Face recognition method and system suitable for outdoor electronic instrument

Abstract

The present application relates to the technical field of face recognition, and discloses a face recognition method and system suitable for outdoor electronic instruments, aiming at the low recognition accuracy of the prior art under complex illumination, posture change and occlusion conditions, S1, image data is collected and preprocessed; S2, face region detection and face key point positioning are performed on the illumination enhanced image; S3, affine transformation is performed based on the face detection result set to align the face; S4, a face feature vector is extracted based on the ResNet100 structure; S5, posture guided feature reconstruction processing is performed on the face feature vector to obtain a posture enhanced face feature vector; S6, feature normalization and identity classification are performed on the posture enhanced face feature vector, and a face recognition result is output; S7, a final identity recognition result and a confidence level are output. The present application can provide an efficient and scientific optimization scheme in outdoor face recognition, bringing significant technical value and economic benefits to practical applications.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of face recognition, and in particular to a face recognition method and system suitable for outdoor electronic instruments. BACKGROUND

[0002] Under the background of rapid development of face recognition technology, algorithms based on deep learning have been widely applied to security monitoring, mobile terminals and identity verification scenes, however, the existing mainstream face recognition methods mostly rely on stable and controlled light and shooting environment, mainly facing indoor scenes or recognition tasks under controllable image quality conditions, when applied to outdoor electronic instruments such as unmanned aerial vehicles, outdoor monitoring, intelligent inspection terminals and other equipment, limited by multiple factors such as strong light, backlight, rain and fog, posture deflection and equipment computing power, the traditional methods are obviously insufficient in face detection accuracy, alignment accuracy and recognition robustness.

[0003] Current methods attempt to introduce lightweight networks and attention mechanisms, but under uneven light and multi-pose interference, it is still difficult to ensure recognition stability, especially under low-quality image input, with large recognition error and high misjudgment rate, which cannot meet the dual requirements of practicality and stability in outdoor environments, therefore, there is an urgent need for a face recognition method for outdoor complex environments, combining multi-source image processing, adaptive light compensation, high-robust feature extraction and recognition optimization mechanism, to realize a face recognition system with stronger adaptability to environmental changes. SUMMARY

[0004] One object of the present application is to provide a face recognition method and system suitable for outdoor electronic instruments, which can provide an efficient and scientific optimization scheme in outdoor face recognition, bringing significant technical value and economic benefits to practical applications.

[0005] According to the face recognition method suitable for outdoor electronic instruments according to the embodiment of the present application, the following steps are included:

[0006] S1, collecting image data in an outdoor environment, the image data including RGB images and infrared images, and pre-processing the original input image to obtain a light feature vector;

[0007] S2, using a face detection model based on an improved MobileViT structure to detect the face region and locate the face key points of the light enhanced image, and outputting a face detection result set containing the face bounding box and the five facial feature point coordinates;

[0008] S3, performing affine transformation based on the face detection result set to align the face, and obtaining a standardized aligned face image;

[0009] S4, input the standardized aligned face image into a face recognition model based on a ResNet100 structure, the face recognition model embeds a double-channel spatial attention mechanism to extract an attention map, obtains a weighting coefficient by fusing the attention map, and outputs a face feature vector;

[0010] S5, performing pose-guided feature reconstruction processing on the face feature vector, extracting head angle information through a pose estimation module, and using the head angle information for re-weighted fusion of the face feature vector to obtain a pose-enhanced face feature vector;

[0011] S6, performing feature normalization and identity classification on the pose-enhanced face feature vector using an improved ArcFace loss function, and outputting a face recognition result;

[0012] S7, calculating a final recognition score based on face detection confidence and feature matching score, and outputting a final identity recognition result and confidence through a fusion scoring mechanism.

[0013] Optionally, the S1 includes the following steps:

[0014] S11, synchronously collecting RGB images and infrared images through an image collection module configured by an outdoor electronic instrument;

[0015] S12, performing histogram equalization processing on each channel of the RGB image and the infrared image respectively, and fusing the processed three-channel image and the single-channel image into an enhanced RGB image;

[0016] S13, performing illumination estimation based on the enhanced RGB image to extract an illumination feature vector, and performing normalization processing on the illumination feature vector.

