the iris boundary can be represented by a radial distance function r(θ), that is, the relationship between the distance from the boundary point to the iris center O and the angle θ. The Zernike polynomial basis function V_nm(ρ, θ) is used to expand r(θ) to obtain the Zernike moment coefficient Z_nm: Z_nm=(n+1) / π∫∫_D[V_nm(ρ, θ)]*r(θ)ρdρdθ, where n is the order, m is the number of repetitions, and * represents the complex conjugate. Select low-order Zernike moment coefficients (such as order n=1-12) to form the iris shape feature vector f_shape. Extract the first 30 moments to obtain a 30-dimensional feature vector.

[0059] Fig.10This is an exemplary flow chart for obtaining texture statistical features according to some embodiments of this specification. For iris statistical feature extraction, the system uses a method based on Markov random field (MRF) to extract the statistical features of iris texture. The MRF model takes into account the spatial correlation between pixels and can characterize the local statistical characteristics of texture. The segmented iris image I_seg is divided into K overlapping local image blocks, each of which is S×S in size. An MRF model is established for each image block, and the pixel grayscale value is the observed state, assuming that it obeys the Gibbs distribution. The MRF energy function E is: E=Σ_iV_1(y_i)+Σ_iΣ_j∈N_iV_2(y_i, y_j), where V_1 is the unary potential function of the pixel point yi, which is related to the pixel grayscale distribution; V_2 is the binary potential function of the pixel point yi and its neighboring pixels yj, reflecting the spatial correlation between pixels. Given the observed pixel value, the maximum likelihood estimation is used to solve the MRF model parameter θ to obtain the probability distribution P(y; θ) of the local texture. Statistical features are extracted from P(y;θ), including mean, variance, entropy, etc. The statistical features of K image blocks are cascaded to obtain the iris statistical feature vector f_stat with a dimension of K×3.

[0060] The extracted iris texture feature f_texture, iris shape feature f_shape and iris statistical feature f_stat are concatenated to obtain the final iris feature vector: f_iris = [f_texture, f_shape, f_stat]. Considering that the scales and distributions of different types of features may vary greatly, the feature vector needs to be normalized to balance the contribution of each feature subspace. Common normalization methods are: Min-Max normalization: linearly scale the feature to the interval [0, 1]. Let the i-th dimension component of the feature vector be x_i, then the normalized component x_i' is: x_i' = (x_i-min(x)) / (max(x)-min(x)), Z-score normalization: subtract the mean from the feature component and divide it by the standard deviation to make it obey the standard normal distribution N(0, 1). The normalized component x_i' is: x_i' = (x_i-μ) / σ, where μ is the mean of the feature component and σ is the standard deviation. L2 normalization: Divide the eigenvector by its L2 norm to make it a unit vector. The normalized eigenvector f' is: f' = f / ||f||_2, where ||*||_2 represents the L2 norm, and ||f||_2 = sqrt(Σ_ix_i^2).

[0061] After normalization, the final iris feature vector f_iris is obtained. The Hamming distance is used as the similarity metric, and the iris feature to be verified is compared with the iris features in the registration library one by one to obtain the similarity score. Suppose the iris feature to be verified is f_1, and the i-th iris feature in the registration library is f_2^i. The Hamming distance d_i between the two features is defined as: d_i = (f_1XORf_2^i) / N, where XOR is an exclusive OR operation and N is the dimension of the feature vector. The value range of d_i is [0, 1], which indicates the dissimilarity of the two binary feature vectors. The smaller the distance, the more similar they are. By setting the similarity threshold τ, it can be determined whether the iris to be verified is a registered iris. If min(d_i)<τ, it is considered that the iris to be verified matches an iris in the registration library and passes the verification; otherwise, the verification fails. The selection of the threshold τ requires a balance between the security and convenience of the iris recognition system, and can be tuned on the development set through methods such as cross-validation.

[0062] To further improve the accuracy and efficiency of iris recognition, some optimization strategies can be adopted, such as: Quality assessment: Before feature extraction, the quality of the collected iris images is assessed to remove low-quality samples such as blur, occlusion, and reflection to improve recognition performance. Feature selection: Through feature sorting, subset search and other methods, the most discriminative iris feature subset is screened to reduce feature dimension and matching complexity. Index acceleration: Use data structures such as hash index and tree index to quickly retrieve high-dimensional iris features and shorten recognition delay. Multimodal fusion: Combine multiple biometric features such as iris, fingerprint, and face for fusion recognition to further improve system security.

