Face recognition method and system based on image analysis

Abstract

This application provides a face recognition method and system based on image analysis, relating to the field of image analysis technology. It determines inter-frame difference vectors by using the difference information between adjacent frames in consecutive image frames during face recognition. Pose fusion is then performed on the consecutive image frames using all inter-frame difference vectors to obtain pose fusion features. These pose fusion features are mapped to a facial contour recognition space based on the spatial difference angle during face recognition. The pose fusion features in the facial contour recognition space are then deconvolved, and the local contour features obtained from the deconvolution determine the pose mapping cost during face recognition. Furthermore, recognition compensation is performed on the face recognition process based on ambient light intensity and the mapping cost to obtain compensation feature quantities. The matching confidence of face recognition is then determined using the compensation feature quantities, and the face recognition results are filtered based on the matching confidence. This application can achieve dynamic feature adjustment and light compensation during the face recognition process to improve the accuracy of face recognition.

Description

Technical Field

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

[0002] Image analysis is a technology based on computer vision and digital image processing, which aims to automatically extract, identify and analyze key information in images through algorithms. The core technologies of image analysis include edge detection, feature extraction, pattern recognition, deep learning, etc., and it can be used in many fields such as object recognition, target detection, and scene understanding. Among them, image analysis technology plays a crucial role in the field of face recognition.

[0003] Facial recognition is an identity verification method based on biometric identification technology. It determines an individual's identity by analyzing and comparing facial features. Its main goal is to enable computers to recognize and verify faces like humans. Facial recognition technology is widely used in security monitoring, access control systems, and other fields due to its non-contact nature and high accuracy. Traditional facial recognition methods rely on key point detection and template matching. However, real-time changes in facial pose can lead to the loss of facial pose features, and varying ambient lighting intensities can cause uneven facial brightness, making it difficult for existing facial matching algorithms to effectively match facial features. Therefore, how to achieve dynamic feature adjustment and light compensation during the facial recognition process to improve accuracy remains a challenge for the industry. Summary of the Invention

[0004] This application provides a face recognition method and system based on image analysis, which can realize dynamic feature adjustment and light compensation in the face recognition process to improve the accuracy of face recognition.

[0005] In a first aspect, this application provides a face recognition method based on image analysis, the face recognition method comprising:

[0006] Capture continuous image frames during face recognition;

[0007] Based on the difference information between adjacent frames in the continuous image frames, the inter-frame difference vector is determined, and the pose fusion features are obtained by performing pose fusion on the continuous image frames using all the inter-frame difference vectors.

[0008] Based on the spatial difference angle during face recognition, the pose fusion features are mapped to the facial contour recognition space. The pose fusion features in the facial contour recognition space are deconvolved, and the pose mapping cost during face recognition is determined by the local contour features obtained from the deconvolution.

[0009] The ambient light intensity during face recognition is obtained, and the face recognition process is compensated based on the ambient light intensity and the mapping cost to obtain the compensation feature quantity.

[0010] The matching confidence of face recognition is determined by the compensation feature quantity, and then the face recognition results are filtered based on the matching confidence.

[0011] In this embodiment, a depth camera is used to capture continuous image frames during face recognition.

[0012] In this embodiment, determining the inter-frame difference vector based on the difference information between adjacent frames in the continuous image frames specifically includes:

[0013] Extract key feature maps from each of the consecutive image frames;

[0014] The difference images of adjacent frames in the continuous image frames are determined, and then the inter-frame difference vector of the adjacent frames is determined by the difference images of the adjacent frames and all key feature maps.

[0015] In this embodiment, the pose fusion feature obtained by performing pose fusion on the consecutive image frames using all inter-frame difference vectors specifically includes:

[0016] Construct a pose fusion matrix based on all inter-frame difference vectors;

[0017] The attitude fusion features are determined by the eigenvalues ​​of the attitude fusion matrix.

[0018] In this embodiment, mapping the pose fusion features to the facial contour recognition space based on the spatial difference angle during face recognition specifically includes:

[0019] The mapping angle is determined based on the spatial difference angle;

[0020] Based on the mapping angle, the pose fusion features are linearly mapped to obtain a facial contour recognition space.

