Visual hidden near-infrared face identity recognition method based on edge computing chip

Abstract

The application discloses a visual hidden near-infrared face identity recognition method based on an edge computing chip, and an electronic device, and the method is applied to a face identity recognition device, and the face identity recognition device comprises a near-infrared camera; the face identity recognition device comprises a first feature detection model and a second feature detection model, and the method comprises the following steps: acquiring a to-be-recognized image by using the near-infrared camera, performing feature detection on the to-be-recognized image by using the first feature detection model, and obtaining feature information corresponding to a detection target; performing feature detection on a visual hidden image by using the second feature detection model, and obtaining a registration feature; performing identity recognition based on the feature information and the registration feature; wherein the visual hidden image is obtained by performing visual disturbance on an original registration image and is stored in the face identity recognition device, and the original registration image is not stored in the face identity recognition device. The method disclosed by the application can reduce the information leakage risk in the face recognition process.

Description

Technical Field

[0001] This invention belongs to the field of facial recognition technology, specifically relating to a visually hidden near-infrared facial identity recognition method and electronic device based on an edge computing chip. Background Technology

[0002] Facial recognition has been widely used in smart access control, financial payment, security monitoring and other scenarios, but existing solutions generally suffer from privacy leaks.

[0003] Currently, mainstream facial recognition system architectures are mainly divided into two categories: "cloud-based comparison" and "edge-based comparison." Cloud-based comparison solutions require uploading raw facial images captured by cameras or extracted plaintext features to a remote server via the network for verification. This architecture makes sensitive biometric data highly vulnerable to man-in-the-middle attacks or data theft during transmission, and the system will completely fail in weak network or offline environments. While traditional edge-based comparison solutions achieve offline operation, they have serious vulnerabilities in their data processing mechanisms: devices typically need to store raw facial images or plaintext feature templates locally. Once the device is physically hacked or its firmware is reverse-engineered, massive biometric databases will face the risk of leakage. Furthermore, to achieve visually hidden data storage, existing technologies often require decrypting the visually hidden feature templates before comparison, or uploading the collected features to a Trusted Execution Environment (TEE) for decryption. This not only increases the system's computational overhead and processing latency but also causes the decrypted plaintext data to be briefly exposed in memory, posing a risk of side-channel attacks. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a visually concealed near-infrared face recognition method and electronic device based on an edge computing chip.

[0005] The technical problem to be solved by this invention is achieved through the following technical solution: On one hand, a visually hidden near-infrared face recognition method based on an edge computing chip is provided. This method is applied to a face recognition device, which includes a near-infrared camera and a first feature detection model and a second feature detection model. The method includes: The near-infrared camera is used to acquire an image to be identified, and the first feature detection model is used to perform feature detection on the image to be identified to obtain feature information corresponding to the target. The second feature detection model is used to perform feature detection on the visually hidden image to obtain registration features; Identity recognition is performed based on the feature information and the registration features; wherein, the visually hidden image is obtained by visually perturbing the original registration image and stored in the face recognition device, and the original registration image is not stored in the face recognition device.

[0006] In one embodiment, the method further includes: The original registration image was acquired using the near-infrared camera. The original registration image is visually perturbed to obtain the visually hidden image, and the visually hidden image is stored in the face recognition device; Delete the original registration image.

[0007] In one embodiment, deleting the original registered image includes: The underlying memory overwrite interface is invoked to overwrite the data in the physical memory area where the original registration image is stored, thereby deleting the original registration image; The data overlay includes writing all-zero data or writing noisy data.

[0008] In one embodiment, the method further includes: After acquiring the image to be identified using the near-infrared camera, the image is stored in volatile memory; or, After obtaining the feature information corresponding to the target to be detected, the image to be identified is deleted.

[0009] In one embodiment, the facial recognition device further includes a facial detection model; The first feature detection model is used to perform feature detection on the image to be identified to obtain feature information corresponding to the target, including: The image to be identified is preprocessed to obtain the image to be processed; The face detection model is used to perform face detection on the image to be processed, and the predicted detection box and confidence score corresponding to the detection target are determined. The actual detection box corresponding to the detection target is determined based on the predicted detection box and the confidence score. The first feature detection model is used to perform feature detection based on the actual detection box corresponding to the detection target, thereby obtaining the feature information corresponding to the detection target.