[0063] The cloud server inputs the voice similarity and iris similarity into the Bayesian network fusion model, and combines the two biometric verification results to obtain the user identity authenticity verification result. On the basis of the static Bayesian network, the time dimension is introduced to construct a dynamic Bayesian network (DBN) model. DBN represents random variables that change over time through time slices, and time slices are connected by state transition probabilities to form a temporal relationship. Each time slice t contains 5 nodes: voice recognition similarity node V^t, iris recognition similarity node I^t, voice credibility node T_V^t, iris credibility node T_I^t and identity verification result node R^t. The dependency relationship between the nodes is as follows: V^t and I^t are observation nodes, indicating the similarity of voice and iris recognition at time t; T_V^t and T_I^t are latent variable nodes, indicating the credibility of the voice and iris recognition results at time t, with a value range of {0, 1}; R^t is a latent variable node, indicating the true result of identity authentication at time t, with a value range of {0, 1}, where 0 indicates identity authentication failure and 1 indicates pass; R^t depends on T_V^t and T_I^t, reflecting the impact of voice and iris credibility on identity authentication results; there is a conditional dependency between T_V^t and T_I^t, reflecting the correlation between the two biometric verification processes; there is a state transfer dependency between nodes with the same name in adjacent time slices, such as R^(t-1) to R^t, reflecting the dynamic changes in identity authenticity.

[0064] Model parameter learning, based on user historical verification data D = {(V_n, I_n, R_n)}, uses the expectation-maximization (EM) algorithm to train the parameters of DBN, including: the conditional probability distribution (CPD) of each node, such as P(R^t|T_V^t, T_I^t), which represents the probability distribution of identity authentication results under the condition of given voice and iris credibility; state transition probability distribution, such as P(R^t|R^(t-1)), which represents the transition probability of identity authentication results between the previous and next two time slices. The EM algorithm estimates parameters in an iterative manner. Each iteration includes two steps: Step E: Estimate the posterior probability distribution of hidden variables based on the current parameters, that is, P(T_V^t, T_I^t, R^t|V_n, I_n); Step M: Update the model parameters based on the posterior probability of hidden variables, maximizing the likelihood function Q(θ|θ^old)=ΣP(T_V^t, T_I^t, R^t|V_n, I_n;θ^old)logP(V_n,I_n,T_V^t,T_I^t,R^t;θ). In Step E, the forward-backward algorithm is used to calculate the posterior probability of hidden variables. The algorithm recursively calculates the forward probability α and the backward probability β to obtain the joint posterior probability γ. In Step M, the Junction Tree algorithm is used for precise reasoning, and the marginal probability of hidden variables is calculated through message passing, and the CPD and transition probability are updated.

[0065] Authentication reasoning, in the authentication stage, obtain the voice recognition similarity {V_1, ..., V_K} and iris recognition similarity {I_1, ..., I_K} of the user's K consecutive interactions, and construct the observation sequence O = {(V_1, I_1), ..., (V_K, I_K)}. Use the Viterbi algorithm to perform reasoning on DBN to find the most likely hidden state sequence H* = {(T_V^1*, T_I^1*, R^1*), ..., (T_V^K*, T_I^K*, R^K*)}. The Viterbi algorithm uses dynamic programming ideas to recursively calculate the optimal hidden state of each time slice and finally obtain the global optimal path. According to the authentication result R^K* of the last time slice, combined with the global threshold τ and decision rules, it is determined whether the user identity is verified: if P(R^K*=1|O)≥τ, the user identity is considered to be authenticated; otherwise, the identity authentication is considered to fail and the user access request is rejected.

[0066] Decision optimization and model update. After obtaining the identity authentication result, corresponding processing is performed for different situations: S Identity authentication passed: Add the current verification data (V_K, I_K, R_K) to the user's historical data set D, and use the incremental learning algorithm to update the parameters of DBN. Incremental learning makes the model adapt to changes in user behavior patterns and improves verification accuracy by locally adjusting CPD and transition probability. Identity authentication failed: Record suspicious behaviors, including interaction time, device information, IP address, etc., for subsequent security audit analysis. At the same time, dynamically adjust the global threshold τ and decision rules to improve the ability to distinguish suspicious users. When optimizing decisions, comprehensively consider factors such as historical verification results and the degree of abnormal user behavior, and use a weighted voting mechanism to dynamically adjust the identity pass standard. Design an adaptive weight allocation strategy to assign different credibility weights to verification results from different sources to enhance the robustness of the system.

