Face recognition method of intelligent robot

Abstract

The invention relates to the technical field of intelligent robots, and discloses a face recognition method of an intelligent robot, which comprises the following steps: step 1, lightweight face detection and feature extraction: performing face detection and feature extraction through a lightweight neural network structure, and combining model compression and edge calculation to reduce calculation complexity; and step 2, illumination and attitude adaptability is realized through multi-modal fusion, and face processing under different illumination and attitude conditions is carried out based on the extracted face features by utilizing fusion of an RGB image and an infrared image, adaptive illumination enhancement and 3D attitude correction. Through a multi-modal fusion technology, 3D attitude correction and self-adaptive illumination enhancement, different illumination conditions and attitude changes can be adapted, then the recognition stability of the system in a complex environment is remarkably improved, it is ensured that face recognition has high accuracy and robustness under the conditions of insufficient illumination, strong light or multi-angle offset, and the face recognition accuracy and robustness are improved. And the application requirements of the intelligent robot in multiple scenes are met.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent robots, and in particular to a face recognition method for an intelligent robot. Background Art

[0002] Intelligent robots are widely used in security monitoring, service robots, medical care, financial identity authentication and other fields, and the demand for identity authentication and human-computer interaction based on face recognition technology is growing. As an important part of biometric identification, face recognition technology has become an important means for intelligent robots to achieve efficient identity recognition and personalized services due to its convenience, non-contact and natural interaction advantages. However, existing face recognition methods face many challenges in changing usage environments, affecting the recognition stability and reliability of intelligent robots in complex environments.

[0003] Existing face recognition systems usually rely on RGB images, which are easily disturbed under complex lighting conditions, resulting in reduced recognition accuracy. In addition, intelligent robots will encounter target faces with different postures during face recognition, which puts higher requirements on the system's recognition under multi-angle conditions. However, existing face recognition technology that lacks lighting adaptability and posture robustness cannot meet the application needs of intelligent robots in complex scenes.

[0004] Traditional face recognition systems are mainly based on the extraction of static image features and are vulnerable to disguised attacks such as photos or videos. This poses a security risk, especially in scenarios with high security requirements such as security and finance. In addition, existing technologies lack an effective liveness detection mechanism and cannot distinguish between real faces and forged faces, which in turn affects the reliability of the system.

[0005] In practical applications, especially in intelligent robot applications for specific groups of people, there is a lack of sample data, which limits the effective training of face recognition models. Traditional face recognition methods are prone to overfitting when there are insufficient samples, resulting in reduced recognition performance.

[0006] Facial data, as sensitive private information of users, is vulnerable to the risk of leakage or illegal use during the collection, storage and transmission process. The existing centralized training and data transmission methods make it impossible to guarantee data privacy security, especially in scenarios where the network environment is unsafe or there is a need to process sensitive data, and user privacy protection needs cannot be met.

[0007] Therefore, those skilled in the art provide a face recognition method for an intelligent robot to solve the above-mentioned problems. Summary of the invention

[0008] In view of the deficiencies in the prior art, the present invention provides a face recognition method for an intelligent robot to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: A face recognition method for an intelligent robot, comprising:

[0010] Step 1: Lightweight face detection and feature extraction: face detection and feature extraction are performed through a lightweight neural network structure, combined with model compression and edge computing to reduce computational complexity;

[0011] Step 2: Multimodal fusion realizes illumination and posture adaptability. Based on the extracted facial features, the fusion of RGB images and infrared images, adaptive illumination enhancement and 3D posture correction are used to process faces under different illumination and posture conditions.

[0012] Step 3: Liveness detection and anti-spoofing attack: Based on the fusion features in step 2, the authenticity of the target face is determined by detecting facial micro-expressions, depth information and dynamic features to avoid photo or video disguise;

[0013] Step 4: Few-sample learning and transfer learning. Based on the features in steps 1 and 2, the pre-trained model is adjusted through transfer learning, and the few-sample learning model is combined to realize face recognition under small sample conditions.

[0014] Step 5: Data privacy protection and security. Based on the data processing flow from step 1 to step 4, federated learning and differential privacy technologies are used to protect user data, and encryption algorithms are used during data transmission and storage to ensure security.