[0010] In one embodiment, identity recognition based on the feature information and registration features includes: Calculate the similarity between the feature information and each of the registered features; The identity information of the detected target is determined based on the similarity.

[0011] In one embodiment, acquiring an image to be identified using the near-infrared camera includes: The near-infrared camera is used to acquire n consecutive images of the image to be identified within a preset time window; Determining the identity information of the detected target based on the similarity includes: If the similarity of m consecutive images to be identified within a preset time window is less than the similarity threshold, the identity recognition fails and an alarm is triggered.

[0012] On the other hand, an electronic device is provided, comprising a memory and a processor coupled to each other, the processor being used to execute program instructions stored in the memory, and the processor being used to execute program data to implement the steps in the visually hidden near-infrared face recognition method based on an edge computing chip as described above.

[0013] In another aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the visually hidden near-infrared face recognition method based on an edge computing chip as described above.

[0014] This invention provides a visually hidden near-infrared face recognition method based on an edge computing chip. The method is applied to a face recognition device, which includes a near-infrared camera and a first feature detection model and a second feature detection model. The method includes acquiring an image to be recognized using the near-infrared camera; performing feature detection on the image using the first feature detection model to obtain feature information corresponding to the target; performing feature detection on a visually hidden image using the second feature detection model to obtain registration features; and performing identity recognition based on the feature information and the registration features. The visually hidden image is obtained by visually perturbing the original registration image and stored in the face recognition device, while the original registration image is not stored in the face recognition device. This method can reduce the risk of information leakage during face recognition.

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the visually hidden near-infrared face recognition method based on an edge computing chip according to this application. Figure 2 yes Figure 1 A flowchart illustrating an embodiment of step S11; Figure 3This is a schematic diagram of the image to be identified and the actual detection bounding box in this application; Figure 4 This is a schematic diagram of the process for obtaining registration features in the visual hidden near-infrared face recognition method based on edge computing chip in this application; Figure 5 This is a schematic diagram of the original registered image in the method of this application; Figure 6 This is a schematic diagram of the visually hidden image in the method of this application; Figure 7 This is a flowchart illustrating another embodiment of the visually hidden near-infrared face recognition method based on edge computing chips in this application; Figure 8 This is a schematic diagram illustrating the registration feature extraction and plaintext removal in the method of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0017] Facial recognition has been widely used in smart access control, financial payment, security monitoring and other scenarios, but existing solutions generally suffer from privacy leaks.

[0018] Currently, mainstream facial recognition system architectures are mainly divided into two categories: "cloud-based comparison" and "edge-based comparison." Cloud-based comparison solutions require uploading raw facial images captured by cameras or extracted plaintext features to a remote server via the network for verification. This architecture makes sensitive biometric data highly vulnerable to man-in-the-middle attacks or data theft during transmission, and the system will completely fail in weak network or offline environments. While traditional edge-based comparison solutions achieve offline operation, they have serious vulnerabilities in their data processing mechanisms: devices typically need to store raw facial images or plaintext feature templates locally. Once the device is physically hacked or its firmware is reverse-engineered, massive biometric databases will face the risk of leakage. Furthermore, to achieve visually hidden data storage, existing technologies often require decrypting the visually hidden feature templates before comparison, or uploading the collected features to a Trusted Execution Environment (TEE) for decryption. This not only increases the system's computational overhead and processing latency but also causes the decrypted plaintext data to be briefly exposed in memory, posing a risk of side-channel attacks.

[0019] In view of this, this application provides a visually hidden near-infrared face recognition method based on an edge computing chip. It should be noted that the method of this application can utilize the heterogeneous computing characteristics of the CPU and NPU in the edge chip to achieve a one-way, non-penetrating video stream transient processing architecture and "non-decrypted feature comparison". The invention will be further described in detail below with reference to specific embodiments, but the implementation of the invention is not limited thereto.

[0020] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the visually concealed near-infrared face recognition method and electronic device based on edge computing chips according to this application. The face recognition method of this application is applied to face recognition devices, such as access control devices, alarm notification devices, human-computer interaction devices, etc. The face recognition device includes a near-infrared camera, and further includes a first feature detection model and a second feature detection model. It should be noted that the outputs of the first feature detection model and the second feature detection model are vector representations of the same dimension.