[0021] In this embodiment, the pose mapping cost for determining face recognition from the local contour features obtained by deconvolution specifically includes:

[0022] A facial pose fusion map is determined based on the pose fusion features in the facial contour recognition space;

[0023] The pose loss vector for face recognition is determined using the facial pose fusion map and all local contour features.

[0024] The pose mapping cost during face recognition is determined by the convolution scale of the deconvolution performed on the pose loss vector during face recognition and the pose fusion features in the facial contour recognition space.

[0025] In this embodiment, a preset neural network model is used to deconvolve the pose fusion features in the facial contour recognition space.

[0026] In this embodiment, the ambient light intensity during face recognition is obtained through the light sensor of the depth camera.

[0027] In this embodiment, the face recognition process is compensated based on the ambient light intensity and the mapping cost to obtain the compensated feature quantity, which specifically includes:

[0028] The feature correction amount in the face recognition process is determined based on the mapping cost;

[0029] The illumination compensation coefficient is determined based on the ambient light intensity.

[0030] Obtain the brightness reference value of the facial contour recognition space;

[0031] Based on the feature correction amount, the illumination compensation coefficient, and the brightness reference value, the face recognition process is subjected to feature adjustment and illumination compensation, thereby obtaining the compensation feature amount.

[0032] Secondly, this application provides an image analysis-based face recognition system for performing an image analysis-based face recognition method, the face recognition system comprising:

[0033] The image acquisition module is used to acquire continuous image frames during face recognition.

[0034] The pose fusion module is used to determine the inter-frame difference vector based on the difference information between adjacent frames in the continuous image frames, and to perform pose fusion on the continuous image frames through all the inter-frame difference vectors to obtain pose fusion features;

[0035] The feature processing module is used to map the pose fusion features to the facial contour recognition space based on the spatial difference angle during face recognition, perform deconvolution on the pose fusion features in the facial contour recognition space, and then determine the pose mapping cost during face recognition from the local contour features obtained by deconvolution.

[0036] The illumination compensation module is used to obtain the ambient light intensity during face recognition, and to perform recognition compensation on the face recognition process based on the ambient light intensity and the mapping cost to obtain the compensation feature quantity.

[0037] The matching and filtering module is used to determine the matching confidence of face recognition through the compensation feature quantity, and then filter the face recognition results based on the matching confidence.

[0038] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0039] The process involves: acquiring continuous image frames during face recognition; determining inter-frame difference vectors based on the difference information between adjacent frames; performing pose fusion on the continuous image frames using all inter-frame difference vectors to obtain pose fusion features; mapping the pose fusion features to a facial contour recognition space based on the spatial difference angle during face recognition; performing deconvolution on the pose fusion features in the facial contour recognition space; determining the pose mapping cost during face recognition based on the local contour features obtained from the deconvolution; acquiring the ambient light intensity during face recognition; compensating for the face recognition process based on the ambient light intensity and the mapping cost to obtain compensation feature quantities; determining the matching confidence of face recognition based on the compensation feature quantities; and then filtering face recognition results based on the matching confidence.

[0040] Therefore, this application can realize dynamic feature adjustment and light compensation in the face recognition process. First, by acquiring continuous image frames, the dynamic features of the face can be captured, reducing recognition errors caused by instantaneous changes and providing more image feature information. Furthermore, the inter-frame difference vector reflects the changing trend of the face in a short period, facilitating dynamic correction of pose errors and reducing feature deviations caused by head rotation or tilt, thus making feature extraction more accurate and improving face matching accuracy under different poses. Second, by adjusting the pose fusion features through spatial difference angles, it maps them to a standardized facial contour recognition space, reducing the impact of individual differences on the facial features. The effects of differences or changes in shooting angle can be mitigated by using deconvolution operations to restore local contour details, making facial features clearer and improving the accuracy of feature matching. Then, by acquiring ambient light information in real time, images under low or strong light conditions can be dynamically adjusted, and compensation can be made by combining pose mapping costs, which can effectively reduce the impact of lighting changes on feature extraction and improve the system's adaptability to different lighting conditions. Finally, by compensating for feature quantity, the final matching result of face recognition can be optimized, improving the system's adaptability to dynamic changes, reducing recognition errors, and improving the accuracy of face recognition.