[0067] Based on the identity authenticity verification result, the cloud server returns the identity authentication result to the e-commerce platform. If the identity authentication is passed, the user is allowed to continue to access the platform service; otherwise, the user request is rejected and the preset exception handling process is initiated.

Claims

1. A digital e-commerce platform based on cloud computing, including: Collect multiple frames of face images; Preprocessing the collected multiple frames of face images; A pre-trained convolutional neural network is used to extract features from the pre-processed multi-frame face images to obtain feature vectors; The obtained feature vector is used to calculate the similarity with the feature vector in the pre-stored authorized personnel face feature library to obtain a similarity score; If the calculated similarity score is greater than the threshold, the identity verification is passed and the user is allowed to access the platform; If the calculated similarity score is less than or equal to the threshold, the identity verification fails; When identity authentication fails, verification is performed according to a pre-set second identity authentication method.

2. The digital e-commerce platform based on cloud computing according to claim 1, characterized in that: The second identity verification method includes: Extract user voice data, use deep neural network model to perform voiceprint recognition, and obtain voice recognition similarity; Collect the user's iris image, extract the iris features, calculate the similarity with the pre-stored iris features, and obtain the iris recognition similarity; Input the voice recognition similarity and iris recognition similarity into the Bayesian network fusion algorithm to obtain the user identity authenticity verification result; Control user access to the platform based on the results of user identity authenticity verification.

3. The digital e-commerce platform based on cloud computing according to claim 2, characterized in that: Get the speech recognition similarity, including: Collect user voice data; Use a speech noise reduction algorithm based on a deep neural network to denoise the collected user speech data; Performing endpoint detection and silence removal on the denoised speech data to obtain pre-processed speech data; The linear prediction analysis method is used to extract the linear prediction cepstral coefficient features of the preprocessed speech data; The formant estimation algorithm is used to extract the formant frequency characteristics of the preprocessed speech data; The extracted linear prediction cepstrum system features and formant frequency features are concatenated to obtain a voiceprint feature vector; The obtained voiceprint feature vector is used as input, and the pre-trained deep neural network model is used to obtain the speech similarity.

4. The digital e-commerce platform based on cloud computing according to claim 3 is characterized by: The linear prediction analysis method is used to extract the linear prediction cepstral coefficient features of the preprocessed speech data, including: Performing frame processing on the preprocessed speech data to obtain a speech frame sequence; A Hamming window function is used to perform windowing processing on the speech frame sequence to obtain a windowed speech frame; Calculating the autocorrelation function of the windowed speech frame to obtain the autocorrelation function coefficient; According to the autocorrelation function coefficient, the Levinson-Durbin recursive algorithm is used to calculate the linear prediction coefficient; Perform fast Fourier transform on the linear prediction coefficients to obtain frequency domain prediction coefficients; Take the logarithm and inverse Fourier transform of the frequency domain prediction coefficients to obtain the cepstrum coefficients; The multidimensional cepstral coefficients are extracted as linear prediction cepstral coefficient features.

5. The digital e-commerce platform based on cloud computing according to claim 4 is characterized by: The formant estimation algorithm is used to extract the formant frequency characteristics of the preprocessed speech data, including: Calculate the linear prediction spectrum according to the obtained linear prediction coefficient; The linear prediction spectrum is smoothed by using a triangular window function to obtain a smoothed linear prediction spectrum; Obtain the local maximum point of the smoothed linear prediction spectrum as the initial resonance peak frequency; Perform harmonic detection on the initial resonance peak frequency, remove the harmonic peak, and obtain the optimized resonance peak frequency; Obtain the top N resonance peak frequencies with the highest energy among the optimized resonance peak frequencies as resonance peak frequency features; Perform discrete Fourier transform on the resonance peak frequency characteristics to obtain the resonance peak characteristics in the frequency domain; Performing triangular Mel filter bank filtering on the frequency domain resonance peak features to obtain the Mel frequency domain resonance peak features; Perform logarithmic transformation and discrete cosine transformation on the Mel frequency domain resonance peak features to obtain the MFCC features of the resonance peak; The MFCC features of the resonance peak are concatenated with the corresponding first-order and second-order difference parameters to obtain the final resonance peak frequency features.