[0015] Step 6: Adaptive system integration and scene recognition. According to the detection and recognition information from steps 1 to 5, the system parameters are adjusted to adapt to different environmental conditions, and the environmental adaptability is improved through multi-target tracking and multi-sensor fusion.

[0016] Preferably, the lightweight face detection and feature extraction includes:

[0017] The MobileNetV3 neural network structure is used to reduce the computational complexity through deep separable convolution. The convolution calculation formula is:

[0018] O conv =D k ×D k ×M×N×D f ×D f ,

[0019] Among them, D k is the convolution kernel size, M is the number of input channels, N is the number of output channels, and D f is the width and height of the feature map;

[0020] The model is compressed through pruning and quantization techniques. The weight score calculation formula in the pruning process is:

[0021] S i =∑ j |W i，j |,

[0022] Among them, S i Represents the weight W i，j The importance score, W i，j is the weight between the i-th node and the j-th node in the network, |W i，j | is the absolute value of the weight.

[0023] Preferably, the multimodal fusion to achieve illumination and posture adaptability includes:

[0024] The feature fusion of RGB and infrared images, the fusion formula is:

[0025] F fused =α·F RGB +(1-α)·F IR ,

[0026] Among them, F fused is the fused image feature, F RGB is the RGB image feature, F IR is the infrared image feature, α is the adaptive weight coefficient;

[0027] The face rotation matrix R and displacement vector T are calculated through the posture estimation algorithm, and the image is corrected for posture. The correction formula is:

[0028] p′=R·p+T,

[0029] Among them, p is the coordinate of the original face key point, p′ is the coordinate of the corrected face key point, R is the rotation matrix, and T is the translation vector.

[0030] Preferably, the illumination compensation adopts adaptive histogram equalization, and the grayscale conversion formula is:

[0031]

[0032] Among them, f(x, y) is the gray value of the pixel, min(f) and max(f) are the minimum and maximum gray values ​​of the image, and L is the gray level.

[0033] Preferably, the liveness detection and anti-spoofing attack include:

[0034] Micro-expressions and head movement features are detected by facial key points, and the key point movement amplitude calculation formula is:

[0035]

[0036] Among them, d t is the motion amplitude of the key point at time t, (x t ,y t ) is the key point coordinate, x t and t represents the x and y coordinates of the point at time t, x t+1 and t+1 represents the x and y coordinates of the point at time t+1;

[0037] Depth images and thermal imaging data are combined to detect the 3D structure and temperature changes of the face for liveness detection.

[0038] Preferably, the dynamic feature time series analysis uses a long short-term memory network to analyze the time series features in continuous frame images, and its hidden state update formula is:

[0039] h t =f(W h ·x t +U h ·h t-1 +b h ),

[0040] Among them, h t is the hidden state at time t, W h and U h is the weight matrix, x t is the input, b h is the bias, h t-1 is the hidden state vector at the t-1th moment, and f is the activation function.

[0041] Preferably, the few-shot learning optimizes the model through a triplet loss function, and the loss function is:

[0042] L = max(||f(x a )-f(x p )|| 2 -||f(x a )-f(x n )|| 2 +α, 0),

[0043] Among them, x a is the anchor point sample, x p is a positive sample, x n is a negative sample, α is a margin hyperparameter, L is the loss value of the triplet loss function, f(x a ) is the anchor sample x a The embedding vector of f(x p ) is a positive sample x pThe embedding vector of f(x n ) is a negative sample x n The embedding vector of .

[0044] Preferably, the data privacy protection includes:

[0045] Based on the federated learning framework, the model is updated locally, and the update formula of the model parameter θ is:

[0046]

[0047] Among them, θ (t+1) is the updated parameter, η is the learning rate, is the gradient of the loss function, θ (t) is the model parameter at the tth iteration;

[0048] Differential privacy noise is added during the gradient update process to protect data privacy.

[0049] Preferably, the encrypted transmission and storage adopts AES encryption algorithm, and the encryption formula is:

[0050]

[0051] Among them, Enc(X) is the encrypted data, X is the original data, is an XOR operation, and K is the key.