[0017] Optionally, the S2 includes the following steps:

[0018] S21, the face detection model takes the illumination feature vector as input, and obtains a feature map through an initial convolution downsampling layer;

[0019] S22, input the feature map into a multi-stage improved MobileViT block, the MobileViT block includes the following modules:

[0020] a lightweight depth separable convolution module for extracting local spatial features , ;

[0021] a local residual enhancement module for performing cross-layer feature superposition and channel attention modulation based on the local spatial features , and outputting enhanced features​ The specific calculation process is as follows:

[0022] First, the local spatial features are globally averaged pooled, and channel weights are generated through two fully connected layers , that is:

[0023] ;

[0024] wherein, is a weight matrix, is the image height, is the image width, is a ReLU activation function, is a Sigmoid function;

[0025] Then, the local spatial features are weighted and fused by applying channel attention to obtain enhanced features , that is:

[0026] ;

[0027] By adopting the local residual enhancement module, the robustness of the features under the conditions of occlusion or complex background is enhanced;

[0028] The face region explicit guidance module performs pixel-by-pixel weighting on the enhanced features according to a predefined or pre-trained face key region heat map, and outputs a region-guided feature map ;

[0029] The region-guided feature map is divided into local blocks, and each local block is expanded into a vector , which is input into a low-rank attention Transformer module to model attention, and an enhanced representation is output ;

[0030] All enhanced representations are reconstructed into an enhanced feature map ;

[0031] An illumination modulation function is adopted, and an illumination-modulated feature map is generated based on an illumination intensity matrix , that is:

[0032] ;

[0033] wherein, is an illumination modulation intensity control coefficient, is an illumination intensity matrix, representing the local illumination intensity of an input image at a pixel position , a full image pixel mean of the illumination intensity matrix ,a standard deviation of the illumination intensity matrix , compared with the original MobileViT, the face detection model using the illumination modulation function and embedding the Transformer can automatically adjust the feature expression mode in extreme environments such as strong light and backlight, and significantly improve the robustness and adaptability of actual outdoor detection;

[0034] S23, performing multi-scale regression and classification processing on the feature map after illumination modulation to obtain a face bounding box set , , and outputting a corresponding face key point set based on an independent key point regression head ;

[0035] S24, using a confidence threshold to filter and simplify the candidate detection results, and outputting a face detection result set .

[0036] Optionally, the S3 includes the following steps:

[0037] S31, based on the face detection result set , taking the centers of the two eyes in the key point coordinates as the affine transformation reference points, defining a standard reference coordinate set , and using the standard reference coordinate set to construct a face standard geometric model;

[0038] S32, calculating an affine transformation matrix , the affine transformation matrix satisfies:

[0039] ;

[0040] wherein, is the jth actual key point coordinate, is the standard key point coordinate corresponding to the jth actual key point coordinate;

[0041] based on the affine matrix , performing geometric transformation on the face region in the original image to generate an aligned face image ;

[0042] S33, uniformly scaling the generated face alignment image to a preset input size to obtain a standardized alignment image .

[0043] Optionally, the S4 includes the following steps:

[0044] S41, standardizing and aligning the image to a face recognition model constructed based on a ResNet100 structure, the face recognition model including a backbone network, a double-channel spatial attention module, and an output embedding head module;

[0045] S42, the backbone network of the face recognition model performs multi-scale hierarchical feature extraction on the input standardized and aligned image through multiple residual blocks to obtain a deep feature tensor wherein, , is the spatial size of the feature map, is the number of channels;

[0046] S43, the deep feature tensor is input to the double-channel spatial attention module to perform two attention mechanism calculations:

[0047] The first channel is significant region extraction based on gradient response, and an edge-guided attention map is calculated ;

[0048] The second channel is region weighting based on a face region heat map, and a structure-sensitive attention map is generated ;

[0049] S44, the edge-guided attention map and the structure-sensitive attention map are fused into a final spatial attention map :

[0050] ;

[0051] S45, a weighted operation is performed on the deep feature tensor to obtain an enhanced feature tensor :

[0052] ;

[0053] S46, the enhanced feature tensor is input to a global average pooling layer and an embedding mapping layer to obtain a face feature vector wherein, is the feature dimension.

[0054] Optionally, the S5 includes the following steps:

[0055] S51, the face feature vector and the standardized and aligned image are input to a pose estimation module;

[0056] In the pose estimation module, a multi-branch regression network is used to perform convolution processing on the image to output a three-dimensional face pose angle vector;

[0057] S52. Normalize the face pose angle vector to obtain the normalized pose vector. ;

[0058] S53. Normalize the attitude vector The input is fed into the pose-guided feature enhancement module to generate a weight vector. ;

[0059] S54, Facial Feature Vectors With weight vector Element-level fusion is performed to obtain pose-enhanced facial feature vectors. .

