Face recognition system and method based on AI technology

Abstract

The application provides a face recognition system and method based on AI technology, and belongs to the field of image recognition.The system is suitable for security monitoring of smart communities or smart villages, and effectively improves the implementation efficiency and accuracy of security protection by realizing accurate face recognition and behavior analysis.The system mainly utilizes an incremental convolutional neural network and a sparse autoencoder.The whole system has the characteristics of stability, rapidity and accuracy, and can be widely used in the fields of security monitoring of smart communities and smart villages.The application solves some main problems in the application of face recognition technology, and provides a new method for effectively training on large-scale and dynamic data sets.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, and more specifically relates to a face recognition system based on AI technology. Background Technology

[0002] In today's society, smart communities or smart villages have become an important component of intelligent city construction. Among these, security monitoring systems, as a crucial element of this construction, play an irreplaceable and key role. Facial recognition technology, as the core technology of security monitoring systems, has been widely applied in various fields such as access control, real-time monitoring, abnormal event early warning, and tracking of lost elderly people and children.

[0003] In smart communities and smart villages, numerous surveillance devices are distributed throughout the area, operating 24 / 7 to acquire massive amounts of image information. Based on this information, facial recognition technology can help administrators understand the security situation within the community / village in real time and react promptly. In some cases, it can even analyze pedestrian behavior patterns and paths to provide early warnings and address potential security issues, significantly improving the efficiency of community / village security management.

[0004] However, while existing facial recognition technology can meet the needs in some scenarios, it still exposes many problems and challenges in large-scale and dynamically changing smart community and smart village environments. Issues such as large data volumes, complex environments, and constantly changing facial information place high demands on the accuracy and stability of facial recognition technology. Furthermore, how to process massive amounts of image information and correctly identify faces in the shortest possible time is also a problem that current technology urgently needs to solve. Summary of the Invention

[0005] This invention utilizes an incremental convolutional neural network and a sparse autoencoder to design a novel face recognition system, which aims to effectively handle large-scale dynamic data processing, improve face recognition accuracy, mine the internal structure of data, and adapt well to environmental changes. It is suitable for deployment in large-scale practical scenarios such as smart communities and smart villages.

[0006] To achieve the above objectives, the present invention employs the following technical solution: the method includes...

[0007] Step 1: First, acquire images for face recognition;

[0008] Step 2: The acquired images are processed by preprocessing algorithms to remove noise and normalize them, so that the images reach the best recognition conditions;

[0009] Step 3: The preprocessed image is input into the Incremental Convolutional Neural Network (InncCNN) to extract features from the image; as the model is trained, InncCNN will continuously update and add new hidden layers.

[0010] Step 4: The extracted high-dimensional features are reduced in dimensionality and encoded using a sparse autoencoder (SAE).

[0011] Step 5: The system matches the dimensionality-reduced encoded features with the existing face feature database to perform face recognition;

[0012] Step 6: If a match is successful, output the corresponding recognition result. If no match is found, output "No match result".

[0013] Furthermore, the feature extraction in step 3 is as follows:

[0014] (1) Convolution operation: Each layer of a convolutional neural network has multiple convolution kernels to perform convolution operations on the input image. Convolution is performed on a 2D image, and the convolution process can be represented as:

[0015] Y {i,j} =σ((k*X) {i,j} +b) (1)

[0016] Where X represents the input image, k represents the convolution kernel, * represents the convolution operation, and (k*X) {i,j} Y represents the convolution result at position (i,j), σ represents the activation function, and Y represents the position (i,j). {i,j} This represents the pixel value at position (i,j) in the generated feature image;

[0017] (2) Use an activation function to process the feature map. The ReLU function is linear at (x>0) and 0 at (x<0);

[0018] g(x) = max(0,x) (2)

[0019] Where g(x) is the output of the activation function and x is the input of the activation function; through the above steps, the input image is transformed into a set of feature maps, thus completing feature extraction.

[0020] Furthermore, the detailed steps for adding a new hidden layer in step 3 are as follows:

[0021] S301 initializes a simple convolutional neural network model, which contains a certain number of convolutional layers and / or fully connected layers;

[0022] S302 trains and initializes the neural network model, calculates the loss function, and updates the weight parameters through backpropagation.

[0023] S303 evaluates the performance of the current network model on the validation set;

[0024] S304 If the performance on the validation set has reached the set target or there is still no significant improvement after a certain number of epochs of training, then the network capacity (i.e. model complexity) is insufficient and new hidden layers need to be added; otherwise, training continues.

