Depression angle face recognition method and system applied to stadium entering

Abstract

The invention relates to the technical field of computer vision, in particular to a depression angle face recognition method and system applied to entering a venue. Comprising the following steps: S1, acquiring depression angle face data and corresponding high-definition front face image data; s2, performing three-dimensional face reconstruction on the front face image data by using a three-dimensional reconstruction network, re-mapping a rotated three-dimensional face model to two dimensions, and performing local cutting on two-dimensional depression angle face image data and a front face gallery set; s3, performing weighted fusion on the processed two-dimensional depression angle face image data and front face gallery set data by using a weight prediction network; s4, using the fused data to train a depression angle face recognition model; and S5, inputting depression angle gallery set data into the trained depression angle face recognition model, and counting the recognition precision. According to the invention, depression angle face recognition is applied to the scene of entering the venue, the accuracy of depression angle face recognition is improved, and the efficiency of entering the venue is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method and system for low-angle face recognition applied to entering a venue. Background Art

[0002] In today's era, various venues such as football stadiums, concert venues, and museums have widely used face recognition technology in entrance management. However, the existing face recognition methods have exposed many problems in actual scene applications.

[0003] From the perspective of recognition, most venues currently use head-up recognition. When multiple people appear in a picture at the same time, due to the limitation of the head-up angle, it is easy to be disturbed by the complex background environment, and the person behind cannot be recognized or mistakenly recognized, causing the person in front to enter and the person behind to be unable to repeat the recognition and enter the venue for the second time. Moreover, in terms of entry speed, existing face recognition technology usually requires visitors to take off their masks for recognition, which has low entry efficiency. In venues with dense traffic, it is also easy to cross-infect influenza. In addition, for people of different heights, such as taller visitors or children, in order to adapt to head-up recognition, they may need to squat or be picked up. This further slows down the entry speed. The application of low-angle face recognition in venue entry can reduce environmental interference, speed up recognition and entry speed, and usually does not need to take off the mask. In another case, the camera in the venue is usually installed at a high place for detection, and when it is used for face recognition, the low-angle face image is recognized, which can be compared with the data in the face database of the entrants, and the person to be found can be quickly found, such as looking for children.

[0004] In terms of the algorithm for face recognition at low angles, the current multi-pose face recognition scheme faces many challenges when applied to face recognition at low angles. The multi-pose face rotation is mainly based on the principle of face rotation and multi-frame information complementarity. Among them, the generative adversarial network (GAN) has become the mainstream solution for face rotation. It uses the principle of left-right symmetry of the face to correct the side face image to the front face posture to improve the recognition accuracy. However, the face at low angles lacks the left-right symmetry that can be used by the side face, which makes it difficult to accurately estimate the self-occluded part of the chin. Therefore, directly applying the GAN multi-pose face recognition scheme to face recognition at low angles often fails to achieve the expected results. In addition, training the GAN network requires a huge amount of data. However, in practical applications, it is often difficult to obtain such a large-scale training data set. On the one hand, collecting a large number of images of face images at low angles takes a lot of time and resources; on the other hand, due to the relatively small application scenarios of face recognition at low angles, the number of face images at low angles in the existing data sets is also relatively limited. When the training data set is too small, it will seriously affect the accuracy of face recognition at low angles. Furthermore, face recognition based on video sequences uses multiple frames of face images with complementary information to synthesize a single recognition feature, but this method requires multiple frames of images at different angles for feature fusion. However, in scenarios such as stadium entrance, it is usually necessary to quickly and accurately identify a single low-angle face, and this method obviously cannot meet this requirement. For example, during the entrance of a football match stadium, the audience needs to quickly enter the stadium through the face recognition device, and it is impossible for them to stay in front of the camera for a long time waiting for the acquisition and processing of multiple frames of images. In addition, the video sequence solution may also be affected by factors such as lighting changes and motion blur, further reducing the accuracy and reliability of recognition.

[0005] To sum up, the existing face recognition technology for venue entry has problems in both entering the venue and the algorithm. A method for rapid entry through low-angle face recognition is needed. At the same time, a technical solution for low-angle face recognition in the venue entry environment is also needed to solve the problem of face recognition, thereby improving the efficiency and accuracy of venue entry management and enhancing the visitor experience. Summary of the invention

[0006] In order to solve the above-mentioned problem of applying low-angle face recognition to entering a venue, improve the efficiency of entering the venue, and solve the problems of speed and accuracy of low-angle face recognition, the present invention provides a low-angle face recognition method and system for entering a venue.

