A pose-robust face recognition method and apparatus and device
By acquiring multi-angle facial pose data and constructing a multi-representation subspace vector network, and employing pose-aware triplet loss and soft-maxout similarity calculation, the problem of low accuracy in face recognition under large pose changes is solved, thus improving the recognition effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAMEN ZHENJING TECH CO LTD
- Filing Date
- 2023-07-24
- Publication Date
- 2026-07-21
AI Technical Summary
Existing facial recognition technologies have low accuracy in scenarios with large changes in posture, making it difficult to effectively identify features of the same face in different postures.
By acquiring multi-angle facial pose data from different video sequences, a multi-representation subspace vector face recognition network is constructed. The robustness and recognition ability of the model are improved by using pose-aware triplet loss and soft-maxout similarity calculation methods.
It improves the accuracy and recognition rate of face recognition under large pose changes and enhances the ability to match faces in different poses.
Smart Images
Figure CN117133031B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of facial recognition technology, and more particularly to a pose-robust facial recognition method, apparatus, and device. Background Technology
[0002] Since the advent of the deep learning era, metric learning, especially in face recognition, has made tremendous progress. This progress stems from several aspects. First, deep learning has overcome the limitations of manually designed features in traditional machine learning methods, fully utilizing the powerful learning and representation capabilities of deep neural networks, and can automatically learn facial feature vectors using supervised gradient descent. Second, academia has made significant progress in this field, and industry has proposed many targeted solutions for face recognition. Industry has proposed loss functions based on sample pairs, such as Contrastive Loss and Triplet Loss, and later, to address the shortcomings of these two, researchers proposed Cross-batch Memory Loss, Lifted Structure Loss, and so on. Schemes based on the softmax function have seen explosive progress with the introduction of sphere loss, leading to a series of improved schemes, such as CosFace, ArcFace, and Circle Loss. These academic solutions have made the large-scale application of face recognition algorithms possible. Third, large-scale datasets have been proposed in the field of face recognition, such as VGG Face, MS-Celeb-1M, Green Vision 360k, and Tsinghua University's release of the world's largest face dataset containing 260 million samples. These large-scale datasets make it possible to train extremely powerful models.
[0003] Currently, facial recognition typically uses deep learning models to predict facial feature vectors. These feature vectors are then used to determine the distance between faces using cosine distance or L2 distance, and the distance function is used to determine the performance of downstream tasks in facial recognition. For face verification tasks, it is only necessary to extract the features of the faces to be compared and then determine whether the distance between them is less than a threshold. For face retrieval tasks, the features of different faces are pre-stored in a database. During prediction, the face to be verified is compared with the nearest feature in the database, and then a threshold is determined to see if it meets the requirements.
[0004] While facial recognition algorithms are becoming increasingly mature, their accuracy remains unsatisfactory in scenarios with significant changes in facial pose. Current facial recognition applications typically limit the user's pose to a frontal view and prompt the user to place their face within a designated area. This presents significant limitations in scenarios requiring significant changes in facial pose, such as recognizing and comparing faces in a photo album. This patent aims to address the shortcomings of current facial recognition methods in such scenarios and optimize recognition accuracy for situations with large pose variations. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a pose-robust face recognition method, apparatus, and device that can solve the problems of low accuracy of face recognition under large poses and low recognition rate of the same face under different poses.
[0006] According to one aspect of the present invention, a pose-robust face recognition method is provided, comprising: acquiring different video sequences and extracting face pose data from different angles from each video sequence;
[0007] Face pose data from different angles are grouped, and a three-group training set is built based on the grouping;
[0008] A multi-representation subspace vector face recognition network is constructed, and a multi-representation subspace vector face recognition model is obtained by training it using a three-group training set.
[0009] A multi-representation subspace vector face recognition model is used to recognize faces from different angles.
