Multi-stage depth feature face recognition method based on edge calculation

Abstract

The invention discloses a multi-stage depth feature face recognition method and device based on edge calculation, equipment and a medium, and is suitable for edge equipment with limited resources. The method comprises the following steps: detecting an input image through a lightweight face detection network with a multi-scale feature pyramid structure, and outputting a face bounding box and confidence; cutting the face area according to the bounding box and carrying out standardization processing; inputting the standardized image into a deep residual convolutional neural network, and extracting a normalized feature vector; adopting a distributed vector database to construct a face feature library for partition isolation management to store the vector; and receiving a to-be-recognized face image, obtaining a normalized feature vector of the to-be-recognized face image, performing matching in the feature library based on the feature similarity, and outputting a recognition result. The method does not need to depend on the cloud, improves the recognition real-time performance and the data security, achieves high-precision recognition in a complex scene, and is suitable for edge computing scenes such as security monitoring and access control management.

Description

Technical Field

[0001] This application relates to the fields of computer vision, artificial intelligence and edge computing technology, and in particular to a multi-stage deep feature face recognition method, apparatus, medium and device based on edge computing. Background Technology

[0002] With the continuous development of artificial intelligence technology, facial recognition, as an important biometric identification technology, has been widely applied in fields such as security monitoring, access control, smart parks, financial risk control, and public safety. Existing facial recognition systems typically rely on deep learning models to analyze faces in images or video streams to achieve automatic identification and comparison of individuals. However, in practical applications, facial recognition often needs to be deployed in edge computing environments such as front-end camera devices, smart terminals, or embedded devices. These devices typically have limited computing power, restricted storage resources, and strict power consumption requirements, posing significant challenges to the practical application of high-precision facial recognition algorithms. Some existing solutions still rely on cloud servers for centralized computing, which not only increases system latency but also suffers from strong network dependence, high data transmission costs, and privacy and security risks, making it difficult to meet the demands for real-time and localized data processing.

[0003] On the other hand, facial recognition in complex scenarios faces challenges such as drastic changes in lighting, diverse facial poses, frequent occlusion, inconsistent resolution, and complex backgrounds. Some existing technologies fail to perform refined processing of the facial region during facial recognition, directly extracting features from the entire image or a roughly cropped area. This results in extracted features containing a large amount of irrelevant background information, thus affecting recognition accuracy. Furthermore, the existing facial recognition process lacks a unified and standardized multi-stage collaborative mechanism between processing stages, making it difficult to ensure recognition accuracy while simultaneously maintaining real-time processing capabilities and system stability at the edge. Summary of the Invention

[0004] The main objective of this application is to provide a multi-stage deep feature face recognition method, apparatus, medium, and device based on edge computing. The aim is to provide a face recognition method that is suitable for edge computing devices, has multi-stage processing capabilities, and can operate stably in complex environments, thereby improving the reliability and practicality of face recognition systems in practical applications.

[0005] To achieve the above objectives, this application provides a multi-stage deep feature face recognition method based on edge computing, comprising: performing target detection on an input image using a lightweight face detection network based on a multi-scale feature pyramid structure, outputting a set of face bounding box coordinates and corresponding detection confidence scores; cropping a face region image from the original image based on the face bounding box coordinate set, and standardizing the cropped face region image to obtain a standardized face region image; inputting the standardized face region image into a deep residual convolutional neural network for feature extraction and inference to obtain a normalized feature vector; using a distributed vector database as a persistent storage medium for face features, and establishing a face feature library architecture with matching partition isolation management in the distributed vector database to store the normalized feature vectors; receiving a face image to be recognized, obtaining the normalized feature vector of the face image to be recognized, matching the normalized feature vector of the face image to be recognized in the face feature library architecture based on feature similarity, and outputting a face recognition judgment result.