[0021] Step S11: Use the near-infrared camera to acquire the image to be identified, and use the first feature detection model to perform feature detection on the image to be identified to obtain the feature information corresponding to the target.

[0022] Specifically, the face recognition device of this application includes a near-infrared camera, which acquires the image to be recognized during face recognition. Because of the near-infrared camera, stable face image input can be obtained as the image to be recognized in nighttime, low light, backlight, or sudden changes in lighting conditions, in conjunction with near-infrared supplementary lighting.

[0023] After the system starts, it pulls a video stream from the vision processing library. The raw frames captured by the camera are only used as instantaneous input into the processing pipeline. The first feature detection model is used to perform feature detection on the image to be recognized, thereby obtaining the feature information corresponding to the detected target in the image.

[0024] In one embodiment of this application, combined with Figure 2 The facial recognition device also includes a facial detection model, and step S11 above further includes: Step S111: Preprocess the image to be identified to obtain the image to be processed.

[0025] Specifically, the images captured by the camera are generally in BGR format, but some image processing models require RGB images as input. Therefore, the image to be recognized is preprocessed by converting the original BGR image into an RGB image to obtain the image to be processed.

[0026] In one embodiment, after obtaining the image to be processed, the image to be identified is deleted. This prevents the image to be identified from being physically disclosed.

[0027] Step S112: Use the face detection model to perform face detection on the image to be processed, and determine the predicted detection box and confidence score corresponding to the detection target.

[0028] The image to be processed is fed into a face detection model deployed on an NPU. The NPU quickly outputs the predicted bounding box and confidence score corresponding to the detected target through forward propagation.

[0029] Step S113: Determine the actual detection box corresponding to the detection target based on the predicted detection box and the confidence score.

[0030] After obtaining the predicted detection box and confidence score corresponding to the target, the center coordinate form of the predicted detection box is converted into corner coordinate form. Non-maximum suppression (NMS) is then applied using a set confidence threshold to eliminate redundant detection boxes. Finally, the precise face coordinates (x1, y1, x2, y2) in the image are located as the actual detection box corresponding to the target. Please refer to... Figure 3 , Figure 3 The image on the left is the image to be identified, and the image on the right is the actual detection box (shown in red).

[0031] Step S114: Use the first feature detection model to perform feature detection based on the actual detection box corresponding to the detection target, and obtain the feature information corresponding to the detection target.

[0032] Specifically, to capture more complete facial information and adapt to complex environments, the actual detection box is expanded outward and offset when extracting the face region (ROI) based on the actual detection box, with out-of-bounds protection added. The cropped face is then scaled to 128×128 and fed into the first feature detection model. The first feature detection model is used to perform feature detection based on the actual detection box corresponding to the detection target to obtain the feature information corresponding to the detection target.

[0033] It should be noted that the first feature detection model transforms the pixel matrix of the actual detection box into a high-dimensional abstract feature vector, eliminating the need to use the original image frame.

[0034] In one embodiment, after acquiring the image to be identified captured by the camera, the image is stored in a volatile memory. After power failure or active release by the program, the internally stored data will be completely lost immediately to avoid leakage.

[0035] In another embodiment, the image to be identified can be deleted after obtaining the feature information corresponding to the target. This proactive deletion mechanism also helps prevent leakage.

[0036] Existing edge recognition systems typically employ a serial architecture of "first capture, then cache, then process," where raw image frames captured by the camera are temporarily stored in the device's memory or flash memory. This data processing mechanism results in the raw facial images existing in plaintext on the device, making them highly vulnerable to being captured by malware or exported through physical interfaces. The solution of this invention ensures that the data captured by the camera enters the processing pipeline only as an instantaneous input stream. After feature extraction by the model, it is immediately discarded, and the raw image data is not retained in any non-volatile storage medium on the device, thus physically eliminating the possibility of leakage of raw biometric data.

[0037] Step S12: Use the second feature detection model to perform feature detection on the visual hidden image to obtain the registration features.

[0038] The visually hidden image is obtained by visually perturbing the original registration image and stored in the face recognition device, while the original registration image is not stored in the face recognition device.