[0060] Optionally, S6 includes the following steps:

[0061] S61. Facial feature vectors with enhanced pose Input is fed into the recognition and classification module with an improved ArcFace loss function;

[0062] S62, Represent the set of facial feature vectors of all registered users as follows: ,in, For the first The unit-normalized class center vectors of each identity satisfy:

[0063] ;

[0064] S63. Calculate the input facial feature vector. Cosine similarity with all category centers:

[0065] ;

[0066] in, Indicates the input features and the first The angle between the class center vectors;

[0067] S64, Angle corresponding to the real category label Add angle interval This yields an improved ArcFace discriminant boundary. ,in, The set angular interval constant;

[0068] S65. Scale all similarities to obtain a normalized classification score. :

[0069] ;

[0070] in, This is the scaling factor. Indexing real classes;

[0071] S66, normalizing the classification scores of all classes Inputting a Softmax function to calculate the prediction probability;

[0072] S67, outputting the predicted identity class according to the maximum probability classification criterion

[0073]

[0074] wherein, represents the probability that the identity class is the first identity when the face feature vector is .

[0075] Optionally, the S7 comprises the following steps:

[0076] S71, taking the face detection confidence as a first score factor and taking the feature matching score as a second score factor;

[0077] S72, constructing a final fusion score function

[0078]

[0079] wherein, , represents the weight coefficient of each factor, is the face detection confidence, is the feature matching score;

[0080] S73, setting an identification confidence threshold When , it is confirmed that the output identification result is the predicted identity class , otherwise, outputting a "cannot be identified" flag.

[0081] According to an embodiment of the present application, a face recognition system suitable for outdoor electronic instruments comprises the following modules:

[0082] An image acquisition module is configured to synchronously acquire RGB images and infrared images, perform image channel separation, time alignment and size normalization, and generate a standard image data pair.

[0083] A face detection module is configured to input the enhanced image into a face detection model with an improved MobileViT structure, generate a bounding box and a key point set through local attention enhancement and illumination modulation mechanism.

[0084] A face alignment module is configured to calculate an affine transformation matrix according to face key points, and perform geometric alignment and standardization processing on the face region in the image. ​​​​​

[0085] a feature extraction module for inputting the aligned face image into a ResNet100 network and combining a double-channel spatial attention mechanism to extract a normalized face feature vector;

[0086] a pose enhancement module for estimating a face three-dimensional pose angle, constructing a pose guide weight vector, and re-weighting and fusing the face feature vector to enhance multi-pose robustness;

[0087] an identity classification module for performing normalized matching and identity prediction on the pose-enhanced feature based on an improved ArcFace loss function, and outputting a classification score and an identity label;

[0088] a score fusion module for comprehensively fusing a face detection confidence, a matching similarity and a light adaptation score, calculating a final recognition score, and determining recognition effectiveness according to a threshold value;

[0089] a result output module for outputting a final recognition identity and a confidence score.

[0090] The present application has the following advantages:

[0091] 1. The present application aims at the defect of unstable recognition performance of the prior art in a complex environment, and introduces a light self-adaptive MobileViT improved detection network and an ArcFace recognition model with a fusion pose guide mechanism, so as to realize accurate detection and robust recognition of a face under strong light, backlight, low illumination and large angle deflection conditions.

[0092] 2. The present application uses image light estimation and feature re-weighting means to effectively improve the positioning accuracy of the detection module under non-ideal light, and through the pose guide feature fusion structure, different angle face features are aligned and enhanced, which significantly improves the recognition accuracy and consistency.

[0093] 3. The present application adopts a lightweight network structure design, which takes into account the resource limitation and response speed requirement of embedded devices, improves the processing efficiency under the premise of ensuring the recognition accuracy, is suitable for practical deployment in outdoor terminal devices, and through a confidence score fusion mechanism, enhances the credibility evaluation ability of the recognition result, further reduces the false recognition rate and the missed recognition rate, and the overall system is superior to the prior art in environmental adaptability, recognition robustness and computing efficiency, and has good engineering feasibility and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0094] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0095] Figure 1 is a flowchart of the face recognition method in the present application;

[0096] Figure 2 The principle block diagram for improving the MobileViT structure in the application;

[0097] Figure 3 The principle block diagram for the face recognition system in the application. DETAILED DESCRIPTION

[0098] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which show the basic structure of the application in a schematic manner only, and thus they show only those components which are relevant to the application.