[0006] Optionally, the lightweight face detection network based on a multi-scale feature pyramid structure performs target detection on the input image and outputs a set of face bounding box coordinates and corresponding detection confidence scores. This includes: performing aspect ratio scaling and boundary padding preprocessing on the input image to adjust it to the standard input size required by the detection network, obtaining face images at different scales; performing multi-layer feature map fusion based on the detection network to detect face images at different scales, outputting multi-branch prediction results including bounding box position offsets, target confidence scores, and face key point positions; and performing affine transformation decoding on the bounding box position offsets and anchor box parameters to convert the relative offsets... The bounding box position offsets in the multi-branch prediction results are converted to absolute coordinates through anchor box center point offset and scale transformation, resulting in a set of face bounding box coordinates. Based on the target confidence in the decoded multi-branch prediction results, the set of face bounding box coordinates is filtered to remove candidate bounding boxes with confidence scores below a preset detection threshold, obtaining the candidate bounding boxes with the highest confidence scores for face images at different scales. A non-maximum suppression algorithm with a greedy iterative strategy is used to remove duplicate bounding boxes among the candidate bounding boxes with the highest confidence scores, outputting the final set of face bounding box coordinates and the corresponding detection confidence scores.

[0007] Optionally, the step of cropping a face region image from the original image based on the face bounding box coordinate set and standardizing the cropped face region image to obtain a standardized face region image includes: scaling the cropped face region image to the standard input size required by the feature extraction network using a bilinear interpolation algorithm to obtain a scaled face region image; automatically selecting a color space conversion method based on the number of input channels of the feature extraction network; if the number of input channels is single-channel, performing grayscale conversion on the scaled face region image; if the number of input channels is three-channel, performing color channel order conversion on the scaled face region image to obtain a color-converted face image; and performing pixel value centering and normalization processing on the color-converted face image, converting the normalized face image data into a tensor format required for network inference to obtain a standardized face region image.

[0008] Optionally, the step of inputting the standardized face region image into a deep residual convolutional neural network for feature extraction and inference to obtain a normalized feature vector includes: constructing a deep residual network using a multi-layer residual block stacking structure; learning the deep feature map of the standardized face region image through the skip connections of the deep residual network; reducing the deep feature map to a fixed-dimensional original feature vector based on the global average pooling layer at the end of the deep residual network; and performing L2 norm normalization on the original feature vector to obtain a normalized feature vector with a modulus normalized to a unit length.

[0009] Optionally, the step of performing L2 norm normalization on the original feature vector to obtain a normalized feature vector with a modulus normalized to a unit length includes: calculating the square root of the sum of squares of each element of the original feature vector as the norm value, and dividing each element of the original feature vector by the norm value to obtain the normalized feature vector.

[0010] Optionally, the step of using a distributed vector database as the persistent storage medium for facial features, and establishing a matching partitioned and isolated facial feature library architecture within the distributed vector database to store normalized feature vectors, includes: generating a unique partition name and corresponding partition vector index through a facial library identifier to obtain a distributed vector database for physical data isolation and independent management between different facial libraries; obtaining the normalized feature vector of the facial image to be registered, associating the normalized feature vector of the facial image to be registered with a unique identifier, inserting it into the vector index structure of the corresponding partition in the distributed vector database, using an asynchronous persistence mechanism to complete the vector insertion of the facial image to be registered, and returning the registration result; performing a data flush operation on the inserted vector in a background thread to establish a partition-level data caching mechanism to obtain a partitioned and isolated facial feature library architecture to cache the personnel list information corresponding to each partition and the facial image to be registered; and setting a distributed vector database cache expiration policy to automatically clear the cached data of the corresponding partition when the facial image data to be registered changes.

[0011] Optionally, the lightweight face detection network is used to: determine the feature maps of each layer of the input image based on the downsampling step size of each layer of the feature pyramid and the preset minimum anchor box size configuration; generate multiple anchor boxes of different sizes at each spatial location of each layer of the feature map using a dense sliding window strategy; and merge the anchor boxes generated by each layer of the feature map into a unified anchor box set; wherein, the coordinates of the anchor box center point are calculated based on the feature map position and the downsampling step size, and the width and height of the anchor box are determined based on the preset minimum anchor box size configuration; perform affine transformation decoding on the bounding box position offset and anchor box parameters to convert the relative offset into absolute coordinates; perform inverse transformation on the bounding box coordinates based on the scaling factor recorded in the preprocessing to restore them to the original image coordinate system; process the confidence output with an activation function to obtain probability values, and filter candidate boxes that meet the detection threshold; and use a greedy iterative strategy non-maximum suppression algorithm to traverse the candidate boxes in descending order of confidence, calculate the intersection-union ratio and suppress overlapping boxes.