[0039] Specifically, during the initialization phase of product registration, the original registration images of legitimate users need to be loaded into memory. After extracting the real-time high-dimensional feature vectors, the system directly traverses the base database at the computing resource layer. For more details, please refer to... Figure 4 The method of this application also includes: Step S31: Use the near-infrared camera to acquire the original registration image.

[0040] Combination Figure 8 The original registration images obtained are used as registration set data.

[0041] Step S32: Visually perturb the original registration image to obtain the visually hidden image, and store the visually hidden image in the face recognition device.

[0042] Specifically, after obtaining the original registration image, the original registration image, i.e., the registration set data, is subjected to registration set preprocessing. Registration set preprocessing includes image cropping, size scaling, noise reduction, filtering, etc., and the specifics are not limited.

[0043] After preprocessing the registration set, local domain transformation processing is performed to obtain intermediate representations. Specifically, transform domain processing is applied to the original registration image. In essence, the system converts the original registration image into DCT domain data, transforming its representation from a standard pixel-domain representation to a frequency-domain or transform-domain representation. This step provides an intermediate data foundation for subsequent visual information hiding processing.

[0044] After obtaining the transform domain data, the system reconstructs the data according to preset channel parameters to obtain an intermediate image representation, i.e., a visual hidden image, used for visual perturbation processing. Combined with... Figure 5and Figure 6 , Figure 5 For the original registration image, Figure 6 Hides the image visually.

[0045] In one embodiment, visual perturbation processing may include, for example, encrypting the original registered image. Image encryption may include, for example, pixel-level encryption (symmetric or asymmetric), chaotic encryption, transform domain encryption, reversible information hiding, and visual hiding, etc., and is not specifically limited.

[0046] The visual hiding model is invoked to generate visual hiding data. Specifically, the second feature detection model (i.e., the visual hiding model) is invoked to extract features from the visual hiding image to obtain registration features. Understandably, these registration features are the visual hiding data. After obtaining the visual hiding data, it is saved, and the plaintext data is cleared, as shown in step S33.

[0047] Step S33: Delete the original registration image.

[0048] In one embodiment, after clearing the plaintext data, noise is written to the original plaintext address to reduce the risk of recovery. Specifically, the underlying memory overwrite interface is invoked to overwrite the data in the physical memory area where the original registration image is stored, thereby visually hiding the original registration image. This data overwriting includes writing all-zero data or writing noisy data.

[0049] Step S13: Perform identity recognition based on the feature information and the registration features.

[0050] Specifically, in the scheme of this application, during the local registration phase, after the system receives the original registration image collected by the user on-site, it first performs visual perturbation on the original registration image to generate a visually hidden image that is completely decoupled from the original visual information. Immediately after the visual hiding is completed, an active destruction mechanism is triggered—calling the underlying memory overwrite interface to write a matrix of all zeros or random noise to the physical memory area storing the original face image data, ensuring that the plaintext data is unrecoverable. Subsequently, the system calls the registration feature extraction model (i.e., the second feature detection model) to extract deep features from the visually hidden image to obtain the registration features. No original registration image resides within the device throughout the entire process; the recognition phase only loads the registration features for comparison, without needing to access the original plaintext, thus reducing the risk of information leakage.

[0051] Specifically, the similarity between the feature information and the registration features is calculated; the identity information of the detected target is determined based on the similarity. It should be noted that registration features can be unique or numerous. For example, in financial identification, registration features only include the registration information of a unique user. However, in fields such as access control, there are many registration features. For example, in residential community access control identification, registration features include the residents within the community.

[0052] Therefore, it is necessary to calculate the similarity between the feature information and each registered feature. If a registered feature with a similarity greater than the similarity threshold is identified, the identity verification passes. If no registered feature with a similarity greater than the similarity threshold is identified, the identity verification fails.

[0053] Furthermore, when acquiring images to be identified, the near-infrared camera is used to acquire n consecutive images to be identified within a preset time window; if the similarity of m consecutive images to be identified within the preset time window is less than the similarity threshold, the identity recognition fails and an alarm is triggered.

[0054] In one embodiment, the similarity between the feature information and the registered features is calculated using the following formula;

[0055] in, Representing feature information, Indicates registration characteristics, This represents the L2 norm. This application uses vector norm distance for similarity calculation, without restoring the original image or decrypting registration features, thus overcoming privacy leaks, cloud dependency, and performance bottlenecks, achieving highly secure, fully offline, and low-latency edge recognition capabilities.