[0010] In the above technical solution, to improve the robustness of the algorithm to facial poses, this case first improves the training process. Corresponding data for frontal and side views are collected from multiple perspectives. Different pose data for the same face ID are obtained from video sequences. Second, considering the specific characteristics of the pose data in this case, a training set construction method using pose-aware triplet loss is proposed to strengthen feature learning for different poses. Third, conventional facial feature learning represents facial features as individual vectors and uses simple cosine similarity when comparing faces. This method struggles to represent different poses of the same face, and a simple cosine similarity is insufficient to characterize cross-pose similarity. In this case, features are implicitly divided into n subspaces. Through multiple representation subspace vectors, the network can more easily represent different facial poses. Finally, the model constructed using the above scheme is used to recognize faces from different angles. Unlike conventional single-feature vector schemes, this paper proposes a multi-pose combined feature representation scheme, which greatly improves the applicability of features under different poses. Without sacrificing performance, it applies soft-maxout when calculating feature similarity to improve the recall rate of frontal and profile face recognition. First, this paper significantly expands the sample richness of the face recognition training dataset through an innovative method for acquiring face pose data, greatly and specifically improving the model's robustness to large poses. Second, it proposes a groundbreaking method for representing and learning face feature vectors in multiple subspaces, greatly enhancing the representational power of the feature space. Finally, matching the multi-subspace representation method, a soft-maxout similarity calculation method is proposed, which can greatly improve the matching ability of faces in different poses.
[0011] In some embodiments, facial pose data from different angles includes:
[0012] Using frontal face data as a reference angle, side face data from different angles are obtained in the same video sequence as the face rotates from the left to the right.
[0013] In the aforementioned technical solution, this case collects corresponding data for frontal and side views from multiple perspectives. Different pose data for the same face ID are obtained from video sequences. Due to its inherent temporal continuity, video data possesses a constant ID prior within the same sequence. This can be used to obtain images of the same face in different poses. Within the same sequence, the face ID remains unchanged, and because the face's orientation is constantly changing, different samples of the same face ID can be collected. This greatly enriches the corresponding data for side and frontal views. The method of acquiring face pose data significantly expands the sample richness of the face recognition training dataset, greatly and specifically improving the model's robustness to large poses.
[0014] In some embodiments, the facial pose data includes: facial data in the same sequence, and static frontal facial data randomly extracted from a video sequence.
[0015] The above technical solution can maximize the amount of data on the frontal face, increase the sample size for subsequent model training, and improve model accuracy.
[0016] In some embodiments, the facial pose data includes: profile data derived from static frontal face data; specifically:
[0017] Bind the frontal face texture in the static frontal face data to the 3D mesh;
[0018] By adjusting the pose angle of the 3D mesh, side profile data from different angles can be obtained through re-rendering.
[0019] In the above technical solutions, video data remains the priority, while a large amount of static frontal face data is also available. A 3D approach is used to convert the frontal face data into side profile data, further expanding the side profile data. A schematic diagram of generating side profile poses from frontal face data is shown above. The principle is to bind the texture of the frontal face to a 3D mesh using 3D technology, and then adjust the pose to re-render data into different poses. This allows for the rapid accumulation of side profile data, solving the problem of insufficient side profile data extracted from video sequences.
[0020] In some embodiments, facial pose data from different angles are grouped, and a three-group training set is established based on the groupings. Specifically:
[0021] The faces are divided into groups based on their angle range;
[0022] Randomly select one group from several groups as the Anchor group;
[0023] Select several groups with a large angle difference from the Anchor group as the Positive group, or select a portion of the samples from several groups with a large angle difference from the Anchor group and several groups with a small angle difference as the Positive group.
[0024] Any other face group is designated as the Negative group.
[0025] In the above technical solution, to enhance feature learning across different poses, a pose-aware triplet loss is proposed. Traditional triplet loss often randomly selects anchors and positives during sampling, making it difficult to effectively sample reasonable triples. Therefore, this solution first groups the video sequences based on their angular characteristics, and then uses these groups for targeted triplet sampling to ensure sufficient difference between triples. It's important to understand that this solution does not completely abandon the traditional triplet strategy, but rather selectively constructs pose-aware triples to enhance cross-pose sample learning, thereby improving the model's recognition ability across poses.