[0052] Preferably, the adaptive system integration and scene recognition includes:

[0053] Kalman filtering is used to track multiple targets on target faces, and the state update formula is:

[0054] X t+1 =FX t +BU t +W t ,

[0055] Among them, X t+1 is the state at the next moment, X t is the current state, F is the state transfer matrix, B is the control matrix, U t is the control vector, W t is the process noise;

[0056] The multimodal data of the RGB camera, depth camera and infrared sensor D are fused to obtain fused features.

[0057] The present invention provides a face recognition method for an intelligent robot. It has the following beneficial effects:

[0058] 1. The present invention can adapt to different lighting conditions and posture changes through multimodal fusion technology, 3D posture correction and adaptive lighting enhancement, thereby significantly improving the recognition stability of the system in complex environments, ensuring that face recognition has high accuracy and robustness under conditions of insufficient lighting, strong light or multi-angle offset, so as to meet the application needs of intelligent robots in multiple scenarios.

[0059] 2. The present invention combines facial micro-expression analysis, depth information detection and dynamic feature analysis to achieve liveness detection function, effectively identifies real faces and forged faces, prevents photo or video disguise attacks, greatly improves the anti-counterfeiting and security of the face recognition system, and makes the system have significant advantages in security, finance and other scenarios with high security requirements.

[0060] 3. The present invention can achieve efficient recognition in a small sample data environment through the combination of transfer learning technology and few-sample learning model. Through the application of triplet loss and prototype network, the system can maintain high accuracy and robustness under few-sample conditions, making the system more versatile and adaptable, and particularly suitable for application scenarios with difficult data collection and small number of samples.

[0061] 4. The present invention adopts federated learning and differential privacy technology to enable the system to update the model locally to avoid transmitting original data, and add random noise in the model update to protect user privacy data, meet the requirements of data privacy protection, adapt to the application needs of intelligent robots in network security sensitive scenarios, and provide reliable security protection for user data. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0063] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0064] The present invention is described in detail below in conjunction with the accompanying drawings:

[0065] Example:

[0066] Please refer to the attached Figure 1 The embodiment of the present invention provides a face recognition method for an intelligent robot, comprising:

[0067] Step 1: Lightweight face detection and feature extraction: face detection and feature extraction are performed through a lightweight neural network structure, combined with model compression and edge computing to reduce computational complexity;

[0068] Step 2: Multimodal fusion realizes illumination and posture adaptability. Based on the extracted facial features, the fusion of RGB images and infrared images, adaptive illumination enhancement and 3D posture correction are used to process faces under different illumination and posture conditions.

[0069] Step 3: Liveness detection and anti-spoofing attack: Based on the fusion features in step 2, the authenticity of the target face is determined by detecting facial micro-expressions, depth information and dynamic features to avoid photo or video disguise;

[0070] Step 4: Few-sample learning and transfer learning. Based on the features in steps 1 and 2, the pre-trained model is adjusted through transfer learning, and the few-sample learning model is combined to realize face recognition under small sample conditions.

[0071] Step 5: Data privacy protection and security. Based on the data processing flow from step 1 to step 4, federated learning and differential privacy technologies are used to protect user data, and encryption algorithms are used during data transmission and storage to ensure security.

[0072] Step 6: Adaptive system integration and scene recognition. According to the detection and recognition information from steps 1 to 5, the system parameters are adjusted to adapt to different environmental conditions, and the environmental adaptability is improved through multi-target tracking and multi-sensor fusion.

[0073] Benefits of step 1: By adopting a lightweight neural network structure for face detection and feature extraction, combined with model compression and edge computing, the computational complexity of the system is reduced, the real-time performance of face detection is improved, and the intelligent robot can efficiently perform face detection tasks in a resource-constrained environment;

[0074] Benefits of step 2: Through the fusion of RGB images and infrared images, adaptive illumination enhancement and 3D posture correction, robust face recognition under various lighting and posture conditions is achieved, so that the system can adapt to complex scenes of strong light, dark light and different angles, and improve the stability of face recognition;

[0075] Benefits of step 3: By detecting facial micro-expressions, depth information and dynamic features, real faces and fake faces can be effectively distinguished to prevent photo or video disguise attacks. The liveness detection process significantly improves the security of the system, making it advantageous in security and identity authentication applications with high security requirements.