[0025] S305 adds new layers: Based on the existing network, new hidden layers are added at appropriate locations; the new layers are either convolutional layers or fully connected layers.

[0026] S306 Repeat: Use the new network with added hidden layers as the current network and repeat the process of S302-S305.

[0027] Furthermore, the detailed steps of step 4, feature encoding, are as follows: The feature encoding stage mainly uses a sparse autoencoder (SAE) to reduce the dimensionality and encode high-dimensional features;

[0028] The S401 encoding process inputs the high-dimensional features extracted from the face into the SAE, and after passing through an encoding function f, obtains a low-dimensional hidden representation h; this process is represented as follows:

[0029] h=f(W*x+b) (3)

[0030] Where W is the weight matrix, b is the bias vector, f is the non-linear activation function, x is the input face feature, and h is the encoded low-dimensional feature;

[0031] S402 Decoding Process: The hidden representation h is further processed by the decoding function g to obtain a reconstructed feature x'. This process is represented as follows;

[0032] x'=g(W'*h+b') (4)

[0033] Where W' is the weight matrix of the decoder, b' is the bias vector of the decoder, g is the activation function of the decoder, and x' is the reconstructed feature after decoding;

[0034] S403 optimization process: SAE updates W, b, W', b' by optimizing the objective function J to further improve the coding effect. The optimization objective of the sparse autoencoder includes two modules: one is to minimize the reconstruction error, and the other is to satisfy the sparsity constraint, which can be expressed by mathematical formula as follows:

[0035] J(W,b,W',b')=Σ||x-x'|| 2 +λ*Σ|h j | (5)

[0036] Where, Σ||x-x'|| 2This is the reconstruction error term, representing the MSE error between the original input x and the reconstructed input x'; λ*Σ|h j | represents the sparsity constraint, which is the sum of the absolute values ​​of the hidden nodes. λ is a parameter that controls the sparsity penalty, and j is the index of the hidden layer node. The objective task is to iteratively update the values ​​of W and b using optimization methods to minimize the objective function J.

[0037] Furthermore, step 5, matching and identification, specifically includes the following steps:

[0038] S501 feature comparison compares the features encoded by a sparse autoencoder (SAE) with a recorded facial feature database; it uses cosine similarity calculation between feature vectors.

[0039] The formula can be expressed as:

[0040]

[0041] Where X·Y represents the dot product of X and Y, and ||X|| and ||Y|| represent the magnitudes of X and Y, respectively, obtained by calculating the Euclidean distance of each vector;

[0042] For Euclidean distance, the formula can be expressed as:

[0043]

[0044] Where X represents the facial feature to be identified, Y represents a facial feature in the facial feature database, i represents the dimension of the feature vector, and Σ i This indicates summing over all dimensions;

[0045] S502 Threshold Calculation: Set a threshold T. When the Euclidean distance or cosine similarity between features is higher than this threshold, the face to be identified is the same person as a face in the face database.

[0046] Furthermore, an AI-based facial recognition system, the system being applicable to the method described, the system comprising:

[0047] Data input module, data preprocessing module, incremental convolutional neural network IncCNN module, sparse autoencoder (SAE) module, feature matching module, output module;

[0048] The data input module is connected to the data preprocessing module: the face image data collected by the data input module is sent to the data preprocessing module for preprocessing;

[0049] The data preprocessing module is connected to the incremental convolutional neural network IncCNN module: the preprocessed image data is input into IncCNN for feature extraction;

[0050] The incremental convolutional neural network IncCNN module is connected to the sparse autoencoder (SAE) module: the features extracted by IncCNN will be entered into SAE for dimensionality reduction and encoding;

[0051] The sparse autoencoder (SAE) module is connected to the feature matching module: the dimensionality-reduced encoded features output by the SAE will be used by the feature matching module to match with an existing face feature database.

[0052] The feature matching module is connected to the output module: the feature matching module determines the output content of the output module based on the matching result. If a matching feature is found, the corresponding recognition result is output; if no matching feature is found, no matching result is output.

[0053] Beneficial effects of this invention:

[0054] Improving recognition efficiency and accuracy: The combination of incremental convolutional neural networks and sparse autoencoders provides a novel solution for face recognition, improving the accuracy of face recognition and enhancing recognition efficiency, adapting to large-scale and dynamic datasets.