[0056] In one embodiment, the method of this application performs identity recognition based on the similarity of n consecutive image frames within a preset time window. Specifically, assuming there are m image frames within the preset time window, if the similarity of n consecutive image frames is less than a similarity threshold, the recognition fails, and an alarm is triggered.

[0057] Specifically, the system uses Z-Score ( The relative dispersion is used to select the most specific Top-1 candidate for a single frame or to determine that the identification fails.

[0058] Subsequently, the system sends the single-frame result to a sliding window queue (i.e., a preset time window) for voting. Only the identity receiving more than half of the votes wins. If the winner fails the identification, the system will further trigger a bistable blanking counter. It must be confirmed as a failed identification for two consecutive frames before it will finally be downgraded to a red box, completely eliminating screen flicker. The final result is directly output and linked to edge devices to open doors or sound alarms. The memory is released immediately after the current frame is processed, and the system smoothly transitions to the next frame.

[0059] Specifically, a sliding window is set up, which contains 5 frames of images. Voting is performed based on the similarity of these 5 frames. If 4 of the frames pass the identification, the identity recognition is successful. If 2, 3, or 4 consecutive frames fail the vote, the identity recognition fails.

[0060] Specific combination Figure 7 This application provides a detailed explanation of its method. Specifically, after system startup, a video stream is retrieved from the visual processing library. The raw frames captured by the camera are only used as instantaneous input into the processing pipeline. The CPU first performs preprocessing, adjusting the resolution to adapt to the face detection model, converting BGR to RGB, and adding dimensions to align the tensor input. After this step, the raw image data is immediately discarded and not written to any non-volatile storage, thus preventing physical leakage. Specifically, the video stream captured by the camera can also undergo frame skipping processing; that is, if the video stream contains 10 images, only the odd-numbered images are processed, or only the even-numbered images are processed; or, processing is performed every 2-3 images, with no specific limitation. Preprocessing includes image cropping in addition to BGR to RGB conversion.

[0061] The preprocessed tensor data is fed into a face detection model deployed on an NPU. The NPU quickly outputs predicted prior bounding boxes and confidence scores through forward propagation. Subsequently, the data is returned to the CPU for post-processing: the system converts the bounding boxes from center coordinates to corner coordinates and performs non-maximum suppression (NMS) based on a set confidence threshold to eliminate redundant detection boxes, ultimately accurately locating the face coordinates (x1, y1, x2, y2) in the image.

[0062] To capture more complete facial information and adapt to complex environments, the detection bounding box was expanded outward and protected against out-of-bounds errors when extracting the Region of Interest (ROI). The cropped face was then scaled to 128×128 and fed into the first feature detection model (i.e., the probe face feature extraction model). The model converts the pixel matrix into a high-dimensional abstract feature vector as feature information (i.e., probe face features), and then no longer touches the original visual data.

[0063] During the initialization phase, the system loads the visual hidden images of legitimate users into memory. Real-time high-dimensional feature vectors are extracted from these visual hidden images and used as registration features.

[0064] Specifically, during the local registration phase, after receiving the original face registration image captured by the user on-site, the system first performs visual perturbation on the original registration image to generate a visually hidden image completely decoupled from the original visual information. Immediately after visual hiding is completed, an active destruction mechanism is triggered—calling the underlying memory overwrite interface to write a matrix of all zeros or random noise to the physical memory area storing the original face image data, ensuring that the plaintext data is unrecoverable. Subsequently, the system calls the second feature detection model (i.e., the visually hidden face feature extraction model) to extract deep features from the visually hidden image to obtain the visually hidden face features (i.e., registration features). No original face image resides within the device throughout the entire process; the recognition phase only loads the registration features for comparison, without needing to access the original plaintext.

[0065] In one embodiment, this application can also bypass the garbage collection mechanism of the programming language itself. Instead of relying on the feature extraction process of the image frame itself, it immediately calls the underlying system interface to actively write all "0"s or a random noise matrix to the physical memory block storing the original face data, achieving a higher level of data fragmentation. In another embodiment, video stream acquisition and preprocessing are not performed in the operating system's conventional rich execution environment, but are directly sent to the chip's TrustZone physical isolation area. After feature extraction, the image pointer is directly destroyed within the TEE, ensuring that even if the device system is rooted, it is absolutely impossible to capture transient raw video frames in memory.