[0007] In the first aspect, the present invention provides a method for low-angle face recognition for entering a venue, which adopts the following technical solution: A method for low-angle face recognition applied to entering a venue, comprising: S1. Obtaining the depression angle face data and the corresponding high-definition frontal face image data, processing the image, and constructing a gallery set including a depression angle gallery set and a frontal face gallery set; S2. Reconstruct the frontal face image data into three-dimensional face using a three-dimensional reconstruction network to obtain a three-dimensional face model, rotate the three-dimensional face model according to the depression angle, and remap the rotated three-dimensional face model to two dimensions to obtain two-dimensional depression angle face image data, and perform local excision processing on the two-dimensional depression angle face image data and the frontal face gallery set; S3. using a weight prediction network, weightedly fusion the processed two-dimensional depression angle face image data and the front face gallery data to obtain fused data; S4. construct a low-angle face recognition model, and use the fused data to train the low-angle face recognition model; S5. Input the high-angle gallery data into the trained high-angle face recognition model and calculate the recognition accuracy.

[0008] Furthermore, the image processing includes cropping, elastic deformation and illumination transformation of the gallery set data, adjusting the data size and performing normalization processing, and using an image annotation tool to clean and annotate the depression angle gallery set; Furthermore, the construction of the three-dimensional face model includes using a face alignment algorithm to regress the coordinates of key points of the face, and then mapping the two-dimensional face to the three-dimensional face model according to the coordinates of the key points of the face, and performing a top-down conversion of the three-dimensional model established by the frontal face gallery set at 15 degrees to 60 degrees to obtain two-dimensional top-down face image data.

[0009] Furthermore, the two-dimensional depression angle face image data and the front face gallery set are partially excised, including aligning the data using an open source Face-Alignment algorithm, excising the portion below the bridge of the nose, and using an image annotation tool to clean and annotate the sample below the bridge of the nose, thereby obtaining partial depression angle data and partial front face data; Furthermore, the construction of the depression angle face recognition model includes: B1. Introduce multiple convolutional layers to extract features at different levels of the fused data, and add a maximum pooling layer between the convolutional layers to retain key feature information; B2. Introduce a multi-model representation to process the error vector in the fused data, each fused data corresponds to a coefficient vector, and optimize the error vector and the coefficient vector; B3. Introduce the attention mechanism, by learning the importance of different regions in the fused data, assign different weights to different regions, use convolutional neural network to learn spatial attention weights, and weight the features of different regions; B4. Use the fully connected layer and activation function to learn the attention weight of each channel, highlight the important feature channels, and fuse the spatial attention weight and channel attention weight.

[0010] Furthermore, the training of the depression angle face recognition model includes: C1. Use the fused data as the input of the low-angle face recognition model, output the facial features, and label them as identity IDs; C2. Construct and optimize the loss function. The formula for constructing and optimizing the loss function is: Among them, L represents the difference between the model prediction result and the true label calculated using the cross entropy loss function. represents the regularization term, Represents the error model-dependent loss of multi-model representation and minimizes the loss function.

[0011] Furthermore, the difference between the model prediction result and the true label calculated using the cross entropy loss function is expressed as: in, is the sample size, is the number of categories, It is a sample The real label, =1 indicates sample Belongs to category ,otherwise =0, is the model prediction sample Belongs to category probability.

[0012] Furthermore, the regularization term is expressed by the formula: in, is the regularization term, are the model coefficients.

[0013] Further, the optimizing the error vector and the coefficient vector comprises: B21. The test sample is represented as a combination of the training sample and the error term, and the formula is: ,in is the test sample vector, is the training sample matrix, is the coefficient vector, is the error vector; B22. Estimate the coefficient vector and error vector by solving the constrained optimization problem, the formula of the constrained optimization problem is: ,in represents the norm of a vector, Indicates about and constraints.

[0014] Furthermore, the error model related loss formula is: in, is the square of the error norm, is a norm of the error vector, is a hyperparameter that adjusts the strength of the sparsity penalty.