[0012] Furthermore, to achieve the above objectives, this application also provides a multi-stage deep feature face recognition device based on edge computing, comprising: a target detection module, used to perform target detection on the input image based on a lightweight face detection network with a multi-scale feature pyramid structure, and output a set of face bounding box coordinates and corresponding detection confidence; an image normalization processing module, used to crop a face region image from the original image according to the face bounding box coordinate set, and perform normalization processing on the cropped face region image to obtain a normalized face region image; a normalization module, used to input the normalized face region image into a deep residual convolutional neural network for feature extraction and inference to obtain a normalized feature vector; a feature library construction module, used to use a distributed vector database as a persistent storage medium for face features, and to establish a face feature library architecture with matching partition isolation management in the distributed vector database to store the normalized feature vector; and a feature matching module, used to receive a face image to be recognized, obtain the normalized feature vector of the face image to be recognized, match the normalized feature vector of the face image to be recognized in the face feature library architecture based on feature similarity, and output the face recognition judgment result.

[0013] To achieve the above objectives, this application also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the multi-stage deep feature face recognition method based on edge computing provided in the above embodiments.

[0014] To achieve the above objectives, this application also provides an electronic device, the electronic device comprising: at least one processor, a memory, and an input / output unit; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the multi-stage deep feature face recognition method based on edge computing provided in any of the foregoing embodiments.

[0015] This application proposes a multi-stage deep feature face recognition method, apparatus, medium, and device based on edge computing. It employs a lightweight detection network based on a multi-scale feature pyramid structure, combined with a pre-defined anchor box mechanism and a non-maximum suppression algorithm, to achieve efficient face detection at different scales. Secondly, an adaptive model configuration mechanism is introduced, automatically selecting grayscale or color processing mode based on the number of input channels of the feature extraction network, improving the system's versatility and scalability. Thirdly, a partitioned face feature library management architecture is designed, using a distributed vector database to achieve physical data isolation between different face libraries, effectively solving the data organization problem in multi-library management scenarios. Furthermore, an asynchronous persistence mechanism and partitioned caching management strategy are adopted, returning a response immediately after feature insertion, significantly improving registration efficiency. Finally, a deep residual network is used to extract feature vectors and perform L2 norm normalization, combined with cosine similarity measurement and an approximate nearest neighbor retrieval algorithm. This application achieves real-time processing capabilities while ensuring recognition accuracy, making it particularly suitable for edge computing applications such as security monitoring, access control, and smart parks. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the multi-stage deep feature face recognition method based on edge computing in this application.

[0017] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0019] Reference Figure 1 The first embodiment of this application provides a multi-stage deep feature face recognition method based on edge computing, which can: S101. A lightweight face detection network based on a multi-scale feature pyramid structure performs target detection on the input image and outputs a set of face bounding box coordinates and the corresponding detection confidence. In one embodiment of this application, the process of using a lightweight face detection network based on a multi-scale feature pyramid structure to detect objects in an input image and outputting a set of face bounding box coordinates and corresponding detection confidence scores may include the following: S1011. Perform aspect ratio scaling and boundary padding preprocessing on the input image to adjust the input image to the standard input size required by the detection network, thereby obtaining face images at different scales; S1012. Multi-layer feature map fusion based on detection network is used to detect face images of different scales and output multi-branch prediction results including bounding box position offset, target confidence and face key point position. S1013. Perform affine transformation decoding on the bounding box position offset and anchor box parameters, convert the relative offset into absolute coordinates, and convert the bounding box position offset in the multi-branch prediction result into actual bounding box coordinates through anchor box center point offset and scale transformation, to obtain the set of face bounding box coordinates. S1014. Based on the target confidence in the decoded multi-branch prediction results, the set of face bounding box coordinates is filtered to filter out candidate bounding boxes with confidence below a preset detection threshold, so as to obtain the candidate bounding boxes with the highest confidence for face images of different scales. S1015. The non-maximum suppression algorithm with a greedy iterative strategy is used to remove duplicate bounding boxes in the candidate bounding boxes with the highest confidence, and the final set of face bounding box coordinates and corresponding detection confidence is output.