[0055] It provides positive support for the management of smart communities and smart villages: The facial recognition system based on this invention can not only improve the accuracy of resident identification, but also assist in the analysis of abnormal behavior, thereby improving the security level and management level of communities and villages.

[0056] A good balance between efficiency and accuracy: This invention effectively solves the problem that improving recognition accuracy in existing technologies increases computational complexity. The sparse autoencoder can effectively compress features, greatly reducing the amount of computation and improving the speed of operation, enabling the system to achieve relatively fast recognition while maintaining high accuracy. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is a system block diagram of the present invention;

[0059] Figure 3 A new hidden layer step diagram is added to this invention;

[0060] Figure 4 This is a flowchart of the feature encoding process. Detailed Implementation

[0061] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0062] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this invention and in its specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0063] like Figure 1 As shown, the method includes

[0064] Step 1: First, acquire the image that needs to be used for facial recognition and input it into the system.

[0065] Step 2: The acquired images are processed by preprocessing algorithms to remove noise and normalize them, so that the images reach the best recognition conditions.

[0066] Image denoising: During image acquisition and transmission, images may be mixed with various types of noise due to various reasons. Noise will affect the image quality and visual effect, and will also have a significant impact on subsequent processing and analysis. Therefore, denoising is the process of removing or reducing image noise using various algorithms (such as median filtering, Gaussian filtering, etc.). For example, median filtering can be used to denoise images. Median filtering achieves the purpose of noise removal by replacing the gray value of a pixel with the median gray value of its neighborhood. This method is parameterless and can preserve edge information.

[0067] Image normalization: Normalization is the process of compressing the pixel values ​​of an image matrix to 0 to 1 or -1 to 1. Its main purpose is to bring the image to a more standardized state before it enters the model, facilitating model training, improving model convergence speed, and increasing model accuracy. Specifically, it involves subtracting the mean of all image grayscale values ​​from the grayscale value of each pixel, and then dividing by the standard deviation of the grayscale values.

[0068] After these preprocessing steps, the image is in a relatively ideal state and can be further extracted and analyzed.

[0069] Step 3: The preprocessed image is input into the Incremental Convolutional Neural Network IncCNN to extract features from the image; as the model is trained, IncCNN will continuously update and add new hidden layers.

[0070] The feature extraction in step 3 is as follows:

[0071] (1) Convolution operation: Each layer of a convolutional neural network has multiple convolution kernels to perform convolution operations on the input image. Convolution is performed on a 2D image, and the convolution process can be represented as:

[0072] Y {i,j} =σ((k*X) {i,j} +b) (1)

[0073] Where X represents the input image, k represents the convolution kernel, * represents the convolution operation, and (k*X) {i,j} Y represents the convolution result at position (i,j), σ represents the activation function, and Y represents the position (i,j). {i,j} This represents the pixel value at position (i,j) in the generated feature image;

[0074] (2) Use an activation function to process the feature map. The ReLU function is linear at (x>0) and 0 at (x<0);

[0075] g(x) = max(0,x) (2)

[0076] Where g(x) is the output of the activation function and x is the input of the activation function; through the above steps, the input image is transformed into a set of feature maps, thus completing feature extraction.

[0077] like Figure 3 As shown, the detailed steps for adding a new hidden layer in step 3 are as follows:

[0078] S301 initializes a simple convolutional neural network model, which contains a certain number of convolutional layers and / or fully connected layers;

[0079] S302 trains and initializes the neural network model, calculates the loss function, and updates the weight parameters through backpropagation.

[0080] S303 evaluates the performance of the current network model on the validation set;

[0081] S304 If the performance on the validation set has reached the set target or there is still no significant improvement after a certain number of epochs of training, then the network capacity (i.e. model complexity) is insufficient and new hidden layers need to be added; otherwise, training continues.

[0082] S305 adds new layers: Based on the existing network, new hidden layers are added at appropriate locations; the new layers are either convolutional layers or fully connected layers.

[0083] S306 Repeat: Use the new network with added hidden layers as the current network and repeat the process of S302-S305.

[0084] Step 4: The extracted high-dimensional features are reduced in dimensionality and encoded using a sparse autoencoder (SAE).