[0041] In summary, the technical solution adopted in this application can realize dynamic feature adjustment and light compensation in the face recognition process, so as to improve the accuracy of face recognition. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1This is a flowchart of a face recognition method based on image analysis provided in this application;

[0044] Figure 2 This is an exemplary flowchart for determining pose fusion features according to the present application;

[0045] Figure 3 This is an exemplary flowchart of determining the attitude mapping cost according to the present application;

[0046] Figure 4 This is a modular structure diagram of the face recognition system provided in this application. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] This application provides a face recognition method and system based on image analysis. The core of the method involves acquiring continuous image frames during face recognition, determining inter-frame difference vectors based on the difference information between adjacent frames, and performing pose fusion on the continuous image frames using all inter-frame difference vectors to obtain pose fusion features. Next, the pose fusion features are mapped to a facial contour recognition space based on the spatial difference angle during face recognition. The pose fusion features in the facial contour recognition space are then deconvolved, and the local contour features obtained from the deconvolution determine the pose mapping cost during face recognition. Then, the ambient light intensity during face recognition is acquired, and the face recognition process is compensated based on the ambient light intensity and the mapping cost to obtain compensation feature quantities. Finally, the matching confidence of face recognition is determined using the compensation feature quantities, and the face recognition results are then filtered based on the matching confidence.

[0049] Example 1: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 1 As shown, this figure is an exemplary flowchart of a face recognition method based on image analysis according to this embodiment of the present application. The face recognition method includes the following steps:

[0050] In step S1, consecutive image frames are acquired during face recognition.

[0051] In specific implementation, a depth camera is used to capture continuous image frames during face recognition. The continuous image frames refer to the image frames of the face video captured by the depth camera within a specified time period. The specified time period can be set between 3 and 5 seconds. It should be noted that the continuous image frames are composed of multiple image frames in chronological order. The image frames are each image frame in the face video captured by the depth camera within the specified time period, and each image frame corresponds to the unique posture features of the person's face.

[0052] In step S2, the inter-frame difference vector is determined based on the difference information between adjacent frames in the continuous image frames, and the pose fusion is performed on the continuous image frames using all the inter-frame difference vectors to obtain pose fusion features.

[0053] In this embodiment, determining the inter-frame difference vector based on the difference information between adjacent frames in the continuous image frames specifically includes:

[0054] Extract key feature maps from each of the consecutive image frames;

[0055] The difference images of adjacent frames in the continuous image frames are determined, and then the inter-frame difference vector of the adjacent frames is determined by the difference images of the adjacent frames and all key feature maps.

[0056] It should be noted that the key feature maps in this application include: structural center map, contour boundary map, orientation feature map, and dynamic expression map. The key features include: nose, ears, eyes, and mouth. The structural center map is represented by the nose in the face, representing the geometric center of the face, pose estimation, and depth information. The ears in the face can be used as the contour boundary map for side face recognition and head contour matching. The eyes in the face can be used as the orientation feature map for gaze tracking and iris recognition. The mouth in the face can be used as the dynamic expression map for expression analysis, speech recognition, and lip reading detection. Extracting these key feature maps can improve the robustness of face recognition at different angles. It should also be noted that adjacent frames in this application refer to two adjacent image frames.