[0015] In the second aspect, a low-angle face recognition system for entering a venue includes: A data acquisition module is configured to acquire the depression angle human face data and the corresponding high-definition frontal face image data, process the images, and construct a gallery set including a depression angle gallery set and a frontal face gallery set; a three-dimensional reconstruction processing module, configured to perform three-dimensional face reconstruction on the frontal face image data using a three-dimensional reconstruction network to obtain a three-dimensional face model, rotate the three-dimensional face model according to the depression angle adjustment, and remap the rotated three-dimensional face model to two dimensions to obtain two-dimensional depression angle face image data, and perform local excision processing on the two-dimensional depression angle face image data and the frontal face gallery set; The weighted fusion module is configured to perform weighted fusion on the processed two-dimensional depression angle face image data and the frontal face gallery set data by using a weight prediction network to obtain fused data; A model training module is configured to construct a low-angle face recognition model and train the low-angle face recognition model using the fused data; The recognition module is configured to input the low-angle gallery set data into the trained low-angle face recognition model and calculate the recognition accuracy.

[0016] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for low-angle face recognition applied to entering a venue.

[0017] In a fourth aspect, the present invention provides a terminal device including a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being suitable for being loaded and executed by the processor as a method for low-angle face recognition applied to entering a venue.

[0018] In summary, the present invention has the following beneficial technical effects: 1. The present invention proposes a method and system for low-angle face recognition for entering a venue, which avoids the problem that traditional head-up recognition is easily disturbed by complex backgrounds in multi-person scenes. By acquiring low-angle face images from a camera installed at a relatively high position, the influence of the surrounding environment on recognition is reduced. For people of different heights, no special actions are required, and visitors do not need to take off their masks for recognition, which prevents cross-influenza in crowded places, improves the fluency and efficiency of personnel entry, and ensures the safety of the entry process.

[0019] 2. The present invention proposes a method and system for low-angle face recognition for entering a venue. Data acquisition is easy to obtain. The user's frontal face and low-angle face information can be obtained by filling in information and collecting facial information before entering the venue. The low-angle face recognition model proposed by the present invention greatly improves the accuracy of recognition when entering the venue.

[0020] 3. The present invention uses a three-dimensional reconstruction network to reconstruct the frontal face image data in three dimensions, and performs depression angle adjustment and two-dimensional mapping to simulate various possible depression angle face situations, so that the model can learn more comprehensive facial features. At the same time, the two-dimensional depression angle face image data and the frontal face gallery set are locally excised to remove irrelevant information, highlight key features, improve the quality and effectiveness of the data, and thus solve the problem of unrecognizable due to the mask covering the lower half of the face.

[0021] 4. The processing of training samples in the present invention is different from that in the prior art. A three-dimensional reconstruction network is used and the rotated three-dimensional face model is remapped to two dimensions. Local excision processing is performed on the two-dimensional low-angle face image data and the front face gallery set. Then, the processed two-dimensional low-angle face image data and the front face gallery set data are weightedly fused to obtain the fused data as training samples. Because facial information of users will be collected through mobile terminals before entering the venue, the model trained by the present invention is more suitable for face recognition in the environment of entering the venue.

[0022] 5. The present invention introduces multiple convolutional layers to extract different levels of features of the fused data, and adds a maximum pooling layer between the convolutional layers to retain key feature information, so that the model can capture the details and abstract features of the face. The multi-model representation processes the error vector in the fused data, corresponds to a coefficient vector for each fused data, optimizes the error vector and the coefficient vector, and improves the model's ability to fit the data. The attention mechanism is introduced to learn the importance of different regions in the fused data, assign different weights to different regions, highlight important features, and enhance the model's attention to key information. The fully connected layer and activation function are used to learn the attention weight of each channel, highlight important feature channels, and fuse the spatial attention weight and channel attention weight, further improving the model's feature extraction and recognition capabilities.

[0023] 6. The method of the present invention is applicable to various venues, such as football match venues, concert venues, museums, etc. The method of the present invention is not only applicable to the rapid recognition of a single low-angle face, but also can be used to compare the data in the face database of the people who enter the venue, and quickly find the people to be found, such as children, etc. This is especially important in large venues, which can improve the security and management efficiency of the venues and provide an efficient and accurate solution for venue admission management. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of a method for low-angle face recognition applied to entering a venue in Embodiment 1 of the present invention; DETAILED DESCRIPTION The present invention is further described in detail below in conjunction with the accompanying drawings.