6. The digital e-commerce platform based on cloud computing according to claim 2, characterized in that: Get iris recognition similarity, including: The iris detection algorithm based on convolutional neural network is used to locate the iris in the iris image and obtain the iris area; An image segmentation algorithm is used to segment the iris region to obtain a segmented iris image; wherein the image segmentation algorithm constructs an energy function of the iris region to obtain a global optimal segmentation curve to perform iris segmentation; Gabor wavelet transform method is used to extract iris texture features of iris segmentation image; The iris shape features of the iris segmented image are extracted using a method based on Zernike moments. The Markov random field based method is used to extract the iris texture statistical features of the iris segmented image. The extracted iris texture features, iris shape features and iris texture statistical features are fused to obtain an iris feature vector; The iris feature vector is taken as input and the Hamming distance metric is used to calculate the iris recognition similarity.

7. The digital e-commerce platform based on cloud computing according to claim 6, characterized in that: An image segmentation algorithm is used to segment the iris region to obtain a segmented iris image. The image segmentation algorithm constructs an energy function of the iris region to obtain a global optimal segmentation curve to perform iris segmentation, including: The iris region is represented as a weighted undirected graph, where the nodes of the graph are the pixels of the iris region and the weights of the edges are the similarities between adjacent pixels. Construct an energy function of a weighted undirected graph, which includes a region term and a boundary term. The region term represents the probability that a pixel belongs to an iris region or a non-iris region, and the boundary term represents the probability that the labels of adjacent pixels are consistent. The maximum flow minimum cut algorithm is used to minimize the energy function and obtain the global optimal segmentation curve; The iris region is segmented according to the global optimal segmentation curve to obtain a segmented iris image; wherein the inside of the segmentation curve is the iris region, and the outside is the non-iris region.

8. The digital e-commerce platform based on cloud computing according to claim 7 is characterized by: The Gabor wavelet transform method is used to extract the iris texture features of the iris segmented image, including: Constructing Gabor wavelet basis function, wherein the parameters of Gabor wavelet include scale, direction, center frequency and bandwidth; Performing radial normalization processing on the segmented iris image to convert the annular iris image into a rectangular image; The constructed Gabor wavelet basis function is used to perform convolution operation on the obtained rectangular image to obtain Gabor wavelet transform coefficients; Modulo the Gabor wavelet transform coefficients to obtain the Gabor wavelet energy image; The Gabor wavelet energy image is divided into blocks, and the mean and variance of each block are calculated to obtain the block Gabor features; The block Gabor features are combined according to scale and direction to obtain the iris texture feature vector.

9. The digital e-commerce platform based on cloud computing according to claim 8, characterized in that: The iris shape features of the iris segmented image are extracted using a method based on Zernike moments, including: Normalizing the coordinates of the segmented iris image, converting the Cartesian coordinate system into a polar coordinate system, and obtaining a normalized iris image; In the polar coordinate system, the normalized iris image is represented as a radial distance function; wherein the radial distance function represents a function of the distance from the iris boundary to the iris center varying with the angle; The radial distance function is expanded using Zernike polynomial basis functions to obtain Zernike moment coefficients; wherein the Zernike polynomial basis functions are a set of orthogonal basis functions used to describe the radial distance function; A preset low-order Zernike moment coefficient is selected as an iris shape feature vector; wherein the preset low-order coefficient corresponds to a global feature of the iris shape.

10. The digital e-commerce platform based on cloud computing according to claim 9, characterized in that: The iris texture statistical features of the iris segmented image are extracted using a Markov random field-based method, including: Dividing the segmented iris image into a plurality of overlapping local image blocks; A Markov random field model is established for each local image block, wherein the pixels in the image block are field points, and the correlation between pixels is the interaction between field points; The energy function of the Markov random field model is defined using Gibbs distribution; the energy function is composed of pixel grayscale distribution and correlation between pixels; The maximum likelihood estimation method is used to estimate the parameters of the energy function and obtain the parameters of the Markov random field model; The Markov random field model parameters are counted to obtain the texture statistical features of the local image block; wherein the texture statistical features include mean, variance and entropy; The texture statistical features of all local image blocks are cascaded to obtain the iris texture statistical feature vector.

Images