[0020] The execution process of step S1011 may include the following: Calculate the width scaling ratio and height scaling ratio of the input image and the standard input size of the detection network, and select the smaller one as the uniform scaling factor to maintain the aspect ratio of the image; calculate the scaled target size based on the scaling factor, and perform bilinear interpolation scaling on the image; calculate the difference between the scaled image and the standard input size, fill the image boundary with zero values ​​to meet the network input requirements, and record the filling parameters for subsequent coordinate restoration; The execution process of step S1012 may include the following: Based on the downsampling step size of each layer of the feature pyramid and the preset minimum anchor box size configuration, multiple anchor boxes of different sizes are generated at each spatial location of the feature map of each layer. The anchor box generation adopts a dense sliding window strategy. The coordinates of the anchor box center point are calculated based on the feature map position and downsampling step size, and the width and height of the anchor box are determined according to the preset minimum size configuration. The anchor boxes generated by each layer of feature map are merged into a unified anchor box set for subsequent decoding. The execution process of step S1013 may include the following: The bounding box offsets and anchor box parameters output by the network are decoded by affine transformation to convert the relative offsets into absolute coordinates. The bounding box coordinates are then inversely transformed according to the scaling factor recorded in the preprocessing to restore them to the original image coordinate system. The confidence output is processed by an activation function to obtain probability values, and candidate boxes that meet the detection threshold are selected. A non-maximum suppression algorithm with a greedy iterative strategy is used to traverse the candidate boxes in descending order of confidence, calculate the intersection-union ratio and suppress overlapping boxes, and output the final detection result.

[0021] Clearly, through a multi-stage cascaded architecture, this embodiment achieves high-precision recognition of facial targets in complex scenes on edge computing devices, reducing reliance on cloud computing and improving system real-time performance and data security.

[0022] In this embodiment, a lightweight face detection network based on the SCRFD architecture is used to detect faces in the input image. First, the input video stream is decoded using a hardware-accelerated decoder, converting video frames into image data formats stored in the device memory. The decoded image is then scaled proportionally while maintaining its aspect ratio. The smaller of the width and height scaling ratios is selected as the uniform scaling factor. Bilinear interpolation is used to adjust the image to the network's required standard input size of 640×640 pixels, and zero-value padding is applied to the image boundaries. Color space conversion is performed using a visual computing library, converting the YUV420P format to the BGR planar format. Pixel value normalization preprocessing is then performed using a mean of 127.5 and a scaling factor of 0.0078125.

[0023] Then, the preprocessed image data is input into the SCRFD model for forward inference. The SCRFD network uses MobileNet 0.25 as its backbone and achieves multi-scale face detection through a three-layer feature pyramid structure, with corresponding downsampling step sizes of 8, 16, and 32. The network output includes three branches: bounding box position offset, target confidence, and the positions of five facial landmarks.

[0024] Next, a preset anchor box mechanism is used to decode the detection results. The anchor box size is configured with three levels corresponding to face detection at [16,32], [64,128], and [256,512] pixels respectively. The bounding box offsets output by the network are decoded using an affine transformation, with a variance parameter of [0.1, 0.2]. A detection confidence threshold of 0.5 is set for initial screening, and a non-maximum suppression algorithm with an NMS threshold of 0.3 is used to remove duplicate detection boxes. Finally, the coordinates of the face bounding box and the coordinates of five key points are output.

[0025] S102. Based on the set of face bounding box coordinates, crop the face region image from the original image, and standardize the cropped face region image to obtain a standardized face region image. In one embodiment of this application, the execution process of step S102 may include the following: S1021. The cropped face region image is scaled to the standard input size required by the feature extraction network using a bilinear interpolation algorithm to obtain the scaled face region image. S1022. Automatically select the color space conversion method according to the number of input channels of the feature extraction network. If the number of input channels is a single channel, perform grayscale conversion on the scaled face region image. If the number of input channels is three channels, perform color channel sequence conversion on the scaled face region image to obtain the color-converted face image. S1023. Perform pixel value centering and normalization processing on the color-converted face image, and convert the normalized face image data into the tensor format required for network inference to obtain the standardized face region image.