[0025] Example 1 Reference Figure 1 , a method for low-angle face recognition applied to entering a venue in this embodiment includes: S1. Obtaining the depression angle face data and the corresponding high-definition frontal face image data, processing the image, and constructing a gallery set including a depression angle gallery set and a frontal face gallery set; The image processing includes cropping, elastic deformation and illumination transformation of the gallery set data, adjusting the data size and performing normalization processing, and using an image annotation tool to clean and annotate the depression angle gallery set; Before entering the venue, users need to fill in information and perform facial verification on the mobile terminal, so as to prevent scalpers from snatching tickets to a certain extent. During the information filling and verification process, it is necessary to collect facial data from various angles such as frontal and overhead views through mobile terminals such as mobile phones, and then collect the user's frontal and overhead angle data; After collecting frontal and overhead facial data through mobile terminals such as mobile phones, the irrelevant background parts of the image are removed by cropping, and only the area containing the face is retained to reduce the subsequent calculation amount. The image is slightly elastically deformed to simulate the changes in the face at different angles and expressions to increase the diversity of the data. The lighting conditions of the image are adjusted, including brightness and contrast, to simulate facial images in different lighting environments. The image is resized to a uniform size to facilitate subsequent processing and model training. The pixel values ​​of the image are normalized to improve the training efficiency and stability of the model.

[0026] S2. Reconstruct the frontal face image data into three-dimensional face using a three-dimensional reconstruction network to obtain a three-dimensional face model, rotate the three-dimensional face model according to the depression angle, and remap the rotated three-dimensional face model to two dimensions to obtain two-dimensional depression angle face image data, and perform local excision processing on the two-dimensional depression angle face image data and the frontal face gallery set; The construction of the three-dimensional face model includes using a face alignment algorithm to regress the coordinates of key points of the face, and then mapping the two-dimensional face to the three-dimensional face model according to the coordinates of the key points of the face, and performing a top-down conversion of the three-dimensional model established by the frontal face gallery set at 15 degrees to 60 degrees to obtain two-dimensional top-down face image data, and the top-down conversion angle is performed randomly.

[0027] The face alignment algorithm is used to regress the coordinates of the key points of the face. The face alignment algorithm based on deep learning is used, such as the face key point detection algorithm in the Dlib library. The algorithm is based on the convolutional neural network (CNN) structure, which can automatically learn the features of the face and accurately detect the key points of the face, such as the coordinates of the eyes, nose, mouth, eyebrows, etc. The general three-dimensional face model 3DMM (3D Morphable Model) is used to map the two-dimensional face to the three-dimensional face model according to the coordinates of the face key points.

[0028] The 3D face model is transformed into a bird's-eye view using 3D graphics processing technology. Specifically, the bird's-eye view effect is achieved by rotating the 3D face model around the center plane. After the bird's-eye view transformation is completed, the 3D face model is reprojected onto a 2D plane, and the 2D bird's-eye view face image data is obtained by converting the coordinates of each point in the 3D model into pixel coordinates in a 2D image.

[0029] The two-dimensional depression angle human face picture data and the front face gallery set are subjected to partial excision processing, including using an open source Face-Alignment algorithm to align the data, excising the portion below the bridge of the nose, and using an image annotation tool to clean and annotate the sample below the bridge of the nose, thereby obtaining partial depression angle data and partial front face data; Specifically, the frontal face gallery contains 1,000 high-definition frontal face images of different people. Each image in the frontal face gallery is processed using the facial key point detection algorithm in the Dlib library. For each image, the algorithm automatically detects the coordinates of 68 key points, including the four corners of the eyes, eyebrows, brows, forehead, nose bridge, nose tip, nose wings, four corners of the mouth, high cheekbones, etc.

[0030] 3DMM is selected as a universal 3D face model. 3DMM is described by shape parameters and texture parameters, and different 3D face shapes are generated by adjusting these parameters. According to the detected facial key point coordinates, the correspondence between the 2D face image and the 3D face model is established.

[0031] For example, the coordinates of the key points of the eyes in the two-dimensional image are matched with the corresponding positions of the eyes in the three-dimensional face model, and the shape parameters and texture parameters are determined through geometric transformation and optimization algorithms so that the three-dimensional face model matches the two-dimensional face image as much as possible.

[0032] Select a central plane, a plane defined by the nose tip and two earlobes. Around this central plane, rotate the 3D face model at a certain angle. Here, a random angle between 15 and 60 degrees is selected for the top-down conversion. When a 30-degree top-down angle is selected, the 3D face model is rotated 30 degrees around the central plane through the rotation function in the 3D graphics processing library (such as OpenGL).