[0099] Example 1

[0100] The current outdoor face recognition technology generally faces the problems of unstable detection, low recognition accuracy and poor adaptability to strong light, backlight, occlusion and the like in complex environments, especially when deployed on mobile terminals or low-power devices, the traditional model is difficult to balance accuracy and efficiency, and the recognition effect is severely limited. Based on this, the embodiment proposes a face recognition method suitable for outdoor electronic instruments, which combines the illumination adaptive improved MobileViT detection network and the pose guided face recognition algorithm, and shows higher robustness and recognition accuracy in handling strong light interference, multi-pose deflection and edge occlusion and the like. Compared with the existing recognition system which relies on single feature extraction or fixed illumination model, the application not only improves the multi-environment adaptability, but also has the advantages of lightweight structure and edge deployment, and is suitable for various outdoor practical scenes such as security patrol, intelligent terminal and emergency response. As shown in Figure 1 and Figure 2 The method comprises the following steps:

[0101] S1, collecting image data in an outdoor environment, the image data comprising an RGB image and an infrared image, and pre-processing the original input image to obtain an illumination feature vector. The following steps are included:

[0102] S11, synchronously collecting the RGB image and the infrared image through an image collection module configured by the outdoor electronic instrument;

[0103] S12, performing histogram equalization processing on each channel of the RGB image and the infrared image respectively, and fusing the processed three-channel image and single-channel image into an enhanced RGB image;

[0104] S13, performing illumination estimation based on the enhanced RGB image to extract an illumination feature vector, and performing normalization processing on the illumination feature vector.

[0105] ​The improved MobileViT structure is adopted for feature extraction and detection, and a lightness self-adaptive modulation mechanism is combined, so that the system has the ability to stably detect the face region and accurately extract the key points of the five organs under strong light interference or complex background, and the adaptability of the detection module to complex scenes is enhanced.

[0106] S2, using a face detection model based on an improved MobileViT structure, performing face region detection and face key point positioning on the lightness enhanced image, and outputting a face detection result set containing a face boundary box and five organ key point coordinates.

[0107] S21, the face detection model takes the lightness feature vector as input, and obtains a feature map through an initial convolution downsampling layer.

[0108] S22, input the feature map to a multi-stage improved MobileViT block, refer to Figure 2 , the MobileViT block includes the following modules:

[0109] a lightweight depth separable convolution module for extracting local spatial features , ;

[0110] a local residual enhancement module based on local spatial features performs cross-layer feature superposition and channel attention modulation, and outputs enhanced features , the specific calculation process is as follows:

[0111] First, the local spatial features are globally averaged pooled, and the channel weights are generated through two fully connected layers , that is:

[0112] ;

[0113] wherein is a weight matrix, is the height of the image, is the width of the image, is a ReLU activation function, is a Sigmoid function;

[0114] Then apply channel attention to weight and fuse the local spatial features :

[0115] ;

[0116] By employing a local residual enhancement module, the robustness of feature representation under occlusion or complex background conditions is enhanced.

[0117] The explicit facial region guidance module enhances features based on predefined or pre-trained facial key region heatmaps. Perform pixel-by-pixel weighting to output region-guided feature maps. ;

[0118] Region-guided feature map Divide into local blocks, and flatten each local block into a vector. The input is a low-rank attention Transformer module for attention modeling, and the output is an enhanced representation. ;

[0119] All enhanced representations Reconstructed into enhanced feature maps ;

[0120] An illumination modulation function is used to generate a feature map after illumination modulation based on the illumination intensity matrix. ,Right now:

[0121] ;

[0122] in, This is the light modulation intensity control coefficient. This is the illumination intensity matrix, representing the pixel positions of the input image. Local light intensity at the location, Represents the light intensity matrix The average pixel value of the entire image. Represents the light intensity matrix The standard deviation of the face detection model, compared to the original MobileViT's lack of awareness of lighting changes, is significantly improved by adopting a lighting modulation function and embedding a Transformer. This allows the face detection model to automatically adjust its feature representation in extreme environments such as strong light and backlight, thus significantly improving the robustness and adaptability of actual outdoor detection.

[0123] S23. Feature map after illumination modulation Multi-scale regression and classification are performed to obtain a set of face bounding boxes. , It outputs the corresponding set of facial key points based on an independent keypoint regression head. ;

[0124] S24. The candidate detection results are simplified by using confidence threshold filtering, and the face detection result set is output. .

[0125] The application realizes geometric alignment based on key point affine transformation, can effectively eliminate the angle and posture difference of the face generated during shooting, improve the accuracy and consistency of subsequent feature extraction and recognition, and lays a standardized input foundation for deep feature learning.