[0026] In some embodiments, a multi-representation subspace vector face recognition network is constructed, specifically:
[0027] Choose any face recognition network and set the last embedding layer of the face recognition network to output n representation vectors.
[0028] In the aforementioned technical solutions, conventional facial feature learning represents facial features as individual vectors and uses simple cosine similarity when comparing faces. This method struggles to represent different poses of the same face, and a simple cosine similarity is insufficient to characterize cross-pose similarity. This patent implicitly divides features into n subspaces. The overall network structure is similar to conventional face recognition, and mature models such as reset100 can be used. Unlike conventional face recognition networks, the final embedding layer of this network outputs n representation vectors. Through multiple representation subspace vectors, the network can more easily represent different poses of the face, greatly improving the representation capability of the feature space.
[0029] In some embodiments, a multi-representation subspace vector face recognition model is used to recognize faces from different angles, specifically:
[0030] The model extracts facial features from the facial images to be compared;
[0031] The system compares the facial features to be compared with those stored in the database and outputs the similarity results.
[0032] The following calculation method is used when comparing facial features:
[0033]
[0034] In the formula, V1 and V2 represent the face image to be compared and the normalized vector of the face feature representation stored in the database, respectively, and their dimensions are n×D, where D represents the subspace dimension.
[0035] The above formula yields an n×n similarity matrix, where n is the number of elements in the similarity matrix. 2 The weight W for each similarity is obtained by performing SoftMax calculation on each element:
[0036] W = SoftMax(M)
[0037] The final similarity result is for each similarity m. i Weighted average:
[0038]
[0039] In the formula, w i m i These represent elements of W and M, respectively.
[0040] In the above technical solution, since there are multiple representation subspaces, in order to match the representation method of multiple subspaces, a soft-maxout similarity calculation method is proposed, which can greatly improve the matching ability of faces with different poses.
[0041] According to another aspect of the present invention, a pose-robust face recognition device is provided, comprising: an acquisition module, a construction module, a training module, and a recognition module that are sequentially connected in communication;
[0042] The acquisition module is used to acquire different video sequences and extract facial pose data from different angles from each video sequence;
[0043] The construction module is used to group facial pose data from different angles and establish a three-group training set based on the grouping.
[0044] The training module is used to construct a multi-representation subspace vector face recognition network and train it using a three-group training set to obtain a multi-representation subspace vector face recognition model.
[0045] The recognition module is used to recognize faces from different angles using a multi-representation subspace vector face recognition model.
[0046] In the above technical solution, to better utilize the method, different steps are sequentially built into different modules, and these modules are linked together, enabling more efficient training and use of the method. It should be noted that the principles and effects of each step have already been described above and will not be elaborated upon here.
[0047] According to another aspect of the present invention, a pose-robust face recognition device is provided, comprising:
[0048] At least one processor; and,
[0049] A memory communicatively connected to the at least one processor; wherein,
[0050] The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the pose-robust face recognition method described above.
[0051] In the above technical solution, to better operate and process the method, the method is stored in memory, and the processor executes the stored method. It should be noted that the principle and effect of each step have been described above and will not be elaborated upon here.
[0052] According to one aspect of the present invention, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described pose-robust face recognition method.
[0053] In the above technical solution, to better operate and use the method, the method is stored in a computer-readable storage medium and implemented using a processor. It should be noted that the principle and effect of each step have been described above and will not be elaborated upon here. Attached Figure Description
[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0055] Figure 1 This is a flowchart illustrating one embodiment of the posture-robust face recognition method of the present invention;
[0056] Figure 2 This is a schematic diagram of a video sequence, which is one embodiment of the pose-robust face recognition method of the present invention;
[0057] Figure 3 This is a schematic diagram of side face pose generation in one embodiment of the pose-robust face recognition method of the present invention;
[0058] Figure 4 This is a schematic diagram of pose angle grouping in one embodiment of the pose-robust face recognition method of the present invention;
[0059] Figure 5 This is a schematic diagram of the subspace of a multi-representation subspace vector face recognition network, which is one of the embodiments of the pose-robust face recognition method of the present invention. Detailed Implementation
[0060] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0061] This invention provides a posture-robust face recognition method, apparatus, and device, which can solve the problems of low accuracy of face recognition under large postures and low recognition rate of the same face under different postures.