[0033] Using the projection formula, the coordinates of each point in the 3D model are converted to pixel coordinates in the 2D image.

[0034] Assume that the coordinates of a point in the 3D model are , the position and direction of the camera are known, through the projection formula and ,in is the focal length of the camera, calculate the pixel coordinates of the point in the two-dimensional image , perform such a transformation on each point in the three-dimensional face model to obtain two-dimensional depression-angle face image data.

[0035] The obtained two-dimensional low-angle face image data and frontal face gallery data are aligned using the open source Face-Alignment algorithm to ensure that the positions of the key parts of the face in the image are relatively fixed.

[0036] Then, the part below the bridge of the nose is removed. After determining the position of the nose bridge, the image with a vertical coordinate greater than the set value ( ) is removed.

[0037] Finally, the image annotation tool is used to clean and annotate the samples below the nose bridge, remove noise and abnormal data, and assign a label with a facial identity information to each sample.

[0038] Use the image annotation tool to clean and annotate the samples below the nose bridge, use the image annotation tool LabelImg to annotate the cleaned samples, and assign a label to each sample, which is the identity information of the face.

[0039] S3. using a weight prediction network, weightedly fusion the processed two-dimensional depression angle face image data and the front face gallery data to obtain fused data; The weight prediction network uses the convolutional neural network (CNN) in deep learning. The convolution layer is used to extract the features of the image, the pooling layer is used to reduce the dimension of the feature map, and the fully connected layer is used to output the weight prediction results. The processed two-dimensional tilted face image data and the frontal face gallery data are used as the input of the weight prediction network. These data can be expressed in matrix form. The training goal of the weight prediction network is to learn how to predict the appropriate weights based on the input data so that the fused data can better represent the features of the face.

[0040] S4. construct a low-angle face recognition model, and use the fused data to train the low-angle face recognition model; The method of constructing a low-angle face recognition model includes: B1. Introduce multiple convolutional layers to extract features at different levels of the fused data, and add a maximum pooling layer between the convolutional layers to retain key feature information; B2. Introduce a multi-model representation to process the error vector in the fused data, each fused data corresponds to a coefficient vector, and optimize the error vector and the coefficient vector; The error vector represents the difference between the test sample and the training sample, and the coefficient vector represents the weight of the training sample when representing the test sample. Each fused data corresponds to a coefficient vector, and by optimizing the error vector and the coefficient vector, the model can better fit the data. In this embodiment, the training sample is the fused data, and the test sample is the depression gallery set.

[0041] The optimizing the error vector and the coefficient vector comprises: B21. The test sample is represented as a combination of the training sample and the error term, and the formula is: ,in is the test sample vector, is the training sample matrix, is the coefficient vector, is the error vector; The coefficient vector and error vector are estimated by solving the constrained optimization problem, the formula of which is: ,in represents the norm of a vector, Indicates about and constraints.

[0042] For the coefficient vector, the sum of its elements is required to be 1, that is, This constraint ensures that when representing the test sample, the sum of the weights of all training samples is 1, avoiding unreasonable weights.

[0043] For the error vector, its norm is required to be less than or equal to a given value ,Right now This constraint can limit the size of the error so that the model does not make too large an error when fitting the data.

[0044] In order to solve this constrained optimization problem, the multi-gradient descent method is used to iteratively update the coefficient vector and the error vector so that the objective function gradually decreases while satisfying the constraints. The gradient of the objective function with respect to the coefficient vector and the error vector is calculated, and then the coefficient vector and the error vector are updated according to the direction and magnitude of the gradient. The coefficient vector update formula is: in, is the learning rate, which controls the step size of the update, The error vector update formula is: in, is the learning rate, represents the number of iterations, Through continuous iterative updates until the convergence condition is met, that is, when the change of the objective function is less than the threshold or the maximum number of iterations is reached, the estimated coefficient vector and error vector are obtained. The test sample is represented as a combination of training samples and error terms, and the coefficient vector and error vector are estimated by solving the constrained optimization problem, so that the model can better fit the data and improve the accuracy of low-angle face recognition.

[0045] B3. Introduce the attention mechanism, by learning the importance of different regions in the fused data, assign different weights to different regions, use convolutional neural network to learn spatial attention weights, and weight the features of different regions; By introducing the attention mechanism and using convolutional neural networks to learn spatial attention weights and weight the features of different regions, the model can pay more attention to the important areas in the fused data and improve the recognition accuracy and robustness of the model.