[0057] In specific implementation, firstly, key features in all image frames can be extracted by setting the size of the Haar-like input features (i.e., the size of the key features) in the Haar cascade detector. Then, all key features in each image frame are combined into a 400*400 pixel image according to the contour position of the face. The combined image is used as the key feature map, thus obtaining the key feature maps of all image frames. Secondly, a group of adjacent frames is selected, and the group of adjacent frames is convolved and fused using a convolutional neural network to obtain a difference image. The Euclidean distance between the two image frames in the group of adjacent frames and the difference image is calculated, and the rotation angle between the two image frames in the group of adjacent frames and the difference image is also calculated (i.e., the rotation angle between the two image frames in the group of adjacent frames and the difference image). The arctangent of the ratio of the difference in the vertical coordinate to the difference in the horizontal coordinate of the same pixel in the key feature map (the calculation process can be provided by the Pandas library). Then, the vector composed of the Euclidean distance and the rotation angle is used as the inter-frame difference vector of facial pose between adjacent frames, that is, inter-frame difference vector = (Euclidean distance, rotation angle). Repeat the above steps to obtain the inter-frame difference vector of the remaining adjacent frames. The inter-frame difference vector refers to the difference features between adjacent frames due to the change in the pose of the key facial parts of the person. By calculating the Euclidean distance and rotation angle between the key feature maps in all adjacent frames, the changes in facial pose between the key facial parts of the person in a specified time period can be analyzed.

[0058] Preferably, in this embodiment, reference Figure 2 As shown, this figure is an exemplary flowchart for determining pose fusion features in an embodiment of this application. In this embodiment, pose fusion is performed on the consecutive image frames using all inter-frame difference vectors to obtain pose fusion features. This can be achieved through the following steps:

[0059] First, in step S21, a pose fusion matrix is ​​constructed based on all inter-frame difference vectors;

[0060] Then, in step S22, the attitude fusion features are determined by the eigenvalues ​​of the attitude fusion matrix.

[0061] In practice, firstly, all inter-frame difference vectors are normalized, and then all normalized inter-frame difference vectors are arranged in a column according to the size of different key facial features from smallest to largest. The matrix formed by sorting and combining all inter-frame difference vectors is then used as the pose fusion matrix. The pose fusion matrix reflects the overall pose difference between key feature maps in all adjacent frames. Normalizing all inter-frame difference vectors can eliminate the dimensional differences between different components. Then, all eigenvalues ​​of the pose fusion matrix are obtained using the Pandas library, and the array formed by randomly combining all eigenvalues ​​is used as the pose fusion feature.

[0062] It should be noted that the pose fusion feature in this application refers to all feature values ​​of a person's face under multiple poses and angles. The pose fusion feature is obtained by fusing the inter-frame difference vectors generated by the same recognition target due to different pose changes. By determining the pose fusion feature, the dynamic changes of facial pose position and angle can be analyzed, providing strong data support for subsequent pose correction.

[0063] In step S3, the pose fusion features are mapped to the facial contour recognition space based on the spatial difference angle during face recognition. The pose fusion features in the facial contour recognition space are deconvolved, and the pose mapping cost during face recognition is determined by the local contour features obtained from the deconvolution.

[0064] In this embodiment, mapping the pose fusion features to the facial contour recognition space based on the spatial difference angle during face recognition specifically includes:

[0065] The mapping angle is determined based on the spatial difference angle;

[0066] Based on the mapping angle, the pose fusion features are linearly mapped to obtain a facial contour recognition space.

[0067] It should be noted that the spatial difference angle in this application refers to the horizontal angle between the position of the webcam and the position of the person's face. The spatial difference angle reflects the degree of directional offset when the webcam is shooting. By using the spatial difference angle when the webcam is shooting to perform contour mapping on the pose fusion features, the facial contour recognition space of the person can be obtained more accurately.

[0068] In specific implementation, firstly, the position of the person's face and the position of the webcam are obtained, and the line connecting the person's face and the webcam is determined. The spatial difference angle is then rotated to 0 degrees, and the adjusted angle is used as the mapping angle. Next, a key facial feature is initialized as the center of the corresponding recognition space. For example, the person's nose can be initialized as the center of the corresponding recognition space. In other embodiments, other key facial features can also be set as the center of the recognition space, which is not limited here. Secondly, all local contour features in the recognition space are linearly combined from largest to smallest. Finally, the recognition space composed of all the initialized and linearly combined local contour features is used as the facial contour recognition space. The facial contour recognition space contains the local contour features corresponding to the combination image of all key facial features of the person. The local contour features in the facial contour recognition space can more precisely describe the shape and structure of the face. Different local contour features can represent the differences in facial features under different poses. Constructing the facial contour recognition space is beneficial for calculating the mapping cost when mapping all local contour features.