[0026] Specifically, the execution process of step S1022 may include the following: During model initialization, the input tensor shape configuration of the feature extraction network is read, and the specific values ​​of batch dimension, channel number dimension, height dimension, and width dimension are obtained. The input type of the model is determined based on the channel number dimension: when the number of channels is 1, it is marked as grayscale input mode, and when the number of channels is 3, it is marked as color input mode. When the input mode is grayscale, a weighted grayscale conversion is performed on the cropped face region, and the weighting coefficient is determined according to the sensitivity of the human eye to different color channels. When the input mode is color mode, the color channel order is adjusted to match the channel configuration during network training. Specifically, the execution process of step S1022 may include the following: A centering linear transformation is used to normalize the pixel values. The original pixel values ​​are subtracted from the center offset and then multiplied by a scaling factor so that the normalized values ​​are distributed within a standard range centered at zero. The normalization parameters are consistent with the network training configuration.

[0027] In this embodiment, for each face bounding box detected in step S101, the crop function from the visual computing library is used to crop the face region from the original image, ensuring the validity of the crop coordinates. The cropped region is then adjusted to the standard input size required by the feature extraction network using a bilinear interpolation algorithm.

[0028] The standardized processing supports adaptive model configuration. During model initialization, it automatically reads the shape of the input tensor of the feature extraction network and determines the model input type based on the number of channels. When the number of input channels is 1, it is marked as grayscale input mode, and a weighted grayscale conversion is performed with weight coefficients of 0.299 for the red channel, 0.587 for the green channel, and 0.114 for the blue channel. When the number of input channels is 3, it is marked as color input mode, and a BGR to RGB color channel order conversion is performed.

[0029] A centered linear transformation is used to normalize pixel values. The original pixel values ​​are subtracted by 127.5 and then multiplied by a scaling factor of 0.0078125, resulting in normalized values ​​ranging from -1 to 1. Finally, the image data is converted from HWC format to NCHW format as the network input tensor.

[0030] S103. Input the standardized face region image into a deep residual convolutional neural network for feature extraction and inference to obtain the normalized feature vector: The process of inputting the standardized face region image into a deep residual convolutional neural network for feature extraction and inference to obtain the normalized feature vector can include the following: S1031. A deep residual network is constructed using a multi-layer residual block stacking structure. The deep feature map of the standardized face region image is learned through the skip connections of the deep residual network. S1032. The global average pooling layer at the end of the deep residual network reduces the deep feature map to the original feature vector of fixed dimension. S1033. Perform L2 norm normalization on the original feature vector to obtain a normalized feature vector with the modulus standardized to a unit length.

[0031] For example, the face region after standardization in step S1022 is input into a deep residual convolutional neural network based on the ArcFace architecture for feature extraction and inference. The ArcFace network uses a ResNet residual network as its backbone architecture, and achieves efficient learning of deep features through multi-layer residual block stacking and skip connections. At the end of the network, a global average pooling layer compresses the feature map into a fixed-length feature vector of 512 dimensions.

[0032] The extracted original feature vectors are normalized using the L2 norm. The square root of the sum of the squares of all elements in the feature vector is calculated as the L2 norm value. Each element is then divided by this norm value to make the normalized feature vector have a magnitude of 1. L2 normalization eliminates the influence of differences in feature magnitudes, making subsequent cosine similarity calculations more stable and reliable.

[0033] S104. A distributed vector database is used as the persistent storage medium for facial features, and a matching partitioned isolation management facial feature library architecture is established in the distributed vector database to store the normalized feature vectors. Specifically, the execution process of step S104 may include the following: The process of using a distributed vector database as a persistent storage medium for facial features, and establishing a matching, partitioned, and isolated facial feature library architecture within the distributed vector database to store normalized feature vectors, can include the following: S1041. Generate a unique partition name and corresponding partition vector index through the face database identifier to obtain a distributed vector database for physical data isolation and independent management between different face databases; S1042. Obtain the normalized feature vector of the face image to be registered, associate the normalized feature vector of the face image to be registered with the unique identifier, insert it into the vector index structure of the corresponding partition in the distributed vector database, use an asynchronous persistence mechanism to complete the vector insertion of the face image to be registered, and return the registration result. S1043. Perform a data flushing operation on the inserted vector in the background thread, establish a partition-level data caching mechanism, and obtain the face feature library architecture for partition isolation management, so as to cache the list information of personnel corresponding to each partition and the face image to be registered. S1044. Set a distributed vector database cache expiration policy to automatically clear the cached data of the corresponding partition when the face image data to be registered changes.