[0126] S3, performing affine transformation based on the face detection result set to perform face alignment, and obtaining a standardized aligned face image. Comprising the following steps:

[0127] S31, based on the face detection result set , the center points of the two eyes in the key point coordinates are taken as the affine transformation reference points, and a standard reference coordinate set is defined to construct a face standard geometric model;

[0128] S32, calculating an affine transformation matrix , the affine transformation matrix satisfies:

[0129] ;

[0130] Wherein, is the jth actual key point coordinate, is the standard key point coordinate corresponding to the jth actual key point coordinate;

[0131] Based on the affine matrix , the face region in the original image is geometrically transformed to generate an aligned face image ;

[0132] S33, the generated face alignment image is uniformly scaled to a preset input size to obtain a standardized alignment image .

[0133] The application adopts a double-channel spatial attention mechanism to weight the features of the face region, effectively strengthens the expression ability of the key region (such as eyes, nose bridge and mouth), and improves the feature discrimination and robustness under complex background and noise interference.

[0134] S4, inputting the standardized aligned face image into a face recognition model based on the ResNet100 structure, the face recognition model embedding a double-channel spatial attention mechanism to extract an attention map, obtaining a weighting coefficient by fusing the attention map, and outputting a face feature vector. Comprising the following steps:

[0135] S41, inputting the standardized aligned image into a face recognition model constructed based on the ResNet100 structure, the face recognition model comprising a backbone network, a double-channel spatial attention module and an output embedding head module;

[0136] S42, the backbone network of the face recognition model extracts multi-scale hierarchical features from the input standardized alignment image through multiple residual blocks to obtain a deep feature tensor , wherein, , is the spatial size of the feature map, is the number of channels;

[0137] S43, input the deep feature tensor to a dual-channel spatial attention module to perform two attention mechanism calculations:

[0138] The first channel is significant area extraction based on gradient response, and an edge-guided attention map is calculated ;

[0139] The second channel is region weighting based on a face region heat map, and a structure-sensitive attention map is generated ;

[0140] S44, the edge-guided attention map and the structure-sensitive attention map are fused into a final spatial attention map :

[0141] ;

[0142] S45, the deep feature tensor is subjected to a weighting operation to obtain an enhanced feature tensor :

[0143] ;

[0144] S46, input the enhanced feature tensor to a global average pooling layer and an embedding mapping layer to obtain a face feature vector , wherein, is the feature dimension.

[0145] The present application can compensate for feature distortion caused by large-angle deflection or side face by estimating head pose and constructing a pose-guided feature fusion method, enhance the recognition ability of the system under multi-angle and multi-view conditions, and reduce the misrecognition rate caused by pose deviation.

[0146] S5, the face feature vector is subjected to pose-guided feature reconstruction processing, head angle information is extracted through a pose estimation module, and is used for re-weighted fusion of the face feature vector to obtain a pose-enhanced face feature vector. Including the following steps:

[0147] S51, the face feature vector and the standardized alignment image Input to the attitude estimation module;

[0148] In the pose estimation module, a multi-branch regression network is used to convolve the image and output a three-dimensional face pose angle vector.

[0149] S52. Normalize the face pose angle vector to obtain the normalized pose vector. ;

[0150] S53. Normalize the attitude vector The input is fed into the pose-guided feature enhancement module to generate a weight vector. ;

[0151] S54, Facial Feature Vectors With weight vector Element-level fusion is performed to obtain pose-enhanced facial feature vectors. .

[0152] This invention utilizes an improved ArcFace loss function to enhance the ability to distinguish inter-class boundaries and achieves high-precision identity classification based on feature normalization, enabling face recognition to maintain high stability and scalability in complex environments and improving overall recognition accuracy.

[0153] S6. The improved ArcFace loss function is used to normalize the features of the pose-enhanced face feature vector and perform identity classification, outputting the face recognition result. This includes the following steps:

[0154] S61. Facial feature vectors with enhanced pose Input is fed into the recognition and classification module with an improved ArcFace loss function;

[0155] S62, Represent the set of facial feature vectors of all registered users as follows: ,in, For the first The unit-normalized class center vectors of each identity satisfy:

[0156] ;

[0157] S63. Calculate the input facial feature vector. Cosine similarity with all category centers:

[0158] ;

[0159] in, Indicates the input features and the first The angle between the class center vectors;

[0160] S64, Angle corresponding to the real category label add angle interval , get improved ArcFace discriminant boundary wherein, is a set angle interval constant;

[0161] S65, scale all similarities to get normalized classification scores :

[0162] ;

[0163] wherein, is a scaling coefficient, is a real class index;

[0164] S66, input normalized classification scores of all classes into a Softmax function to calculate prediction probabilities;

[0165] S67, output a predicted identity class according to a maximum probability classification criterion :

[0166] ;

[0167] wherein, represents a probability that the identity class is the i-th identity when the face feature vector is .