[0069] It should be noted that in this embodiment, contour mapping of pose fusion features is performed by spatial difference angle, which can accurately obtain the facial contour recognition space of a person. At the same time, considering the different standing positions of the person, the recognition space can be adjusted according to different situations, making the methods described in some embodiments of this application more adaptable.

[0070] Additionally, it should be noted that in this application, the pose fusion features in the facial contour recognition space are deconvolved using a preset neural network model. Specifically, the transposed convolution kernel is initialized by setting a scale, and deconvolution is performed using a convolutional neural network with preset kernels. For example, the scale of deconvolution is 4, and the size and number of kernels can be set to 4×4. The scale in this application is determined based on the number of key facial features. That is, when the number of key facial features in the combined image of key facial features is set to 4, the scale of deconvolution is 4. Due to different application scenarios, different numbers of key facial features in this application can be set according to different application scenarios, which is not limited here.

[0071] Preferably, in this embodiment, reference Figure 3 As shown, this figure is an exemplary flowchart of determining the pose mapping cost in an embodiment of this application. In this embodiment, the pose mapping cost for face recognition determined by the local contour features obtained from deconvolution can be implemented using the following steps:

[0072] First, in step S31, a facial pose fusion map is determined based on the pose fusion features in the facial contour recognition space;

[0073] Then, in step S32, the pose loss vector for face recognition is determined using the facial pose fusion map and all local contour features;

[0074] Finally, in step S33, the pose mapping cost during face recognition is determined by the convolution scale of the deconvolution between the pose loss vector during face recognition and the pose fusion features in the facial contour recognition space.

[0075] In specific implementation, the MATLAB discrete image method in the existing technology can be used to construct all feature values ​​of the pose fusion feature into a scatter point grayscale image with a pixel size of 400*400, and use this grayscale image as the facial pose fusion map of the person. The facial pose fusion map represents the image obtained by mapping all features in the pose fusion feature. The facial pose fusion map can reflect all the initial features of the person's facial contour. In particular, mapping the local contour features obtained by deconvolution of the pose fusion feature to the facial contour recognition space is more conducive to subsequent processing and reduces the recognition error caused by the multi-form changes of the person.

[0076] In addition, in specific implementation, the initial position of each facial key part combination image is obtained during face recognition. The displacement and rotation angle of each facial key part combination image and the facial pose fusion image are calculated respectively. The calculated displacement and rotation angle are linearly combined to obtain the pose loss vector of the facial pose fusion image when the pose changes in each adjacent frame, that is, pose loss vector = (displacement, rotation angle). The displacement can be determined by calculating the average slope of all corresponding pixels in all facial key part combination images and the facial pose fusion image due to pose changes. Then, all pose loss vectors are randomly arranged into a column to form a pose loss matrix. The average of the first column data of the pose loss matrix and the average of the second column data of the pose loss matrix are used as the pose compensation array. It should be noted that the pose compensation array in this application refers to the standard array of pose loss vectors of the facial key part combination image. The pose compensation array can represent the overall trend of the pose change of the facial key parts of a person.

[0077] In specific implementation, firstly, the first value of the pose compensation array (i.e., the average displacement of all facial key part combination images) and the second value of the pose compensation array (i.e., the average rotation angle of all facial key part combination images) are used as the direction references for transposed convolution. Finally, transposed convolution is performed on all pose loss vectors using the set convolution kernel and direction reference to obtain multiple local contour features of the person. Then, the Euclidean distance between each local contour feature in the facial contour recognition space and the facial pose fusion image is used as the corresponding local contour feature displacement deviation. The displacement deviation represents the feature difference of the target face due to pose, expression and other factors when the pose fusion feature is mapped to the recognition space, and the average of all displacement deviations is used as the pose mapping cost when mapping the contour.

[0078] It should be noted that the pose mapping cost in this application represents the degree of deviation when the pose fusion feature is mapped to a local contour feature. The larger the pose mapping cost, the greater the degree of deviation when the pose fusion feature is mapped to a local contour feature. The smaller the pose mapping cost, the smaller the degree of deviation when the pose fusion feature is mapped to a local contour feature. Correcting facial feature changes caused by different postures, expressions and other factors through the pose mapping cost can reduce the error of face recognition.