[0076] The benefits of step 4 are to adjust the pre-trained model in combination with transfer learning, and use the few-sample learning model to maintain high recognition accuracy under limited data conditions. This is suitable for situations where sample collection is difficult or training data is limited, ensuring that the face recognition system has good adaptability and generalization capabilities;

[0077] Benefits of step 5: Federated learning and differential privacy technologies are used to protect user data. Encryption algorithms are used to ensure the security of data transmission and storage. This allows the system to meet privacy protection requirements when processing user data and avoid the risk of user data leakage. This is especially suitable for use in security and privacy-sensitive application scenarios.

[0078] The benefit of step 6 is that the system parameters are adjusted in real time through multi-target tracking and multi-sensor fusion, so that the system can automatically adapt to different environmental conditions, further improve the environmental adaptability of face recognition, ensure the recognition stability of the system in dynamic or multi-target scenes, and meet the application requirements in complex environments.

[0079] Lightweight face detection and feature extraction include:

[0080] The MobileNetV3 neural network structure is used to reduce the computational complexity through deep separable convolution. The convolution calculation formula is:

[0081] O conv =D k ×D k ×M×N×D f ×D f ,

[0082] Among them, D k is the convolution kernel size, M is the number of input channels, N is the number of output channels, and D f is the width and height of the feature map;

[0083] The model is compressed through pruning and quantization techniques. The weight score calculation formula in the pruning process is:

[0084] S i =∑ j |W i，j |,

[0085] Among them, S i Represents the weight W i，j The importance score, W i，j is the weight between the i-th node and the j-th node in the network, |W i，j | is the absolute value of the weight.

[0086] By adopting the MobileNetV3 neural network structure and using deep separable convolution to reduce the computational complexity of convolution operations, compared with traditional convolution, deep separable convolution decomposes the calculation into convolution operations, reduces the number of parameters and the amount of calculation, and makes face detection and feature extraction more efficient, meeting the requirements of intelligent robots for real-time and efficient calculation;

[0087] The model is compressed through pruning and quantization technology. Pruning technology removes redundant weight connections in the network, and quantization reduces the floating-point precision of weights to 8 or 16 bits, which further reduces the model size, reduces the model storage requirements, and reduces computing power consumption, making the face recognition model suitable for running on resource-constrained embedded devices and intelligent robots;

[0088] By compressing the model, memory usage and processor load are reduced, and the battery life of the device is extended. This is especially suitable for portable or battery-powered intelligent robots. The energy-saving advantage is particularly evident on devices with limited resources, and helps maintain stability during long-term operation.

[0089] In general, the use of the MobileNetV3 structure and the pruning and quantization compression model make face detection and feature extraction efficient and suitable for implementation on resource-constrained embedded devices such as intelligent robots, providing support for real-time face recognition applications.

[0090] Multimodal fusion to achieve illumination and posture adaptability includes:

[0091] The feature fusion of RGB and infrared images, the fusion formula is:

[0092] F fused =α·F RGB +(1-α)·F IR ,

[0093] Among them, F fused is the fused image feature, F RGB is the RGB image feature, F IR is the infrared image feature, α is the adaptive weight coefficient;

[0094] The face rotation matrix R and displacement vector T are calculated through the posture estimation algorithm, and the image is corrected for posture. The correction formula is:

[0095] p′=R·p+T,

[0096] Among them, p is the coordinate of the original face key point, p′ is the coordinate of the corrected face key point, R is the rotation matrix, and T is the translation vector.

[0097] By fusing the features of RGB and infrared images, the system can utilize visible light information and infrared information. RGB images provide clear details under normal lighting conditions, and infrared images maintain clarity under low light or complex lighting conditions. This allows the system to maintain consistent recognition results in complex lighting environments such as strong light, dark light, and shadows, thereby improving the robustness of face recognition.

[0098] The rotation matrix R and displacement vector T of the face are calculated through posture estimation, and posture correction is performed on face images at different angles. The deflected face is corrected to a frontal view, so that the face recognition system maintains high recognition accuracy under multi-angle conditions and reduces recognition errors caused by posture changes.