[0062] Example 1
[0063] Please see Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the pose-robust face recognition method of the present invention. It should be noted that if substantially the same result is obtained, the method of the present invention does not necessarily require further elaboration. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, the method includes the following steps:
[0064] S101. Obtain different video sequences and extract facial pose data from different angles from each video sequence;
[0065] In this embodiment, facial pose includes frontal face, side face, and facial data depending on other angles.
[0066] In this embodiment, facial pose data from different angles includes:
[0067] Using frontal face data as a reference angle, side face data from different angles are obtained in the same video sequence as the face rotates from the left to the right.
[0068] In this embodiment, corresponding data for frontal and side views are collected from multiple perspectives. Different pose data for the same face ID are obtained from video sequences. Due to its inherent temporal continuity, video data possesses a constant ID prior within the same sequence. This can be used to obtain images of the same face in different poses. Within the same sequence, the face ID remains unchanged, and because the face's orientation is constantly changing, different samples of the same face ID can be collected. This greatly enriches the corresponding data for side and frontal views. The method of acquiring face pose data significantly expands the sample richness of the face recognition training dataset, greatly and specifically improving the model's robustness to large poses.
[0069] In this embodiment, the facial pose data includes: facial data from the same sequence, and static frontal facial data randomly extracted from the video sequence. This embodiment maximizes the amount of data on the frontal facial features, increasing the sample size for subsequent model training and improving model accuracy.
[0070] In this embodiment, the facial pose data includes: profile data converted from static frontal face data; specifically:
[0071] Bind the frontal face texture in the static frontal face data to the 3D mesh;
[0072] By adjusting the pose angle of the 3D mesh, side profile data from different angles can be obtained through re-rendering.
[0073] In this embodiment, video data is still prioritized, while a large amount of static frontal face data is also available. A 3D approach is used to convert the frontal face data into side profile data, further expanding the side profile data. A schematic diagram of generating side profile poses from frontal face data is shown above. The principle is to bind the texture of the frontal face to a 3D mesh using 3D technology, and then adjust the pose to re-render data into different poses. This allows for the rapid accumulation of side profile data, solving the problem of insufficient side profile data extracted from video sequences.
[0074] S102. Group the facial pose data from different angles and establish a three-group training set based on the grouping;
[0075] In this embodiment, facial pose data from different angles are grouped, and a three-group training set is established based on these groups. Specifically:
[0076] The faces are divided into groups based on their angle range;
[0077] Randomly select one group from several groups as the Anchor group;
[0078] Select several groups with a large angle difference from the Anchor group as the Positive group, or select a portion of the samples from several groups with a large angle difference from the Anchor group and several groups with a small angle difference as the Positive group.
[0079] Any other face group is designated as the Negative group.
[0080] In this embodiment, triplet loss is a widely used metric learning loss, and many subsequent metric learning methods are derived from triplet loss. To enhance feature learning across different poses, pose-aware triplet loss is proposed. Traditional triplet loss often randomly selects anchors and positives during sampling, making it difficult to effectively sample reasonable triples. Therefore, in this case, the angular characteristics of the video sequence are first used to group the data, and then targeted triplet sampling is performed using these groups to ensure that the differences between triples are sufficiently large. It is important to understand that this approach does not completely abandon the traditional triplet strategy, but rather enhances cross-pose sample learning by selectively constructing pose-aware triples, thereby improving the model's recognition ability across poses.