[0079] In step S4, the ambient light intensity during face recognition is obtained, and the face recognition process is compensated based on the ambient light intensity and the mapping cost to obtain the compensation feature quantity.

[0080] In practice, the ambient light intensity during face recognition is obtained through the light sensor of the depth camera. That is, the light intensity sensor that works together with the network camera is connected to obtain the ambient light intensity data of the corresponding time period of face recognition, and the average value of the ambient light intensity data in that time period is used as the ambient light intensity when the network camera takes pictures.

[0081] In this embodiment, the face recognition process is compensated based on the ambient light intensity and the mapping cost to obtain the compensated feature quantity, which specifically includes:

[0082] The feature correction amount in the face recognition process is determined based on the mapping cost;

[0083] The illumination compensation coefficient is determined based on the ambient light intensity.

[0084] Obtain the brightness reference value of the facial contour recognition space;

[0085] Based on the feature correction amount, the illumination compensation coefficient, and the brightness reference value, the face recognition process is subjected to feature adjustment and illumination compensation, thereby obtaining the compensation feature amount.

[0086] In specific implementation, firstly, the feature correction amount can be determined by the following formula: Feature correction amount = 1 / (1 + mapping cost); then, the overall brightness during face recognition is read from the video playback software, and the ratio of the ambient light intensity to the overall brightness is used as the brightness error rate during face video acquisition. The brightness error rate represents the degree of deviation between the ambient light intensity and the actual video brightness when acquiring face video. The value range of the brightness error rate is between 0 and 1, that is, the larger the brightness error rate, the greater the deviation between the ambient light intensity and the actual video brightness, and the smaller the brightness error rate, the smaller the deviation between the ambient light intensity and the actual video brightness. Then, the illumination compensation coefficient can be determined by the difference between the preset face occlusion standard and the brightness error rate. As a preferred embodiment, the value of the preset face occlusion standard can be set to 1, that is: Illumination compensation coefficient = 1 - brightness error rate.

[0087] It should be noted that the illumination compensation coefficient in this application represents the deviation between the ambient light intensity and the overall brightness of the face video. The illumination compensation coefficient can take into account the ambient light intensity when the face video is captured. By the difference between the ambient light intensity and the overall brightness of the face video, it is determined whether there is occlusion in the captured face video, which provides a basis for subsequent illumination compensation of facial contour features. The overall brightness of the face video can be used as the overall brightness of continuous image frames.

[0088] In specific implementation, firstly, the `adjust_brightness` function in Python can be used to calculate the brightness reference value in the facial contour recognition space. This brightness reference value refers to the standard brightness in the facial contour space, typically ranging from 100 to 180 nits. Then, light compensation is applied to the face recognition process using both the illumination compensation coefficient and the brightness reference value. The amount of light compensation can be adjusted based on the ambient light intensity and the illumination compensation coefficient. In a preferred embodiment, the amount of light compensation can be determined using the following formula: Light compensation amount = Brightness reference value - Ambient light intensity × Illumination compensation coefficient. Light compensation improves the accuracy of the face recognition system under different lighting conditions. Finally, the light compensation amount and the feature correction amount are weighted and summed to obtain the compensated feature amount, which further improves the accuracy of face recognition.

[0089] It should be noted that the compensation feature quantity in this application refers to the fusion quantity of feature adjustment and light compensation. In particular, by adjusting the brightness of the face recognition process through the light compensation quantity, the impact of light changes on facial features can be reduced, so that the face recognition can still extract image features with consistent brightness even when the light intensity changes greatly, thereby improving the accuracy of face recognition. Feature adjustment can reduce the mapping cost in the face recognition process, which is conducive to improving the accuracy of face recognition.

[0090] In step S5, the matching confidence of face recognition is determined by the compensation feature quantity, and then the face recognition results are filtered based on the matching confidence.