[0066] The comparison principle uses vector norm distance to calculate similarity (feature distance calculation):

[0067] The system employs vector norms derived from asymmetric matrix expansion for high-speed, fully dense-state alignment, without reconstructing the original image or requiring any decryption of the underlying database features. After calculating the distance across the entire database, the system uses Z-Score (… The relative dispersion is used to filter out the most specific Top-1 candidates for a single frame or to determine them as Unknown. In one embodiment, Euclidean distance can also be used to measure the similarity in the feature space.

[0068] Understandably, when performing similarity calculations, the registered feature (i.e., the registration set ID) that is closest to the feature information is selected as the current recognition result.

[0069] Furthermore, multi-frame voting can be implemented. The system sends the results of each frame to a sliding window queue for voting. Only the identity that receives more than half of the votes wins. If the winner is an unregistered identity, the system will further trigger a bistable blanking counter. Only after two consecutive frames confirming that the identity is unregistered will it be downgraded to a red box, completely eliminating screen flicker. The final result is directly output and linked to edge devices for door opening or alarm activation. Memory is released immediately after the current frame is processed, and the system smoothly transitions to the next frame.

[0070] The method described in this application calculates the distance between real-time features and registered features locally and generates a decision result without restoring the original image of the registered face, achieving blind recognition where "data is available but not visible." This design, through a security compensation mechanism of "visual hiding (encryption) followed by destruction," ensures the absolute security of the base database data while granting edge devices complete offline autonomous decision-making capabilities without relying on external security media.

[0071] Finally, existing security systems often separate "front-end data acquisition," "feature extraction," and "visual concealment computation" into independent functional modules, which requires data to be transferred and converted multiple times between different modules, increasing system latency and power consumption.

[0072] The method of this invention can be applied to edge chips, fully utilizing the heterogeneous computing characteristics of the edge chip's NPU and CPU to pipeline the acquisition stream, feature extraction stream, and visual hiding computation stream. The system directly extracts and compares the acquired data in memory in real time, eliminating the overhead of data storage and repeated reading in intermediate links, and realizing millisecond-level real-time processing of the entire process from image acquisition to feature comparison, achieving a high-security, fully offline face recognition system.

[0073] Specifically, the facial recognition device of this application includes the following functional features: Transient video stream access module: Responsible for connecting to external cameras or RTSP video streams, and directly reading video frames into volatile memory for preliminary parsing.

[0074] Preprocessing module: Uses the CPU to perform lightweight operations such as image scaling and color space conversion, and organizes the image into the tensor format required by the NPU.

[0075] Multi-path NPU model inference module: It contains three independent lightweight deep learning models—a face detection model and two feature extraction models, all of which are hardware-accelerated inferences on the NPU.

[0076] Direct comparison module (no decryption): It directly compares the real-time visible features with the features of the visual hidden registration set through vector distance calculation, without decryption, and then performs multi-frame voting.

[0077] This application proposes a one-way, non-distributed, real-time video stream processing method to address privacy compliance requirements in facial recognition. It comprehensively applies real-time frame parsing technology, an instantaneous destruction mechanism, and a storage-free processing protocol. Raw video data is read as streaming data only in volatile memory for edge-side preprocessing and face coordinate localization. Once processing is complete, the current frame image data is immediately released from memory. This architecture also covers the plaintext destruction process during the registration phase, ensuring that the original face image is "not stored, leaves no trace, and is unrecoverable" throughout the entire system lifecycle, physically isolating the risk of privacy leakage.

[0078] To address privacy and security concerns in the feature comparison process, a direct comparison (without decryption) scheme based on irreversible high-dimensional features is proposed. A lightweight pre-trained model on the device side is used to "vectorize" the transiently stored face regions in memory, generating high-dimensional abstract features completely decoupled from the original visual information. During authentication, the vector distance between the real-time features and the registered features in the loaded registration set is calculated directly in memory, and the result is output according to a security threshold. The system never decrypts the original photos in the registration set; the entire process involves direct feature-level matching and visual hiding-level processing, eliminating the need for image reconstruction.