[0081] S103. Construct a multi-representation subspace vector face recognition network and train it using a three-group training set to obtain a multi-representation subspace vector face recognition model.
[0082] In this embodiment, a multi-representation subspace vector face recognition network is constructed, specifically:
[0083] Choose any face recognition network and set the last embedding layer of the face recognition network to output n representation vectors.
[0084] In this embodiment, the face recognition network can be any existing conventional face recognition network, and this embodiment is not limited thereto.
[0085] In this embodiment, the number of representation vectors output by the embedding layer is preferably three, which can balance the representation capability and the computational complexity of the subsequent soft-maxout.
[0086] In this embodiment, conventional facial feature learning represents facial features as individual vectors and uses simple cosine similarity when comparing faces. This method struggles to represent different poses of the same face, and simple cosine similarity is insufficient to characterize cross-pose similarity. This patent implicitly divides features into n subspaces. The overall network structure is similar to conventional face recognition, and mature models such as reset100 can be used. Unlike conventional face recognition networks, the final embedding layer of this network outputs n representation vectors. Through multiple representation subspace vectors, the network can more easily represent different facial poses, greatly improving the representation capability of the feature space. This invention first significantly expands the sample richness of the face recognition training dataset through an innovative method for acquiring facial pose data, greatly and specifically improving the model's robustness to large poses. Second, it proposes a groundbreaking method for representing and learning multi-subspace facial feature vectors, greatly enhancing the representation capability of the feature space. Finally, in line with the multi-subspace representation method, a soft-maxout similarity calculation method was proposed, which can greatly improve the matching ability of faces with different poses.
[0087] S104. Use a multi-representation subspace vector face recognition model to recognize faces from different angles.
[0088] In this embodiment, a multi-representation subspace vector face recognition model is used to recognize faces from different angles. Specifically:
[0089] The model extracts facial features from the facial images to be compared;
[0090] The system compares the facial features to be compared with those stored in the database and outputs the similarity results.
[0091] The following calculation method is used when comparing facial features:
[0092]
[0093] In the formula, V1 and V2 represent the face image to be compared and the normalized vector of the face feature representation stored in the database, respectively, and their dimensions are n×D, where D represents the subspace dimension.
[0094] The above formula yields an n×n similarity matrix, where n is the number of elements in the similarity matrix. 2 The weight W for each similarity is obtained by performing SoftMax calculation on each element:
[0095] W = SoftMax(M)
[0096] The final similarity result is for each similarity m.i Weighted average:
[0097]
[0098] In the formula, w i m i These represent elements of W and M, respectively.
[0099] In this embodiment, since there are multiple representation subspaces, a soft-maxout similarity calculation method is proposed to match the representation method of multiple subspaces, which can greatly improve the matching ability of faces with different poses.
[0100] Based on the method described in one of the embodiments, the specific algorithm flow is as follows:
[0101] To improve the robustness of the algorithm to human facial poses, this patent first improves the training process by collecting corresponding data from multiple aspects of the front and side faces.
[0102] 1. Obtain different pose data of the same face ID from video sequences.
[0103] Video data, due to its inherent temporal continuity, possesses a constant ID prior within the same sequence. This can be used to obtain images of the same face in different poses. See also... Figure 2 , Figure 2 This is a schematic diagram of a video sequence. It demonstrates that within the same sequence, the face ID remains constant, and because the face's orientation is constantly changing, different samples with the same face ID can be collected. This greatly enriches the corresponding data for side profiles and frontal views.
[0104] 2. Video data remains the priority, but there is also a large amount of static frontal face data; please refer to [link / reference]. Figure 3 This method uses a 3D approach to transform the frontal face data into a side profile, further expanding the side profile data. The diagram above illustrates the generation of side profile poses from frontal face data. The principle is to bind the texture of the frontal face to a 3D mesh using 3D technology, and then adjust the pose to re-render data into different poses.