[0099] The combination of multimodal fusion and posture correction technology enables the system to work in a variety of application scenarios, including indoor low-light, night monitoring, outdoor strong light and complex angle scenes, broadening the scope of face recognition application scenarios;

[0100] The fusion of RGB and infrared images and posture correction reduce recognition errors caused by scattered light or posture deviation, and improve the accuracy and stability of face recognition. It is particularly suitable for scenarios such as security and outdoor monitoring that have high requirements for recognition accuracy and environmental adaptability.

[0101] Light compensation uses adaptive histogram equalization, and the grayscale conversion formula is:

[0102]

[0103] Among them, f(x, y) is the gray value of the pixel, min(f) and max(f) are the minimum and maximum gray values ​​of the image, and L is the gray level.

[0104] Through adaptive histogram equalization, the grayscale of the image can be adjusted to improve the contrast of the image under low light or scattered light conditions. The adjustment process makes the facial features in the image clearer, which is helpful for subsequent face detection and feature extraction, and improves the recognition accuracy.

[0105] Adaptive histogram equalization can locally adjust the grayscale of the image and enhance the dark areas in the image to avoid affecting the details of the bright areas. The local processing method effectively reduces the image distortion caused by scattered light and ensures that reliable facial features can be extracted under complex lighting conditions.

[0106] At night or in low-light environments, reduced image brightness and contrast will result in poor face recognition results. Adaptive histogram equalization can significantly improve the brightness of low-light images, making faces clearly visible in low-light conditions and improving the system's recognition capabilities in low-light environments.

[0107] Adaptive histogram equalization can retain more image details while improving contrast, avoiding over-enhancement and detail loss caused by traditional global histogram equalization, and improving the accuracy of feature extraction;

[0108] Adaptive histogram equalization automatically adjusts the contrast according to the local illumination conditions of the image without the need to manually set parameters, enabling the system to obtain optimized image quality in different scenarios, reducing the need for human intervention and improving the convenience and adaptability of the system.

[0109] Liveness detection and anti-spoofing attack include:

[0110] Micro-expressions and head movement features are detected by facial key points, and the key point movement amplitude calculation formula is:

[0111]

[0112] Among them, d t is the motion amplitude of the key point at time t, (x t ,y t ) is the key point coordinate, x t and t represents the x and y coordinates of the point at time t, x t+1 and t+1 represents the x and y coordinates of the point at time t+1;

[0113] Depth images and thermal imaging data are combined to detect the 3D structure and temperature changes of the face for liveness detection.

[0114] By detecting micro-expressions and micro-movements of the head through facial key points, the system can capture subtle dynamic changes in the face, such as blinking and lip movements. However, fake photos or videos lack subtle dynamic features. Therefore, the analysis of the key point movement amplitude can effectively distinguish between real faces and static fake faces, improving the security of recognition.

[0115] Based on the motion detection and depth information of facial key points, the system can determine whether it is a real face and prevent attacks using disguised means such as photos, videos played on the screen, and even 3D models. It is particularly suitable for scenarios such as financial payment and access control security that require extremely high identity authentication security, and improves the system's anti-counterfeiting capabilities;

[0116] Combining depth images and thermal imaging data, the system can detect the three-dimensional structure and temperature distribution of the face, and further determine whether it is a real face. Fake faces usually lack normal depth and temperature features. Multimodal detection increases the reliability of recognition and improves the security and credibility of the face recognition system.

[0117] The liveness detection function is particularly suitable for application scenarios such as security monitoring, access control, and identity authentication. The scenarios need to ensure that the object being identified is a real face. The system combines multimodal detection of micro-expressions, depth information, and thermal imaging to broaden the application scenarios of face recognition technology and make it highly adaptable in various occasions with high security requirements.

[0118] Dynamic feature time series analysis uses a long short-term memory network to analyze the time series features in continuous frame images. Its hidden state update formula is:

[0119] h t =f(W h ·x t +U h ·h t-1 +b h ),

[0120] Among them, h t is the hidden state at time t, W h and U h is the weight matrix, x t is the input, b h is the bias, h t-1 is the hidden state vector at the t-1th moment, and f is the activation function.