[0168] The present application constructs a multi-factor fusion score model by comprehensively detecting confidence and matching similarity, realizes dynamic evaluation and reasonable output of the recognition result reliability, significantly reduces the misjudgment rate, and improves the reliability and security of the system in actual application.

[0169] S7, calculate a final recognition score based on face detection confidence and feature matching score, and output a final identity recognition result and confidence through a fusion score mechanism. Including the following steps:

[0170] S71, take face detection confidence as a first score factor, and take feature matching score as a second score factor;

[0171] S72, construct a final fusion score function :

[0172] ;

[0173] wherein, , indicates a weight coefficient of each factor, is face detection confidence, is feature matching score;

[0174] ​​S73, set an identification credibility threshold When , confirm that the output identification result is the predicted identity category , otherwise, output the "unable to identify" flag.

[0175] Embodiment Two

[0176] On the basis of the face recognition method proposed in Embodiment One, the face recognition system suitable for outdoor electronic equipment for implementing the face recognition method is provided, which is composed of an image acquisition module, a face detection module, a face alignment module, a feature extraction module, a pose enhancement module, an identity classification module, a score fusion module, and a result output module.

[0177] The image acquisition module is used for synchronously acquiring RGB images and infrared images, performing image channel separation, time alignment, and size normalization, and generating a standard image data pair; the face detection module is used for inputting the enhanced image into the face detection model with the improved MobileViT structure, generating a bounding box and a key point set through local attention enhancement and illumination modulation mechanism; and the face alignment module is used for calculating an affine transformation matrix according to the face key points, and performing geometric alignment and standardization processing on the face region in the image.

[0178] The feature extraction module is used for inputting the aligned face image into the ResNet100 network and combining the dual-channel spatial attention mechanism to extract a normalized face feature vector; the pose enhancement module is used for estimating the three-dimensional pose angle of the face, constructing a pose guide weight vector, and re-weighting and fusing the face feature vector to enhance the multi-pose robustness; the identity classification module is used for performing normalized matching and identity prediction on the pose-enhanced feature based on the improved ArcFace loss function, outputting a classification score and an identity label; the score fusion module is used for comprehensively fusing the face detection confidence, the matching similarity, and the illumination adaptability score, calculating a final recognition score, and determining the recognition effectiveness according to the threshold; and the result output module is used for outputting the final recognition identity and the confidence score.

[0179] The face recognition system proposed in this embodiment adopts a lightweight network structure design, taking into account the resource constraints and response speed requirements of embedded devices, improving the processing efficiency on the premise of ensuring the recognition accuracy, and being suitable for actual deployment in outdoor terminal devices.

[0180] Embodiment Three

[0181] In the construction project of an intelligent patrol system in a coastal city, the research and development team deployed a batch of outdoor electronic instrument terminals with face recognition function, which are used to quickly identify target personnel in the mobile crowd in outdoor mobile scenarios. These outdoor electronic instruments are installed on the top of electric patrol vehicles or related equipment, and have the requirements of low power consumption, high real-time performance, all-weather operation, etc. Typical scenarios include direct sunlight, dense crowds, night duty, and harsh weather (such as fog, high humidity) environments.

[0182] Traditional face recognition methods have significant problems in such application scenarios. For example, in the case of strong sunlight at noon, the face key area information is missing due to overexposure of the image, resulting in a significant decrease in recognition accuracy. In the night or backlit environment, the detection module cannot accurately identify the face area due to insufficient image brightness or light interference. In addition, in the process of dynamic shooting, the face angle changes frequently, and the problems of side face and large pitch angle further exacerbate the difficulty of recognition. Preliminary testing of the project shows that the existing system's accuracy decreases to below 72% when the face deflection angle exceeds 25°, and the false positive rate increases to 18%.

[0183] To solve the above problems, the project team introduced the method proposed in the present application, integrated the improved MobileViT detection model and ArcFace+PGFE recognition module into the terminal system. The system first uses dual cameras to collect RGB and infrared images in parallel, and realizes image preprocessing through time synchronization and illumination estimation to ensure balanced image quality under different lighting conditions. Then, the light-adaptive face detection model is used to realize accurate face positioning, and the face is standardized and aligned according to the key point coordinates. The extracted features are strengthened by a dual-channel spatial attention mechanism, and then input into the pose-guided module for multi-view reconstruction, effectively enhancing the recognition ability of large-angle deflection faces.