[0034] Specifically, the execution process of step S1041 is as follows: The partition name is generated based on the unique business identifier of the face database. The naming rule of concatenating the business identifier with a fixed prefix is ​​used to ensure the uniqueness and resolvability of the partition name. Before performing the face registration operation, the existence of the target partition is checked. If it does not exist, a new partition is created dynamically. Specifically, the execution process of step S1042 is as follows: After feature vector insertion is completed, a separate background daemon thread is started to perform data flushing to disk, and the main thread immediately returns a registration success response; the background thread automatically clears the cached data of the corresponding partition after the flush is completed. Specifically, the execution process of step S1043 is as follows: Maintain an independent personnel list cache and cache timestamp for each partition, and set a uniform cache validity period; when performing a partition query, prioritize checking the cache validity, and if the cache has not expired, return the cached data directly; after performing a data change operation, proactively clear the cache entries for the corresponding partition.

[0035] S105. Receive the face image to be recognized, obtain the normalized feature vector of the face image to be recognized, match the normalized feature vector of the face image to be recognized in the face feature library architecture based on feature similarity, and output the face recognition judgment result.

[0036] In one embodiment of this application, the execution process of step S105 may include the following: S1051. Parse the target face database list parameter in the recognition request, and convert each database identifier into the corresponding partition name; verify whether each partition exists in the vector database, filter invalid partitions and record alarm logs; if all specified partitions are invalid, expand the search scope to all available partitions. S1052. Construct a vector retrieval request, specifying the target partition list and retrieval parameters; use cosine similarity as the distance metric, execute the approximate nearest neighbor retrieval algorithm to return the set of candidate vectors with the highest similarity; the retrieval results include the primary key identifier and similarity score of the candidate vectors; S1053. For single-partition retrieval scenarios, the face database identifier is directly parsed from the partition name as the database to which the hit result belongs. For multi-partition retrieval scenarios, each partition is traversed and its database is determined based on the primary key identifier of the candidate vector. The similarity score is compared with the preset recognition threshold. When the threshold condition is met, the complete recognition result of the identity identifier, database to which it belongs, and similarity score is returned.

[0037] The beneficial results of this invention are as follows: This invention achieves high-precision facial target recognition in complex scenes on edge computing devices through a multi-stage cascaded architecture. First, a lightweight detection network based on a multi-scale feature pyramid structure is adopted, combined with a preset anchor box mechanism and a non-maximum suppression algorithm, to achieve efficient detection of faces at different scales. Second, an adaptive model configuration mechanism is introduced, which automatically selects grayscale or color processing mode according to the number of input channels of the feature extraction network, improving the system's versatility and scalability. Third, a face feature library management architecture based on partitioned isolation is designed, which achieves physical data isolation between different face libraries through a distributed vector database, effectively solving the data organization problem in multi-library management scenarios. In addition, an asynchronous persistence mechanism and a partitioned caching management strategy are adopted, which immediately return a response after feature insertion, greatly improving registration efficiency. Finally, a deep residual network is used to extract feature vectors and perform L2 norm normalization, combined with cosine similarity measurement and approximate nearest neighbor retrieval algorithm, to achieve real-time processing capabilities while ensuring recognition accuracy, making it particularly suitable for edge computing application scenarios such as security monitoring, access control management, and smart parks.