[0046] B4. Use the fully connected layer and activation function to learn the attention weight of each channel, highlight the important feature channels, and fuse the spatial attention weight and channel attention weight.

[0047] The fully connected layer can perform global analysis and learning on the input features to capture the relationship between different channels, and input the feature map after spatial attention weighting into the fully connected layer. The fully connected layer outputs the attention weight of each channel through learning. These weights represent the importance of each channel. The activation function can perform a nonlinear transformation on the channel attention weights output by the fully connected layer to highlight important feature channels. In this embodiment, the ReLU function is used as the activation function. The spatial attention weight and the channel attention weight are element-wise multiplied to obtain the fused attention weight.

[0048] The training of the depression angle face recognition model includes: C1. Use the fused data as the input of the low-angle face recognition model, output the facial features, and label them as identity ID. This label is used to supervise the training of the model so that the model can learn the correspondence between the features of different faces and identities.

[0049] C2. Construct and optimize the loss function. The formula for constructing and optimizing the loss function is: Among them, L represents the difference between the model prediction result and the true label calculated using the cross entropy loss function. represents the regularization term, Represents the error model-dependent loss of multi-model representation and minimizes the loss function.

[0050] The difference between the model prediction result and the true label calculated using the cross entropy loss function is expressed as: in, is the sample size, is the number of categories, It is a sample The real label, =1 indicates sample Belongs to category ,otherwise =0, is the model prediction sample Belongs to category probability.

[0051] The regularization term is expressed as: in, is the regularization term, are the model coefficients.

[0052] The error model related loss formula is: in, is the square of the error norm, is a norm of the error vector, is a hyperparameter that adjusts the strength of the sparsity penalty.

[0053] The goal of the error model-dependent loss is to make the error vector as small as possible so that the model can better fit the data.

[0054] The total loss function is minimized by the optimization algorithm, and the parameters of the model are adjusted to make the model's prediction results closer to the true label. The parameters are updated according to the gradient to gradually reduce the loss function.

[0055] The parameter update formula is: in, are model parameters, is the learning rate, is the loss function with parameters The gradient at .

[0056] Through the above steps, the loss function of the low-angle face recognition model can be constructed and optimized to improve the performance and generalization ability of the model.

[0057] S5. Input the high-angle gallery data into the trained high-angle face recognition model and calculate the recognition accuracy.

[0058] In order to improve the computational efficiency, the images in the depression gallery can be processed in batches. The parallel processing capability of the computer can be used to speed up the prediction of the model. In order to ensure the stability of the precision calculation.

[0059] Example 2 This embodiment provides a low-angle face recognition system for entering a venue, including: A data acquisition module is configured to acquire the depression angle human face data and the corresponding high-definition frontal face image data, process the images, and construct a gallery set including a depression angle gallery set and a frontal face gallery set; a three-dimensional reconstruction processing module, configured to perform three-dimensional face reconstruction on the frontal face image data using a three-dimensional reconstruction network to obtain a three-dimensional face model, rotate the three-dimensional face model according to the depression angle adjustment, and remap the rotated three-dimensional face model to two dimensions to obtain two-dimensional depression angle face image data, and perform local excision processing on the two-dimensional depression angle face image data and the frontal face gallery set; The weighted fusion module is configured to perform weighted fusion on the processed two-dimensional depression angle face image data and the frontal face gallery set data by using a weight prediction network to obtain fused data; A model training module is configured to construct a low-angle face recognition model and train the low-angle face recognition model using the fused data; The recognition module is configured to input the low-angle gallery set data into the trained low-angle face recognition model and calculate the recognition accuracy.

[0060] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for low-angle face recognition applied to entering a venue.

[0061] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to implement a method for low-angle face recognition applied to entering a venue.

[0062] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for low-angle face recognition applied to entering a venue, characterized in that: include: S1. Obtaining the depression angle face data and the corresponding high-definition frontal face image data, processing the image, and constructing a gallery set including a depression angle gallery set and a frontal face gallery set; S2. Reconstruct the frontal face image data into three-dimensional face using a three-dimensional reconstruction network to obtain a three-dimensional face model, rotate the three-dimensional face model according to the depression angle, and remap the rotated three-dimensional face model to two dimensions to obtain two-dimensional depression angle face image data, and perform local excision processing on the two-dimensional depression angle face image data and the frontal face gallery set; S3. using a weight prediction network, weightedly fusion the processed two-dimensional depression angle face image data and the front face gallery data to obtain fused data; S4. construct a low-angle face recognition model, and use the fused data to train the low-angle face recognition model; S5. Input the high-angle gallery data into the trained high-angle face recognition model and calculate the recognition accuracy.