[0184] In the specific test, the team selected a certain area in a city for continuous 7-day field deployment test, with pre-arranged target personnel mixed in the crowd. A total of about 86,000 sample images were collected, covering all-day periods and various weather conditions. The system compared and evaluated the results with the traditional CNN-based system by interfacing with the pre-established database.

[0185] The results show that in the daytime strong light scene, the recognition accuracy of the system is improved from 81.3% of the comparative system to 94.6%, and in the night low light condition, it is improved from 68.7% to 89.2%. Especially in the case of face deflection angle greater than 30°, the recognition accuracy remains above 87.5%, which is more than 15 percentage points better than the traditional system. At the same time, the average recognition time is controlled within 176ms, meeting the real-time response requirements of mobile terminals.

[0186] Through the fusion scoring mechanism, the system also effectively reduces the false recognition rate and the repeated alarm problem. In the high human flow density area (such as the subway entrance and the shopping area entrance), the false recognition rate is reduced from 9.3% of the original system to 3.2%, and the accurate, safe and efficient outdoor face recognition target is realized.

[0187] Table 1. Recognition performance data under different illumination and posture conditions

[0188]

[0189] The above examples and measured data show that the face recognition method has good outdoor adaptability, recognition accuracy and response speed, can effectively solve the problem of unstable recognition performance of the prior art in complex scenes such as strong light, low light and posture deflection, and has significant practicability and popularization value.

[0190] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A face recognition method for outdoor electronic devices, characterized by, The method comprises the following steps: S1, collecting image data in an outdoor environment, the image data comprising an RGB image and an infrared image, and pre-processing the original input image to obtain an illumination feature vector; S2, using a face detection model based on an improved MobileViT structure to detect a face region and locate a face key point of the illumination enhanced image, and outputting a face detection result set comprising a face boundary box and five facial feature point coordinates, specifically comprising the following steps: S21, the face detection model takes the illumination feature vector as input As input, the feature map is obtained through the initial convolution downsampling layer ; S22, the feature map is input to a multi-stage improved MobileViT block, the MobileViT block comprising the following modules: Lightweight depthwise separable convolution module, extracting local spatial features , ; local residual enhancement module based on local spatial features performing cross-layer feature concatenation and channel attention modulation to output enhanced features ; The face region explicit guidance module performs pixel-by-pixel weighting on the enhanced feature according to a predefined or pre-trained face key region heat map to output a region guidance feature map The face region explicit guidance module performs pixel-by-pixel weighting on the enhanced feature according to a predefined or pre-trained face key region heat map to output a region guidance feature map ; Area guide feature map Split into local blocks, each local block is expanded into a vector , input low-rank attention Transformer module for attention modeling, output enhanced representation ; all the enhanced representations reconstructed into an enhanced feature map ; The illumination modulation function is adopted, and the feature map after illumination modulation is generated based on an illumination intensity matrix ; S23, the feature map after illumination modulation Multi-scale regression and classification processing is performed to obtain a face bounding box set , , and an independent key point regression head is used to output a corresponding face key point set ; S24, adopt the confidence threshold to screen the candidate detection result, output the face detection result set ; S3, performing face alignment based on the face detection result set through affine transformation to obtain a standardized aligned face image; S4, inputting the standardized aligned face image into a face recognition model based on a ResNet100 structure, the face recognition model embedding a double-channel spatial attention mechanism to extract an attention map, obtaining a weighting coefficient by fusing the attention map, and outputting a face feature vector; S5, performing pose-guided feature reconstruction processing on the face feature vector, extracting head angle information through a pose estimation module, and using the head angle information to perform re-weighted fusion on the face feature vector to obtain a pose-enhanced face feature vector; S6, performing feature normalization and identity classification on the pose-enhanced face feature vector using an improved ArcFace loss function, and outputting a face recognition result; S7, calculating a final recognition score based on a face detection confidence and a feature matching score, and outputting a final identity recognition result and a confidence through a fusion scoring mechanism.

2. The face recognition method of claim 1, wherein, The S1 comprises the following steps: S11, synchronously collecting an RGB image and an infrared image through an image acquisition module of an outdoor electronic instrument; S12, performing histogram equalization processing on each channel of the RGB image and the infrared image respectively, and fusing the processed three-channel image and the single-channel image into an enhanced RGB image; S13, performing light estimation based on the enhanced RGB image to extract a light feature vector, and performing normalization processing on the light feature vector performed.

3. The face recognition method of claim 1, wherein, The generated illumination-modulated feature map , the expression is: ； wherein, is a light modulation intensity control coefficient, is a light intensity matrix representing the local light intensity of the input image at pixel position , denotes the global pixel mean of the light intensity matrix , denotes the standard deviation of the light intensity matrix .