[0038] Based on the above embodiments, this application also provides a multi-stage deep feature face recognition device based on edge computing, including: a target detection module, used to perform target detection on the input image based on a lightweight face detection network with a multi-scale feature pyramid structure, and output a set of face bounding box coordinates and corresponding detection confidence; an image normalization processing module, used to crop a face region image from the original image according to the face bounding box coordinate set, and perform normalization processing on the cropped face region image to obtain a normalized face region image; a normalization module, used to input the normalized face region image into a deep residual convolutional neural network for feature extraction and inference to obtain a normalized feature vector; a feature library construction module, used to use a distributed vector database as a persistent storage medium for face features, and to establish a face feature library architecture with matching partition isolation management in the distributed vector database to store the normalized feature vector; and a feature matching module, used to receive a face image to be recognized, obtain the normalized feature vector of the face image to be recognized, match the normalized feature vector of the face image to be recognized in the face feature library architecture based on feature similarity, and output a face recognition judgment result.

[0039] Based on the above embodiments, this application also provides a computer-readable storage medium including instructions that, when run on a computer, cause the computer to execute the multi-stage deep feature face recognition method based on edge computing provided in the above embodiments.

[0040] Based on the above embodiments, this application also provides an electronic device, which includes: at least one processor, a memory, and an input / output unit; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the multi-stage deep feature face recognition method based on edge computing provided in any of the foregoing embodiments.

[0041] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multi-stage deep feature face recognition method based on edge computing, characterized in that, include: A lightweight face detection network based on a multi-scale feature pyramid structure performs target detection on the input image and outputs a set of face bounding box coordinates and the corresponding detection confidence. Based on the set of face bounding box coordinates, the face region image is cropped from the original image, and the cropped face region image is standardized to obtain the standardized face region image. The standardized face region image is input into a deep residual convolutional neural network for feature extraction and inference, resulting in a normalized feature vector: A distributed vector database is used as the persistent storage medium for facial features, and a matching partitioned and isolated facial feature library architecture is established in the distributed vector database to store the normalized feature vectors. The system receives a face image to be identified, obtains the normalized feature vector of the face image to be identified, matches the normalized feature vector of the face image to be identified in the face feature database architecture based on feature similarity, and outputs the face recognition result.

2. The multi-stage deep feature face recognition method based on edge computing as described in claim 1, characterized in that, The lightweight face detection network based on a multi-scale feature pyramid structure performs target detection on the input image and outputs a set of face bounding box coordinates and corresponding detection confidence scores, including: The input image is preprocessed by scaling its aspect ratio and padding its boundaries to adjust it to the standard input size required by the detection network, thus obtaining face images at different scales. Multi-layer feature map fusion based on detection network is used to detect face images at different scales, and outputs multi-branch prediction results including bounding box position offset, target confidence and face key point position. Affine transformation decoding is performed on the bounding box position offset and anchor box parameters to convert the relative offset into absolute coordinates. The bounding box position offset in the multi-branch prediction result is converted into actual bounding box coordinates through anchor box center point offset and scale transformation, thus obtaining the set of face bounding box coordinates. The set of face bounding box coordinates is filtered based on the target confidence in the decoded multi-branch prediction results to filter out candidate bounding boxes with confidence below a preset detection threshold, thereby obtaining the candidate bounding boxes with the highest confidence for face images of different scales. A non-maximum suppression algorithm with a greedy iterative strategy is used to remove duplicate bounding boxes among the candidate bounding boxes with the highest confidence, and the final set of face bounding box coordinates and corresponding detection confidence is output.

3. The multi-stage deep feature face recognition method based on edge computing as described in claim 1, characterized in that, The step of cropping a face region image from the original image based on the face bounding box coordinate set, and then standardizing the cropped face region image to obtain a standardized face region image, includes: The cropped face region image is scaled to the standard input size required by the feature extraction network using a bilinear interpolation algorithm to obtain the scaled face region image. The color space conversion method is automatically selected based on the number of input channels of the feature extraction network. If the number of input channels is single, grayscale conversion is performed on the scaled face region image. If the number of input channels is three, color channel sequence conversion is performed on the scaled face region image to obtain the color-converted face image. The color-converted face image is subjected to pixel value centering and normalization processing. The normalized face image data is then converted into a tensor format required for network inference, resulting in a standardized face region image.

4. The multi-stage deep feature face recognition method based on edge computing as described in claim 1, characterized in that, The process of inputting the standardized face region image into a deep residual convolutional neural network for feature extraction and inference to obtain a normalized feature vector includes: A deep residual network is constructed using a multi-layer residual block stacking structure. The deep feature maps of the normalized face region image are learned through the skip connections of the deep residual network. The global average pooling layer at the end of the deep residual network reduces the deep feature map to a fixed-dimensional original feature vector. The original feature vectors are normalized using the L2 norm to obtain normalized feature vectors with the modulus standardized to a unit length.