[0121] By analyzing the time series features in continuous frame images through the long short-term memory network, the system can accurately capture subtle dynamic changes in the face, such as blinking, lip movements, and expression changes. Compared with the static features of a single frame image, the long short-term memory network can better model the temporal feature changes and improve the recognition system's ability to perceive the dynamics of real faces.

[0122] By using the time series analysis capability of the LSTM network, the system can detect continuous dynamic changes in human faces and identify forged videos or static photos disguised as real faces. Forged faces lack natural time series features, and the LSTM network significantly improves the accuracy of liveness detection by learning features, effectively preventing camouflage attacks.

[0123] Long short-term memory networks can process video frame sequences in real time in complex environments and dynamically track changes in faces. The hidden state update formula enables the network to maintain important timing information for a long time while ignoring unimportant short-term fluctuations, thereby achieving robust dynamic feature detection and adapting to complex application scenarios.

[0124] The long short-term memory network structure has memory and forgetting functions, which can effectively filter out noise and short-term interference in the image, allowing the system to focus on the real dynamic features of the face. Its anti-interference ability improves the system's performance in noisy environments or when the camera is unstable, ensuring the stability and reliability of the recognition process.

[0125] Few-shot learning optimizes the model through a triplet loss function, and its loss function is:

[0126] L = max(||f(x a )-f(x p )|| 2 -||f(x a )-f(x n )|| 2 +α, 0),

[0127] Among them, x a is the anchor point sample, x p is a positive sample, x n is a negative sample, α is a margin hyperparameter, L is the loss value of the triplet loss function, f(x a ) is the anchor sample x a The embedding vector of f(x p ) is a positive sample x p The embedding vector of f(x n ) is a negative sample x n The embedding vector of .

[0128] By optimizing the triplet loss function, the system can learn the effective feature distances between different faces and enhance the ability to distinguish faces under small sample conditions. The triplet loss makes the distance between the anchor point sample and the positive sample closer and the distance from the negative sample farther, effectively solving the problem of the traditional method's reduced recognition performance when the sample size decreases.

[0129] The triplet loss learns the relative distance between samples instead of relying on the number of samples of a specific category, so it can effectively generalize under small sample data. In the case of insufficient data, the model can generate high-quality feature embedding vectors, thereby improving the adaptability and accuracy of the recognition system in different environments and scenarios.

[0130] The optimization of the triplet loss function requires the pairing information of anchor points, positive samples, and negative samples, and does not require labeled data. Therefore, it reduces the system's demand for massive samples and reduces the cost of data preparation. It is especially suitable for application scenarios where data collection is difficult and samples are limited.

[0131] The triplet loss function provides a clear optimization goal, allowing the model to quickly converge to an efficient feature representation. By aggregating positive samples and anchor samples and separating negative samples, the gradient direction of the loss function is clear, which significantly improves training efficiency, shortens the convergence time of the model, and speeds up the system deployment cycle.

[0132] Data privacy protection includes:

[0133] Based on the federated learning framework, the model is updated locally, and the update formula of the model parameter θ is:

[0134]

[0135] Among them, θ (t+1) is the updated parameter, η is the learning rate, is the gradient of the loss function, θ (t) is the model parameter at the tth iteration;

[0136] Differential privacy noise is added during the gradient update process to protect data privacy.

[0137] The federated learning framework is used to update the model locally. The system does not need to upload the user's original data to a centralized server, but transmits the model's parameter update information. This method ensures that the user's data is always stored locally, effectively protecting the user's privacy. It is particularly suitable for application scenarios with high privacy sensitivity, such as medical, financial, and security fields.

[0138] Differential privacy noise is introduced in the gradient update process. By adding noise to the gradient, it is ensured that the specific data of a single user cannot be restored when the server aggregates multiple local model updates. The introduction of differential privacy provides a strong data security guarantee for the system, preventing hackers or internal attackers from restoring sensitive information by analyzing parameter updates.

[0139] The privacy budget in differential privacy technology can be flexibly adjusted, allowing the system to determine the balance between data privacy protection and model performance. According to the different needs of application scenarios, the noise is adjusted to obtain the privacy protection level to ensure the recognition accuracy of the model.

[0140] The federated learning framework can perform model updates on multiple distributed devices, allowing the system to participate in training on different devices and locations, thereby protecting data privacy. It enables the system to integrate diverse device data, achieve cross-device recognition optimization, and further enhance the model's generalization capabilities.