2. According to claim 1, a method for low-angle face recognition for entering a venue is characterized in that: The image processing includes cropping, elastic deformation and illumination transformation of the gallery set data, adjusting the data size and performing normalization processing, and using image annotation tools to clean and annotate the depression angle gallery set.

3. The method for low-angle face recognition for entering a venue according to claim 1, characterized in that: The construction of the three-dimensional face model includes using a face alignment algorithm to regress the key point coordinates of the face, and then mapping the two-dimensional face to the three-dimensional face model according to the key point coordinates of the face, and performing a top-down conversion of the three-dimensional model established by the frontal face gallery set at 15 degrees to 60 degrees to obtain two-dimensional top-down face image data.

4. The method for low-angle face recognition for entering a venue according to claim 1, characterized in that: The local excision processing of the two-dimensional depression angle face image data and the front face gallery set includes aligning the data using an open source Face-Alignment algorithm, removing the part below the bridge of the nose, and using an image annotation tool to clean and annotate the samples below the bridge of the nose to obtain partial depression angle data and partial front face data.

5. The method for low-angle face recognition for entering a venue according to claim 1, characterized in that: The method of constructing a low-angle face recognition model includes: B1. Introduce multiple convolutional layers to extract features at different levels of the fused data, and add a maximum pooling layer between the convolutional layers to retain key feature information; B2. Introduce a multi-model representation to process the error vector in the fused data, each fused data corresponds to a coefficient vector, and optimize the error vector and the coefficient vector; B3. Introduce the attention mechanism, by learning the importance of different regions in the fused data, assign different weights to different regions, use convolutional neural network to learn spatial attention weights, and weight the features of different regions; B4. Use the fully connected layer and activation function to learn the attention weight of each channel, highlight the important feature channels, and fuse the spatial attention weight and channel attention weight.

6. The method for low-angle face recognition for entering a venue according to claim 5, characterized in that: The optimizing the error vector and the coefficient vector comprises: B21. The test sample is represented as a combination of the training sample and the error term, and the formula is: ,in is the test sample vector, is the training sample matrix, is the coefficient vector, is the error vector; B22. Estimate the coefficient vector and error vector by solving the constrained optimization problem, the formula of the constrained optimization problem is: ,in represents the norm of a vector, Indicates about and constraints.

7. The method for low-angle face recognition for entering a venue according to claim 1, characterized in that: The training of the depression angle face recognition model includes: C1. Use the fused data as the input of the low-angle face recognition model, output the facial features, and label them as identity IDs; C2. Construct and optimize the loss function. The formula for constructing and optimizing the loss function is: Among them, L represents the difference between the model prediction result and the true label calculated using the cross entropy loss function. represents the regularization term, Represents the error model-dependent loss of multi-model representation and minimizes the loss function.

8. A low-angle face recognition system for entering a venue, characterized in that: include: A data acquisition module is configured to acquire the depression angle human face data and the corresponding high-definition frontal face image data, process the images, and construct a gallery set including a depression angle gallery set and a frontal face gallery set; a three-dimensional reconstruction processing module, configured to perform three-dimensional face reconstruction on the frontal face image data using a three-dimensional reconstruction network to obtain a three-dimensional face model, rotate the three-dimensional face model according to the depression angle adjustment, and remap the rotated three-dimensional face model to two dimensions to obtain two-dimensional depression angle face image data, and perform local excision processing on the two-dimensional depression angle face image data and the frontal face gallery set; The weighted fusion module is configured to perform weighted fusion on the processed two-dimensional depression angle face image data and the frontal face gallery set data by using a weight prediction network to obtain fused data; A model training module is configured to construct a low-angle face recognition model and train the low-angle face recognition model using the fused data; The recognition module is configured to input the low-angle gallery set data into the trained low-angle face recognition model and calculate the recognition accuracy.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing a method for low-angle face recognition applied to entering a venue as described in claim 1.

10. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method for low-angle face recognition applied to entering a venue as described in claim 1.

Images