5. The multi-stage deep feature face recognition method based on edge computing as described in claim 4, characterized in that, The step of performing L2 norm normalization on the original feature vector to obtain a normalized feature vector with a modulus normalized to a unit length includes: The square root of the sum of squares of the elements of the original eigenvector is calculated as the norm value. The normalized eigenvector is obtained by dividing each element of the original eigenvector by the norm value.

6. The multi-stage deep feature face recognition method based on edge computing as described in claim 1, characterized in that, The method employs a distributed vector database as the persistent storage medium for facial features, and establishes a matching, partitioned, and isolated facial feature library architecture within the distributed vector database to store normalized feature vectors, including: By generating unique partition names and corresponding vector indexes for each partition using face database identifiers, a distributed vector database is obtained to enable physical data isolation and independent management between different face databases. Obtain the normalized feature vector of the face image to be registered, associate the normalized feature vector of the face image to be registered with the unique identity identifier, insert it into the vector index structure of the corresponding partition in the distributed vector database, use an asynchronous persistence mechanism to complete the vector insertion of the face image to be registered, and return the registration result. In the background thread, a data flushing operation is performed on the inserted vectors to establish a partition-level data caching mechanism, thereby obtaining a partition-isolated face feature library architecture to cache the list of personnel information corresponding to each partition and the face image to be registered. Configure a distributed vector database cache expiration policy to automatically clear the cached data of the corresponding partition when the face image data to be registered changes.

7. The multi-stage deep feature face recognition method based on edge computing as described in claim 1, characterized in that, The lightweight face detection network is used for: The feature maps of each layer of the input image are determined based on the downsampling step size of each layer of the feature pyramid and the preset minimum anchor box size configuration. At each spatial location of each layer of the feature map, a dense sliding window strategy is used to generate multiple anchor boxes of different sizes, and the anchor boxes generated by each layer of the feature map are merged into a unified anchor box set. The coordinates of the anchor frame center point are calculated based on the feature map position and downsampling step size, while the width and height of the anchor frame are determined based on the preset minimum anchor frame size configuration. The bounding box position offset and anchor box parameters are decoded by affine transformation to convert the relative offset into absolute coordinates; the bounding box coordinates are inversely transformed according to the scaling factor recorded in the preprocessing to restore them to the original image coordinate system; the confidence output is processed by activation function to obtain probability values, and candidate boxes that meet the detection threshold are selected; a non-maximum suppression algorithm with a greedy iterative strategy is used to traverse the candidate boxes in descending order of confidence and calculate the intersection-union ratio to suppress overlapping boxes.

8. A multi-stage deep feature face recognition device based on edge computing, characterized in that, include: The object detection module is used to perform object detection on the input image based on a lightweight face detection network with a multi-scale feature pyramid structure, and outputs the set of face bounding box coordinates and the corresponding detection confidence. The image normalization processing module is used to crop the face region image from the original image based on the face bounding box coordinate set, and to normalize the cropped face region image to obtain a normalized face region image. The normalization module is used to input the standardized face region image into a deep residual convolutional neural network for feature extraction and inference, resulting in a normalized feature vector. The feature library construction module is used to use a distributed vector database as a persistent storage medium for facial features, and to establish a matching partitioned and isolated facial feature library architecture in the distributed vector database to store normalized feature vectors. The feature matching module is used to receive the face image to be identified, obtain the normalized feature vector of the face image to be identified, match the normalized feature vector of the face image to be identified in the face feature library architecture based on feature similarity, and output the face recognition judgment result.

9. A computer-readable storage medium, characterized in that, It includes instructions that, when executed on a computer, cause the computer to perform the multi-stage deep feature face recognition method based on edge computing as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes: At least one processor, memory, and input / output unit; The memory is used to store computer programs, and the processor is used to call the computer programs stored in the memory to execute the multi-stage deep feature face recognition method based on edge computing according to any one of claims 1 to 7.

Images