[0085] like Figure 4 As shown, the detailed steps of step 4, feature encoding, are as follows: The feature encoding stage mainly uses a sparse autoencoder (SAE) to reduce the dimensionality and encode high-dimensional features;

[0086] The S401 encoding process inputs the high-dimensional features extracted from the face into the SAE, and after passing through an encoding function f, obtains a low-dimensional hidden representation h; this process is represented as follows:

[0087] h=f(W*x+b) (3)

[0088] Where W is the weight matrix, b is the bias vector, f is the non-linear activation function, x is the input face feature, and h is the encoded low-dimensional feature;

[0089] S402 Decoding Process: The hidden representation h is further processed by the decoding function g to obtain a reconstructed feature x'. This process is represented as follows;

[0090] x'=g(W'*h+b') (4)

[0091] Where W' is the weight matrix of the decoder, b' is the bias vector of the decoder, g is the activation function of the decoder, and x' is the reconstructed feature after decoding;

[0092] S403 optimization process: SAE updates W, b, W', b' by optimizing the objective function J to further improve the coding effect. The optimization objective of the sparse autoencoder includes two modules: one is to minimize the reconstruction error, and the other is to satisfy the sparsity constraint, which can be expressed by mathematical formula as follows:

[0093] J(W,b,W',b')=Σ||x-x'|| 2 +λ*Σ|h j | (5)

[0094] Where, Σ||x-x'|| 2 This is the reconstruction error term, representing the MSE error between the original input x and the reconstructed input x'; λ*Σ|h j | represents the sparsity constraint, which is the sum of the absolute values ​​of the hidden nodes. λ is a parameter that controls the sparsity penalty, and j is the index of the hidden layer node. The objective task is to iteratively update the values ​​of W and b using optimization methods to minimize the objective function J.

[0095] Step 5: The system matches the dimensionality-reduced encoded features with the existing face feature database to perform face recognition;

[0096] Step 5, matching and identification, specifically includes the following steps:

[0097] S501 feature comparison compares the features encoded by a sparse autoencoder (SAE) with a recorded facial feature database; it uses cosine similarity calculation between feature vectors.

[0098] The formula can be expressed as:

[0099]

[0100] Where X·Y represents the dot product of X and Y, and ||X|| and ||Y|| represent the magnitudes of X and Y, respectively, obtained by calculating the Euclidean distance of each vector;

[0101] For Euclidean distance, the formula can be expressed as:

[0102]

[0103] Where X represents the facial feature to be identified, Y represents a facial feature in the facial feature database, i represents the dimension of the feature vector, and Σ i This indicates summing over all dimensions;

[0104] S502 Threshold Calculation: Set a threshold T. When the Euclidean distance or cosine similarity between features is higher than this threshold, the face to be identified is the same person as a face in the face database.

[0105] Step 6: If a match is successful, output the corresponding recognition result. If no match is found, output "No match result".

[0106] like Figure 2 As shown, an AI-based facial recognition system is described, the system being applicable to the method described above, and the system includes:

[0107] Data input module, data preprocessing module, incremental convolutional neural network IncCNN module, sparse autoencoder (SAE) module, feature matching module, output module;

[0108] The data input module is connected to the data preprocessing module: the face image data collected by the data input module is sent to the data preprocessing module for preprocessing;

[0109] The data acquisition module works as follows:

[0110] Real-time image acquisition: This module captures images in real time using devices such as cameras. When a face appears in the frame, the image is captured immediately. Therefore, the camera used needs to have high resolution to obtain high-quality facial images. Furthermore, depending on the specific scenario and requirements, certain triggering conditions may need to be set, such as acquiring images only within a specified time range, or only acquiring images when a specific action or event is detected.

[0111] Pre-stored image extraction: In some application scenarios, a pre-stored facial image library may be used for recognition. These images may come from a criminal suspect database of public security agencies or a facial photo database of community residents. The data acquisition module needs to provide an interface to extract the required facial images from these image libraries.

[0112] Image preprocessing: After the image is acquired, a series of preprocessing operations are performed, including image noise reduction, background removal, and illumination adjustment, in order to reduce noise and background interference and make the image most suitable for subsequent face recognition operations.

[0113] Face detection and alignment: In the preprocessed image, face detection algorithms are used to locate faces and perform alignment operations, ensuring the faces are centered in the image and vertically aligned. These operations improve the efficiency and accuracy of subsequent face recognition processes.

[0114] Face image saving: After detection and alignment, the face images are saved for subsequent face recognition tasks or as new training data to update the face recognition model.