4. The face recognition method of claim 1, wherein, The S3 comprises the following steps: S31、based on the face detection result set , the two eye center points in the key point coordinates are taken as the affine transformation reference points, and a standard reference coordinate set is defined for constructing a face standard geometric model; S32, calculate the affine transformation matrix , based on the affine matrix performing geometric transformation on the face region in the original image to generate an aligned face image ; S33, generating a face alignment image uniformly scaling to a preset input size to obtain a standardized alignment image .

5. The face recognition method of claim 1, wherein, The S4 comprises the following steps: S41, standardizing the aligned image input into a face recognition model constructed based on a ResNet100 structure, the face recognition model including a backbone network, a double-channel spatial attention module, and an output embedding head module; S42, the backbone network of the face recognition model extracts multi-scale hierarchical features from the input standardized aligned image through multiple residual blocks to obtain a deep feature tensor wherein, , is a spatial size of the feature map, is a number of channels; S43, the depth feature tensor is input to the dual-channel spatial attention module, which performs two attention mechanism calculations: The first channel is a gradient response based salient region extraction, which computes an edge guided attention map ; The second channel is a region weighting based on a face region heat map to generate a structure-sensitive attention map ; S44, fuse the edge-guided attention map and the structure-sensitive attention map into a final spatial attention map ; S45, to the depth feature tensor performing a weighting operation to obtain an enhanced feature tensor ; S46, the enhanced feature tensor Input into the global average pooling layer and the embedding mapping layer to obtain a face feature vector wherein, is the feature dimension.

6. The face recognition method of claim 1, wherein, The S5 comprises the following steps: S51, the face feature vector and the normalized aligned image is input to the pose estimation module; In the pose estimation module, a multi-branch regression network is used to perform convolution processing on the image to output a three-dimensional face pose angle vector; S52, normalizing the face pose angle vector to obtain a normalized pose vector ; S53, normalizing the pose vector input to the pose-guided feature enhancement module to generate a weight vector ; S54, the face feature vector with the weight vector element-level fusion is performed to obtain the pose-enhanced face feature vector .

7. The face recognition method of claim 1, wherein, The S6 comprises the following steps: S61, the posture enhanced face feature vector is input to the recognition classification module with the improved ArcFace loss function; S62, Represent the set of facial feature vectors of all registered users as follows: ,in, For the first Unit-normalized category center vector for each identity; S63, calculate the face feature vector of the input person cosine similarity with all class centers; S64, the angle corresponding to the real class label adding angle intervals to obtain an improved ArcFace discrimination boundary wherein, is a set angle interval constant; S65, scaling all similarities to obtain normalized classification scores ; S66, normalizing the classification scores of all classes an input softmax function to compute the prediction probabilities; S67, output the predicted identity class according to the maximum probability classification criterion .

8. The face recognition method of claim 1, wherein, The S7 comprises the following steps: S71, taking the face detection confidence as a first scoring factor and taking the feature matching score as a second scoring factor; S72, constructing a final fusion score function ; S73, set a recognition confidence threshold When , confirm that the output recognition result is the predicted identity class , otherwise, output an "unrecognizable" flag.

9. A face recognition system suitable for outdoor electronic equipment, which executes the face recognition method according to any one of claims 1 to 8, characterized in that, The method comprises the following modules: An image acquisition module is configured to synchronously collect an RGB image and an infrared image, perform image channel separation, time alignment and size normalization, and generate a standard image data pair; A face detection module is configured to input the enhanced image into a face detection model based on an improved MobileViT structure, generate a boundary box and a key point set through local attention enhancement and illumination modulation mechanism; A face alignment module is configured to calculate an affine transformation matrix according to the face key points, and perform geometric alignment and standardization processing on the face region in the image; A feature extraction module is configured to input the aligned face image into a ResNet100 network and combine a double-channel spatial attention mechanism to extract a normalized face feature vector; A pose enhancement module is configured to estimate a three-dimensional face pose angle, construct a pose-guided weight vector, and perform re-weighted fusion on the face feature vector to enhance multi-pose robustness. An identity classification module is configured to normalize and match the posture enhanced features based on an improved ArcFace loss function, and output classification scores and identity labels; A score fusion module is configured to comprehensively fuse face detection confidence, matching similarity and illumination adaptability scores, calculate a final recognition score, and determine recognition effectiveness according to a threshold value; A result output module is configured to output a final recognized identity and a confidence score.

Images