[0079] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, such as... Figure 9 As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604. Memory 603 is used to store computer programs; When the processor 601 executes the program stored in the memory 603, it implements the steps of the above-mentioned visual hidden near-infrared face recognition method based on edge computing chip.

[0080] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0081] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0082] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0083] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0084] The present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program implements the steps of the above-described method for visually concealed near-infrared face recognition based on an edge computing chip.

[0085] Optionally, the computer-readable storage medium may be non-volatile memory (NVM), such as at least one disk storage device.

[0086] Optionally, the computer-readable storage medium may also be at least one storage device located remotely from the aforementioned processor.

[0087] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform the steps of the above-described visual hidden near-infrared face recognition method based on an edge computing chip.

[0088] It should be noted that, for the embodiments of electronic devices / storage media / computer programs, since they are basically similar to the method embodiments, the description is relatively simple. For relevant parts, please refer to the description of the method embodiments. All embodiments of the above-mentioned visual hidden near-infrared face recognition method based on edge computing chips are applicable to the electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0089] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0091] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0092] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (devices), or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects, all of which are collectively referred to herein as "modules" or "systems." Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The computer program may be stored / distributed in a suitable medium, provided with or as part of other hardware, or may take other distribution forms, such as via the Internet or other wired or wireless telecommunications systems.

[0093] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A visually concealed near-infrared face recognition method based on an edge computing chip, characterized in that, The method is applied to a facial recognition device, which includes a near-infrared camera; the facial recognition device includes a first feature detection model and a second feature detection model, and the method includes: The near-infrared camera is used to acquire an image to be identified, and the first feature detection model is used to perform feature detection on the image to be identified to obtain feature information corresponding to the target. The second feature detection model is used to perform feature detection on the visually hidden image to obtain registration features; Identity recognition is performed based on the feature information and the registration features; wherein, the visually hidden image is obtained by visually perturbing the original registration image and stored in the face recognition device, and the original registration image is not stored in the face recognition device.

2. The method according to claim 1, characterized in that, The method further includes: The original registration image was acquired using the near-infrared camera. The original registration image is visually perturbed to obtain the visually hidden image, and the visually hidden image is stored in the face recognition device; Delete the original registration image.

3. The method according to claim 2, characterized in that, Deleting the original registered image includes: The underlying memory overwrite interface is invoked to overwrite the data in the physical memory area where the original registration image is stored, thereby deleting the original registration image; The data overlay includes writing all-zero data or writing noisy data.

4. The method according to claim 1, characterized in that, The method further includes: After acquiring the image to be identified using the near-infrared camera, the image is stored in volatile memory; or, After obtaining the feature information corresponding to the target to be detected, the image to be identified is deleted.

5. The method according to claim 1, characterized in that, The facial recognition device also includes a facial detection model; The first feature detection model is used to perform feature detection on the image to be identified to obtain feature information corresponding to the target, including: The image to be identified is preprocessed to obtain the image to be processed; The face detection model is used to perform face detection on the image to be processed, and the predicted detection box and confidence score corresponding to the detection target are determined. The actual detection box corresponding to the detection target is determined based on the predicted detection box and the confidence score. The first feature detection model is used to perform feature detection based on the actual detection box corresponding to the detection target to obtain the feature information corresponding to the detection target.

6. The method according to claim 1, characterized in that, Identity recognition based on the aforementioned feature information and registration features includes: Calculate the similarity between the feature information and each of the registered features; The identity information of the detected target is determined based on the similarity.

7. The method according to claim 6, characterized in that, Acquiring an image to be identified using the near-infrared camera includes: The near-infrared camera is used to acquire n consecutive images of the image to be identified within a preset time window; Determining the identity information of the detected target based on the similarity includes: If the similarity of m consecutive images to be identified within a preset time window is less than the similarity threshold, the identity recognition fails and an alarm is triggered.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor coupled to each other. The processor is used to execute program instructions stored in the memory and to execute program data to implement the steps in the visually hidden near-infrared face recognition method based on an edge computing chip as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the visually hidden near-infrared face recognition method based on an edge computing chip as described in any one of claims 1 to 7.

Images