[0141] Encrypted transmission and storage use the AES encryption algorithm, the encryption formula is:

[0142] Enc(X)=X⊕K,

[0143] Among them, Enc(X) is the encrypted data, X is the original data, ⊕ is the XOR operation, and K is the key.

[0144] The AES encryption algorithm is used to encrypt data in transmission and storage, ensuring that the user's facial feature data cannot be directly accessed during transmission and storage. If it is intercepted during network transmission or the storage device is attacked, the encrypted data cannot be cracked, effectively preventing data leakage and ensuring the security of user information.

[0145] AES encryption is a symmetric encryption algorithm with high security. It is especially suitable for encrypting sensitive information such as facial features. The AES algorithm has high anti-cracking capabilities. It is very difficult for attackers with strong computing power to decrypt AES encrypted data. Therefore, it can ensure that data is effectively protected during storage to avoid malicious access or tampering.

[0146] During data transmission, the AES encryption algorithm can encrypt data at the sending end and decrypt it at the receiving end, ensuring that the data always remains encrypted during transmission. In this way, the system can exchange data securely in an open network environment, effectively avoiding network eavesdropping and man-in-the-middle attacks, and improving the security of data transmission.

[0147] Adaptive system integration and scene recognition include:

[0148] Kalman filtering is used to track multiple targets on target faces, and its state update formula is:

[0149] X t+1 =FX t +BU t +W t ,

[0150] Among them, X t+1 is the state at the next moment, X t is the current state, F is the state transfer matrix, B is the control matrix, U t is the control vector, W t is the process noise;

[0151] The multimodal data of the RGB camera, depth camera and infrared sensor D are fused to obtain fused features.

[0152] The Kalman filter is used to track multiple target faces, so that the system can continuously track faces in a dynamic environment. The state update formula of the Kalman filter infers the state of the next moment based on the previous state of the target, and corrects the prediction result in combination with noise factors to ensure that the target face can be accurately tracked in motion and improve the stability of recognition.

[0153] Through multimodal data fusion, the system integrates the data of RGB camera, depth camera and infrared sensor, so that the recognition system can work in a variety of environmental conditions. RGB images provide conventional lighting information, depth camera provides distance and shape information, infrared camera provides low light or night vision information. Multimodal fusion enables the system to adapt to different lighting and complex scene conditions.

[0154] The system performs automatic parameter adjustment through adaptive multimodal data fusion and Kalman filtering. It can automatically adjust tracking and identification parameters according to real-time environmental conditions, reduce the need for human intervention, improve the automation and adaptability of the system in complex and changing environments, and enable it to respond to environmental changes more flexibly.

[0155] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A face recognition method for an intelligent robot, characterized in that: include: Step 1: Lightweight face detection and feature extraction: face detection and feature extraction are performed through a lightweight neural network structure, combined with model compression and edge computing to reduce computational complexity; Step 2: Multimodal fusion realizes illumination and posture adaptability. Based on the extracted facial features, the fusion of RGB images and infrared images, adaptive illumination enhancement and 3D posture correction are used to process faces under different illumination and posture conditions. Step 3: Liveness detection and anti-spoofing attack: Based on the fusion features in step 2, the authenticity of the target face is determined by detecting facial micro-expressions, depth information and dynamic features to avoid photo or video disguise; Step 4: Few-sample learning and transfer learning. Based on the features in steps 1 and 2, the pre-trained model is adjusted through transfer learning, and the few-sample learning model is combined to realize face recognition under small sample conditions. Step 5: Data privacy protection and security. Based on the data processing flow from step 1 to step 4, federated learning and differential privacy technologies are used to protect user data, and encryption algorithms are used during data transmission and storage to ensure security. Step 6: Adaptive system integration and scene recognition. According to the detection and recognition information from steps 1 to 5, the system parameters are adjusted to adapt to different environmental conditions, and the environmental adaptability is improved through multi-target tracking and multi-sensor fusion.