[0105] 3. To enhance feature learning across different poses, a pose-aware triplet loss is proposed. Traditional triplet loss often uses random anchor and positive selection during sampling, making it difficult to effectively sample reasonable triples. During training, faces are grouped according to pose angle, such as... Figure 4As shown. After sampling, pose samples within group ae are randomly selected as anchors. Then, when selecting positive samples, positive samples from the same group are avoided, and samples with poses significantly different from those in the current group are selected. For example, if anchor group c is selected, positive samples from a and e are selected, and a small number of samples from bd are selected.
[0106] 4. Conventional facial feature learning represents facial features as individual vectors and uses simple cosine similarity for face comparison. This method struggles to represent different poses of the same face, and simple cosine similarity is insufficient to characterize cross-pose similarity. This patent implicitly divides features into three subspaces, as shown in Figure 5. The overall network structure is similar to conventional face recognition and can utilize mature models such as reset100. Unlike conventional face recognition networks, this patent's network outputs three representation vectors at the final embedding layer. Through multiple representation subspace vectors, the network can more easily represent different facial poses.
[0107] 5. Due to the existence of multiple representational subspaces, the following calculation method is used when comparing features between faces:
[0108]
[0109] Where V1 and V2 represent the normalized vectors representing two faces, respectively, and their dimensions are 3×D, where D represents the subspace dimension. This patent uses D=512. A 3×3 similarity matrix can be obtained by calculating using the above formula. Softmax is then performed on the 9 elements to obtain the weight of each similarity.
[0110] W = softmax(M)
[0111] The final similarity is for each similarity m. i Weighted average:
[0112]
[0113] Where w i m i These represent elements of W and M, respectively.
[0114] In this embodiment, firstly, an innovative method for acquiring facial pose data significantly expands the sample richness of the face recognition training dataset, greatly and specifically improving the model's robustness to large poses. Secondly, a groundbreaking method for representing and learning facial feature vectors in multiple subspaces is proposed, greatly enhancing the representational power of the feature space. Finally, matching the multiple subspace representation method, a soft-maxout similarity calculation method is proposed, which can significantly improve the matching ability of faces with different poses. This solution addresses the problems of low accuracy in face recognition under large poses and low recognition rate of the same face under different poses.
[0115] Example 2
[0116] A posture-robust face recognition device includes: an acquisition module, a construction module, a training module, and a recognition module that are sequentially connected in communication.
[0117] The acquisition module is used to acquire different video sequences and extract facial pose data from different angles from each video sequence;
[0118] The construction module is used to group facial pose data from different angles and establish a three-group training set based on the grouping.
[0119] The training module is used to construct a multi-representation subspace vector face recognition network and train it using a three-group training set to obtain a multi-representation subspace vector face recognition model.
[0120] The recognition module is used to recognize faces from different angles using a multi-representation subspace vector face recognition model.
[0121] In this embodiment, to better utilize the method of one of the embodiments, different modules are sequentially established for different steps, and these modules are connected in series, enabling more efficient training and use of the method. It should be noted that the principle and effect of each step have been described above and will not be elaborated upon here.
[0122] Example 3
[0123] A posture-robust facial recognition device includes:
[0124] At least one processor; and,
[0125] A memory communicatively connected to the at least one processor; wherein,
[0126] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the pose-robust face recognition method described in one embodiment.
[0127] In the above technical solution, in order to better operate and process the method described in one of the embodiments, the method is stored in a memory, and the stored method is executed by a processor. It should be noted that the principle and effect of each step have been described above and will not be elaborated further here.
[0128] Example 4
[0129] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the pose-robust face recognition method described in one embodiment.
[0130] In the above technical solution, to better operate and use the method described in one of the embodiments, the method is stored in a computer-readable storage medium and implemented using a processor. It should be noted that the principle and effect of each step have been described above and will not be elaborated further here.