[0115] The data preprocessing module is connected to the incremental convolutional neural network (IncCNN) module: the preprocessed image data is input into IncCNN for feature extraction;

[0116] The implementation process of the preprocessing module is as follows:

[0117] Image denoising: First, in order to reduce the impact of noise in the image on recognition, some image denoising algorithms are used, such as Gaussian filtering, median filtering or bilateral filtering, to remove random noise in the image, such as salt vibration noise, speckle noise, etc.

[0118] Background removal: Before extracting facial features, it is necessary to remove background information as much as possible, because the background does not help with facial recognition and can even introduce interference. This step is usually accomplished using color space conversion and gradients to make the facial area in the image more prominent.

[0119] Contrast enhancement: Improving the contrast of an image can make facial features more prominent. For example, histogram equalization is a commonly used technique to enhance image contrast.

[0120] Rotation Alignment: Sometimes, for various reasons, the face in an image may not be upright, requiring rotation alignment. This step typically uses an eye-position-based alignment method, which first identifies the position of the eyes and then rotates the image according to the eye position information so that both eyes are on the same horizontal line.

[0121] Cropping and scaling: After the above preprocessing, the image usually needs to be cropped and scaled, that is, only the face region is retained and scaled to a specific size. This reduces the amount of computation and ensures consistency between different images, thereby improving the accuracy of recognition.

[0122] The incremental convolutional neural network (IncCNN) module is connected to the sparse autoencoder (SAE) module: the features extracted by IncCNN will be entered into SAE for dimensionality reduction and encoding;

[0123] The sparse autoencoder (SAE) module is connected to the feature matching module: the dimensionality-reduced encoded features output by the SAE will be used by the feature matching module to match with an existing face feature database.

[0124] The feature matching module is connected to the output module: the feature matching module determines the output content of the output module based on the matching result. If a matching feature is found, the corresponding recognition result is output; if no matching feature is found, no matching result is output.

[0125] Example:

[0126] First, the system receives a facial image from a surveillance camera.

[0127] Then, the system preprocesses the acquired image. Assuming the original grayscale image pixel values ​​range from [0, 255], the system first uses median filtering for noise reduction, and then performs normalization. This process subtracts the average pixel value from the original image for each pixel value, and then divides it by the standard deviation of the pixels. Assuming the normalized image pixel values ​​range from [-1, 1].

[0128] Next, the preprocessed image is input into the trained incremental convolutional neural network IncCNN, which outputs a feature vector containing 256 elements.

[0129] This feature vector is input into a sparse autoencoder (SAE) for feature dimensionality reduction and encoding, and outputs a feature vector containing 128 elements.

[0130] Next, the system matches the dimensionality-reduced encoded features with an existing facial feature database. Assuming the system's feature database contains 10,000 known facial features, all of which have been appropriately labeled and classified, the system finds the known feature that is closest to the new feature by calculating the cosine similarity or Euclidean distance between the new facial feature and these 10,000 features.

[0131] During the recognition phase, a similarity threshold T is first set, assuming T = 0.8. If the highest matching score is greater than 0.8, the system will determine that the new face is the same person as the corresponding face in the database, and then output the information of this person; if no score is greater than 0.8, the system will determine that there is no matching result.

[0132] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0133] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A face recognition method based on AI technology, characterized in that, include: Step 1: First, acquire images for face recognition; Step 2: The acquired images are processed by preprocessing algorithms to remove noise and normalize them, so that the images reach the best recognition conditions; The preprocessing algorithm specifically includes: denoising the image using median filtering; normalizing the image by subtracting the mean of the gray values ​​of the entire image from the gray value of each pixel and then dividing by the standard deviation of the gray values; removing background information before extracting facial features; improving image contrast using histogram equalization; using an eye-position-based alignment method, i.e., first identifying the position of the eyes and then rotating the image according to the eye position information so that the two eyes are on the same horizontal line; retaining only the facial region and scaling it. Step 3: The preprocessed image is input into the Incremental Convolutional Neural Network (IncCNN) for feature extraction. As the model trains, IncCNN continuously updates and adds new hidden layers; specifically, this includes: S301 initializes a simple convolutional neural network model, which contains a certain number of convolutional layers and fully connected layers; S302 trains and initializes the neural network model, calculates the loss function, and updates the weight parameters through backpropagation. S303 evaluates the performance of the current network model on the validation set; S304 If the performance on the validation set has reached the set target or if there is still no significant improvement after a certain number of epochs of training, then the network capacity is deemed insufficient and new hidden layers need to be added; otherwise, training continues. S305 adds new hidden layers at appropriate locations on the basis of the existing network; the new layers are either convolutional layers or fully connected layers. S306 uses the new network with added hidden layers as the current network and repeats the process of S302-S305. Step 4: The extracted high-dimensional features are reduced in dimensionality and encoded using a sparse autoencoder (SAE). The dimensionality reduction and encoding include: The S401 encoding process inputs the high-dimensional features extracted from the face into the SAE, and obtains a low-dimensional hidden representation h through an encoding function f; In the S402 decoding process, the hidden representation h is further processed by the decoding function g to obtain a reconstructed feature x'. In the S403 optimization process, SAE updates W, b, W', b' by optimizing the objective function J to further improve the coding effect. The optimization objective of the sparse autoencoder includes two modules: minimizing the reconstruction error and satisfying the sparsity constraint, which can be expressed by the mathematical formula as follows: （5） in, It is the reconstruction error term, representing the MSE error between the original input x and the reconstructed input x'; λ is a sparsity constraint, representing the summation of the absolute values ​​of the hidden nodes, where λ is a parameter that controls the sparsity penalty and j is the index of the hidden layer node; the objective task is to iteratively update the values ​​of W and b using optimization methods to minimize the objective function J. Step 5: The system matches the dimensionality-reduced encoded features with the existing face feature database to perform face recognition; Step 6: If a match is successful, output the corresponding recognition result; if no match is successful, output no matching result.