[0091] In this embodiment, the matching confidence of face recognition is determined by the compensated feature quantity, and the face recognition results are then filtered based on the matching confidence. Specifically, this can be done in the following way:

[0092] Obtain all facial features to be matched from the database of the facial recognition system;

[0093] The matching confidence of face recognition is determined by all the face features to be matched and the compensation feature quantity;

[0094] The face recognition results are matched and filtered based on the matching confidence level of the face recognition.

[0095] In specific implementation, firstly, by connecting to the database of the face recognition system, the facial features to be matched of all face images to be matched in the database are extracted using an existing feature extraction model. The database can be a public database or can be self-built according to the actual application scenario; there is no limitation here. The database includes multiple preset face images to be matched. Then, an existing face recognition model (e.g., VGGFace, FaceNet, DeepFace, ArcFace, OpenFace, etc.) can be used to pre-adjust the compensation feature amount as a pre-adjustment for the face recognition process. The adjusted image features are then matched with the face features to be matched of each face image to be matched in the database. The matching confidence of each face feature to be matched is calculated using the corresponding confidence function in the face recognition model. Thus, the face image corresponding to the highest matching confidence is taken as the face recognition result of the person.

[0096] It should be noted that when all matching confidence scores are less than 75%, the face image corresponding to the highest matching confidence score is marked as a suspected match result and a prompt is made that further verification is needed. This indicates that there are no face images to be matched in the database. You can add face images to be matched or change the database selected when performing face recognition based on the actual application scenario.

[0097] Therefore, this application can realize dynamic feature adjustment and light compensation in the face recognition process. First, by acquiring continuous image frames, the dynamic features of the face can be captured, reducing recognition errors caused by instantaneous changes and providing more image feature information. Furthermore, the inter-frame difference vector reflects the changing trend of the face in a short period, facilitating dynamic correction of pose errors and reducing feature deviations caused by head rotation or tilt, thus making feature extraction more accurate and improving face matching accuracy under different poses. Second, by adjusting the pose fusion features through spatial difference angles, mapping them to a standardized facial contour recognition space, the effects of individual differences can be reduced. The impact of XOR shooting angle changes can be mitigated by using deconvolution operations to restore local contour details, making facial features clearer and improving the accuracy of feature matching. Then, by acquiring ambient light information in real time, images under low or high light conditions can be dynamically adjusted, and compensation can be made by combining pose mapping costs, which can effectively reduce the impact of lighting changes on feature extraction and improve the system's adaptability to different lighting conditions. Finally, by compensating for feature quantity, the final matching result of face recognition can be optimized, improving the system's adaptability to dynamic changes, reducing recognition errors, and improving the overall accuracy of face recognition.

[0098] In summary, the technical solution adopted in this application can realize dynamic feature adjustment and light compensation in the face recognition process, so as to improve the accuracy of face recognition.

[0099] Example 2: This application provides a face recognition system based on image analysis, referencing... Figure 4 As shown in the figure, this is a schematic diagram of a face recognition system according to this embodiment of the present application. The face recognition system includes:

[0100] Image acquisition module 100 is used to acquire continuous image frames during face recognition;

[0101] The pose fusion module 200 is used to determine the inter-frame difference vector based on the difference information between adjacent frames in the continuous image frames, and to perform pose fusion on the continuous image frames through all the inter-frame difference vectors to obtain pose fusion features.

[0102] The feature processing module 300 is used to map the pose fusion features to the facial contour recognition space based on the spatial difference angle during face recognition, perform deconvolution on the pose fusion features in the facial contour recognition space, and then determine the pose mapping cost during face recognition from the local contour features obtained by deconvolution.

[0103] The illumination compensation module 400 is used to obtain the ambient light intensity during face recognition, and to perform recognition compensation on the face recognition process based on the ambient light intensity and the mapping cost to obtain the compensation feature quantity.

[0104] The matching and filtering module 500 is used to determine the matching confidence of face recognition through the compensation feature quantity, and then filter the face recognition results based on the matching confidence.