[0131] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A pose-robust face recognition method, characterized in that, include: Acquire different video sequences and extract facial pose data from different angles from each video sequence; The faces are divided into several angle ranges according to their angles. A three-group training set is established based on the grouping. Among them, one group is randomly selected as the Anchor group; several groups with a large angle difference from the Anchor group are selected as the Positive group; or, some samples from several groups with a large angle difference from the Anchor group and several groups with a small angle difference are extracted as the Positive group; and other face groups are used as Negative groups. A multi-representation subspace vector face recognition network is constructed. The last embedding layer of the multi-representation subspace vector face recognition network outputs n representation vectors, each of which corresponds to a subspace, where n is a positive integer greater than or equal to 2. The multi-representation subspace vector face recognition network is then trained using a three-group training set to obtain a multi-representation subspace vector face recognition model. The face recognition model utilizes a multi-representation subspace vector face recognition system to identify faces from different angles. Specifically: The model extracts facial features from the facial images to be compared; The system compares the facial features to be compared with those stored in the database and outputs the similarity results. The following calculation method is used when comparing facial features: In the formula, and These represent the normalized vectors of the face images to be compared and the face feature representations stored in the database, respectively, and their dimensions are... Where D represents the subspace dimension; The above formula can be used to calculate... The similarity matrix is the matrix of similarity. Each element SoftMax Calculate the weight for each similarity. : The final similarity result is for each similarity. Weighted average: In the formula, Represent Element.
2. The pose-robust face recognition method as described in claim 1, characterized in that, Facial pose data from different angles, including: Using frontal face data as a reference angle, side face data from different angles are obtained in the same video sequence as the face rotates from the left to the right.
3. The pose-robust face recognition method as described in claim 2, characterized in that, The facial pose data includes: facial data in the same sequence, and static frontal facial data randomly extracted from the video sequence.
4. The pose-robust face recognition method as described in claim 2, characterized in that, The facial pose data includes: profile data derived from static frontal face data; specifically: Bind the frontal face texture in the static frontal face data to the 3D mesh; By adjusting the pose angle of the 3D mesh, side profile data from different angles can be obtained through re-rendering.
5. The pose-robust face recognition method as described in claim 1, characterized in that, Construct a face recognition network with multiple representation subspace vectors, specifically: Choose any face recognition network and set the last embedding layer of the face recognition network to output n representation vectors.
6. A posture-robust face recognition device, characterized in that, include: The acquisition module, construction module, training module, and recognition module are connected sequentially via communication. The acquisition module is used to acquire different video sequences and extract facial pose data from different angles from each video sequence; divide the facial angles into several angle ranges into groups, and establish a three-group training set based on the groups. Among these groups, one group is randomly selected as the Anchor group; several groups with angles significantly different from the Anchor group are selected as the Positive group, or a portion of samples from several groups with angles significantly different from the Anchor group and several groups with small angle differences are extracted as the Positive group; and other arbitrary facial groups are used as Negative groups. The construction module is used to group facial pose data from different angles and establish a three-group training set based on the grouping. The training module is used to construct a multi-representation subspace vector face recognition network. The last embedding layer of the multi-representation subspace vector face recognition network outputs n representation vectors, each of which corresponds to a subspace, where n is a positive integer greater than or equal to 2. The multi-representation subspace vector face recognition network is trained using a three-group training set to obtain a multi-representation subspace vector face recognition model. The recognition module is used to recognize faces from different angles using a multi-representation subspace vector face recognition model. Specifically: The model extracts facial features from the facial images to be compared; The system compares the facial features to be compared with those stored in the database and outputs the similarity results. The following calculation method is used when comparing facial features: In the formula, and These represent the normalized vectors of the face images to be compared and the face feature representations stored in the database, respectively, and their dimensions are... Where D represents the subspace dimension; The above formula can be used to calculate... The similarity matrix is the matrix of similarity. Each element SoftMax Calculate the weight for each similarity. : The final similarity result is for each similarity. Weighted average: In the formula, Represent Element.
7. A posture-robust face recognition device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the pose-robust face recognition method as described in any one of claims 1 to 5.
8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the posture-robust face recognition method according to any one of claims 1 to 5.