2. The face recognition method based on AI technology according to claim 1, characterized in that, The feature extraction in step 3 is as follows: (1) Convolution operation: Each layer of a convolutional neural network has multiple convolution kernels to perform convolution operations on the input image. Convolution is performed on a 2D image, and the convolution process is represented as follows: （1） Where X represents the input image, k represents the convolution kernel, and * represents the convolution operation. This represents the convolution result at position (i, j), where σ represents the activation function. This represents the pixel value at position (i, j) in the generated feature image; (2) Use an activation function to process the feature map. The ReLU function is linear at (x>0) and 0 at (x<0); （2） Where g(x) is the output of the activation function and x is the input of the activation function; through the above steps, the input image is transformed into a set of feature maps, thus completing feature extraction.

3. The face recognition method based on AI technology according to claim 1, characterized in that, Step 5, matching and identification, specifically includes the following steps: S501 feature comparison compares the features encoded by a sparse autoencoder (SAE) with a recorded facial feature database; it uses cosine similarity calculation between feature vectors. The formula is expressed as: （6） in, Represents the dot product of X and Y. and Let X and Y represent the magnitudes, respectively, obtained by calculating the Euclidean distance between each vector; For Euclidean distance, the formula is expressed as: （7） Where X represents the facial feature to be identified, Y represents a facial feature in the facial feature database, and i represents the dimension of the feature vector. This indicates summing over all dimensions; S502 Threshold Calculation: Set a threshold T. When the Euclidean distance or cosine similarity between features is higher than the threshold, the face to be identified is the same person as a face in the face database.

4. A face recognition system based on AI technology, wherein the system is applicable to the method as described in any one of claims 1-3, characterized in that, include: Data input module, data preprocessing module, incremental convolutional neural network IncCNN module, sparse autoencoder (SAE) module, feature matching module, output module; The data input module is connected to the data preprocessing module: the face image data collected by the data input module is sent to the data preprocessing module for preprocessing; wherein, the data input module captures in real time through a camera, and immediately captures the image when a face appears in the frame; and / or, extracts face images from a pre-stored face image library; and, after acquiring the image, performs face detection and alignment operations, that is, positions the face in the center of the image and keeps it in a vertical position; The data preprocessing module is connected to the IncCNN module: the preprocessed image data is input into IncCNN for feature extraction; specifically, the data preprocessing module performs the following: image noise reduction using median filtering or Gaussian filtering; background removal using color space conversion and gradient methods; contrast enhancement using histogram equalization; rotation alignment using an eye-position-based alignment method; and cropping and scaling. The incremental convolutional neural network IncCNN module is connected to the sparse autoencoder (SAE) module: the features extracted by IncCNN will be entered into SAE for dimensionality reduction and encoding; The sparse autoencoder (SAE) module is connected to the feature matching module: the dimensionality-reduced encoded features output by the SAE will be used by the feature matching module to match with an existing face feature database. The feature matching module is connected to the output module: the feature matching module determines the output content of the output module based on the matching result. If a matching feature is found, the corresponding recognition result is output; if no matching feature is found, no matching result is output.

Images