2. The face recognition method of an intelligent robot according to claim 1, characterized in that: The lightweight face detection and feature extraction includes: The MobileNetV3 neural network structure is used to reduce the computational complexity through deep separable convolution. The convolution calculation formula is: O conv =D k ×D k ×M×N×D f ×D f , Among them, D k is the convolution kernel size, M is the number of input channels, N is the number of output channels, and D f is the width and height of the feature map; The model is compressed through pruning and quantization techniques. The weight score calculation formula in the pruning process is: S i =∑ j |W i，j |, Among them, S i Represents the weight W i，j The importance score, W i，j is the weight between the i-th node and the j-th node in the network, |W i，j | is the absolute value of the weight.

3. The face recognition method of an intelligent robot according to claim 1, characterized in that: The multimodal fusion to achieve illumination and posture adaptability includes: The feature fusion of RGB and infrared images, the fusion formula is: F fused =α·F RGB +(1-α)·F IR , Among them, F fused is the fused image feature, F RGB is the RGB image feature, F IR is the infrared image feature, α is the adaptive weight coefficient; The face rotation matrix R and displacement vector T are calculated through the posture estimation algorithm, and the image is corrected for posture. The correction formula is: p′=R·p+T, Among them, p is the coordinate of the original face key point, p′ is the coordinate of the corrected face key point, R is the rotation matrix, and T is the translation vector.

4. The face recognition method of an intelligent robot according to claim 1, characterized in that: The illumination compensation adopts adaptive histogram equalization, and its grayscale conversion formula is: Among them, f(x, y) is the gray value of the pixel, min(f) and max(f) are the minimum and maximum gray values ​​of the image, and L is the gray level.

5. The face recognition method of an intelligent robot according to claim 1, characterized in that: The liveness detection and anti-spoofing attack include: Micro-expressions and head movement features are detected by facial key points, and the key point movement amplitude calculation formula is: Among them, d t is the motion amplitude of the key point at time t, (x t ,y t ) is the key point coordinate, x t and t represents the x and y coordinates of the point at time t, x t+1 and t+1 represents the x and y coordinates of the point at time t+1; Depth images and thermal imaging data are combined to detect the 3D structure and temperature changes of the face for liveness detection.

6. The face recognition method of an intelligent robot according to claim 1, characterized in that: The dynamic feature time series analysis adopts a long short-term memory network to analyze the time series features in continuous frame images, and its hidden state update formula is: h t =f(W h ·x t +U h ·h t-1 +b h ), Among them, h t is the hidden state at time t, W h and U h is the weight matrix, x t is the input, b h is the bias, h t-1 is the hidden state vector at the t-1th moment, and f is the activation function.

7. The face recognition method of an intelligent robot according to claim 1, characterized in that: The few-shot learning optimizes the model through a triplet loss function, and its loss function is: L=max(||f(x a )-f(x p )|| 2 -||f(x a )-f(x n )|| 2 +α,0), Among them, x a is the anchor point sample, x p is a positive sample, x n is a negative sample, α is a margin hyperparameter, L is the loss value of the triplet loss function, f(x a ) is the anchor sample x a The embedding vector of f(x p ) is a positive sample x p The embedding vector of f(x n ) is a negative sample x n The embedding vector of .

8. The face recognition method of an intelligent robot according to claim 1, characterized in that: The data privacy protection includes: Based on the federated learning framework, the model is updated locally, and the update formula of the model parameter θ is: Among them, θ (t+1) is the updated parameter, η is the learning rate, is the gradient of the loss function, θ (t) is the model parameter at the tth iteration; Differential privacy noise is added during the gradient update process to protect data privacy.

9. The face recognition method of an intelligent robot according to claim 1, characterized in that: The encrypted transmission and storage adopts the AES encryption algorithm, and the encryption formula is: Among them, Enc(X) is the encrypted data, X is the original data, is an XOR operation, and K is the key.

10. The face recognition method of an intelligent robot according to claim 1, characterized in that: The adaptive system integration and scene recognition include: Kalman filtering is used to track multiple targets on target faces, and its state update formula is: X t+1 =FX t +THIS t +W t , Among them, X t+1 is the state at the next moment, X t is the current state, F is the state transfer matrix, B is the control matrix, U t is the control vector, W t is the process noise; The multimodal data of the RGB camera, depth camera and infrared sensor D are fused to obtain fused features.

Images