[0105] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0106] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0107] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A face recognition method based on image analysis, characterized in that, The face recognition method includes: Capture continuous image frames during face recognition; Based on the difference information between adjacent frames in the continuous image frames, the inter-frame difference vector is determined, and the pose fusion features are obtained by performing pose fusion on the continuous image frames using all the inter-frame difference vectors. Based on the spatial difference angle during face recognition, the pose fusion features are mapped to the facial contour recognition space. The pose fusion features in the facial contour recognition space are deconvolved, and the pose mapping cost during face recognition is determined by the local contour features obtained from the deconvolution. The ambient light intensity during face recognition is obtained, and the face recognition process is compensated based on the ambient light intensity and the mapping cost to obtain the compensation feature quantity. The matching confidence of face recognition is determined by the compensation feature quantity, and then the face recognition results are filtered based on the matching confidence.

2. The face recognition method based on image analysis as described in claim 1, characterized in that, A depth camera captures continuous image frames during facial recognition.

3. The face recognition method based on image analysis as described in claim 1, characterized in that, Determining the inter-frame difference vector based on the difference information between adjacent frames in the continuous image frames specifically includes: Extract key feature maps from each of the consecutive image frames; The difference images of adjacent frames in the continuous image frames are determined, and then the inter-frame difference vector of the adjacent frames is determined by the difference images of the adjacent frames and all key feature maps.

4. The face recognition method based on image analysis as described in claim 1, characterized in that, Pose fusion is performed on the consecutive image frames using all inter-frame difference vectors to obtain pose fusion features, specifically including: Construct a pose fusion matrix based on all inter-frame difference vectors; The attitude fusion features are determined by the eigenvalues ​​of the attitude fusion matrix.

5. The face recognition method based on image analysis as described in claim 1, characterized in that, Mapping the pose fusion features to the facial contour recognition space based on the spatial difference angle during face recognition specifically includes: The mapping angle is determined based on the spatial difference angle; Based on the mapping angle, the pose fusion features are linearly mapped to obtain a facial contour recognition space.

6. The face recognition method based on image analysis as described in claim 1, characterized in that, The specific costs of determining pose mapping for face recognition using local contour features obtained from deconvolution include: A facial pose fusion map is determined based on the pose fusion features in the facial contour recognition space; The pose loss vector for face recognition is determined using the facial pose fusion map and all local contour features. The pose mapping cost during face recognition is determined by the convolution scale of the deconvolution performed on the pose loss vector during face recognition and the pose fusion features in the facial contour recognition space.

7. The face recognition method based on image analysis as described in claim 1, characterized in that, The pose fusion features in the facial contour recognition space are deconvolved using a pre-defined neural network model.

8. The face recognition method based on image analysis as described in claim 1, characterized in that, The ambient light intensity during face recognition is obtained through the light sensor of the depth camera.

9. The face recognition method based on image analysis as described in claim 1, characterized in that, The face recognition process is compensated based on the ambient light intensity and the mapping cost, and the compensated feature quantities specifically include: The feature correction amount in the face recognition process is determined based on the mapping cost; The illumination compensation coefficient is determined based on the ambient light intensity. Obtain the brightness reference value of the facial contour recognition space; Based on the feature correction amount, the illumination compensation coefficient, and the brightness reference value, the face recognition process is subjected to feature adjustment and illumination compensation, thereby obtaining the compensation feature amount.

10. A face recognition system based on image analysis, used to execute a face recognition method based on image analysis as described in any one of claims 1 to 9, characterized in that, The facial recognition system includes: The image acquisition module is used to acquire continuous image frames during face recognition. The pose fusion module is used to determine the inter-frame difference vector based on the difference information between adjacent frames in the continuous image frames, and to perform pose fusion on the continuous image frames through all the inter-frame difference vectors to obtain pose fusion features; The feature processing module is used to map the pose fusion features to the facial contour recognition space based on the spatial difference angle during face recognition, perform deconvolution on the pose fusion features in the facial contour recognition space, and then determine the pose mapping cost during face recognition from the local contour features obtained by deconvolution. The illumination compensation module is used to obtain the ambient light intensity during face recognition, and to perform recognition compensation on the face recognition process based on the ambient light intensity and the mapping cost to obtain the compensation feature quantity. The matching and filtering module is used to determine the matching confidence of face recognition through the compensation feature quantity, and then filter the face recognition results based on the matching confidence.

Images