High-discrimination feature extraction and recognition method based on low-quality three-dimensional face

Abstract

The invention provides a high-discrimination feature extraction and recognition method based on a low-quality three-dimensional face. The method is used for improving the recognition precision of the low-quality three-dimensional face. The method comprises the following steps: 1, preprocessing a low-quality three-dimensional face to obtain a normal graph; 2, a soft threshold denoising module is introduced for noise filtering; 3, adding an Inception module at the tail end of the backbone network to enhance the feature extraction capability; 4, extracting multi-scale features by using an attention multi-scale feature fusion module and carrying out key feature identification; 5, training a network model; and 6, testing the network model. According to the high-discrimination feature extraction and recognition method based on the low-quality three-dimensional face, noise interference in low-quality three-dimensional face data can be effectively reduced, face discrimination feature extraction is highlighted, and the recognition process is more accurate and reliable. The introduced Insection module can obtain the feature output of various receptive fields, so that the feature expression is richer and more diversified, and the feature discrimination is improved. In addition, in a multi-scale feature fusion module, an attention mechanism is introduced to perform weighted fusion on extracted features, and feature representation of identity-related information is further enhanced.

Description

Technical Field

[0001] The present application belongs to the technical field of image processing and pattern recognition, and more specifically to a recognition method for low-quality three-dimensional face data. Background Art

[0002] Traditional two-dimensional face recognition technology is easily affected by factors such as lighting and posture changes, while three-dimensional face recognition technology has better robustness due to its use of depth information. However, the high cost of collecting three-dimensional face data limits its application. With the popularization of consumer-grade three-dimensional sensors, it has become possible to obtain low-quality three-dimensional face data. Compared with high-quality three-dimensional faces, low-quality face images have lower data accuracy and a lot of noise, making them more difficult to recognize. Summary of the invention

[0003] 1. Technical issues to be solved

[0004] In view of the shortcomings of the existing technology, this application proposes a high-discriminative feature extraction and recognition method based on low-quality three-dimensional faces, which solves the problems of poor data quality and low discriminability of feature extraction faced by low-quality three-dimensional face recognition and improves recognition accuracy.

[0005] (II) Technical solution

[0006] 1. A method for extracting and recognizing high-discriminative features based on low-quality three-dimensional faces, characterized in that the steps include:

[0007] 1. Preprocess low-quality 3D faces to obtain normal maps;

[0008] Second, introduce the soft threshold denoising module to filter noise;

[0009] 3. Introduce the Inception module to enhance feature extraction capabilities;

[0010] Fourth, use the attention multi-scale feature fusion module to extract multi-scale features and identify key features;

[0011] 5. Network model training;

[0012] 6. Network model testing.

[0013] 2. The method for extracting and recognizing low-quality three-dimensional facial features according to item 1 is characterized in that: the step of preprocessing the low-quality three-dimensional face in step one includes: linear interpolation, nose tip correction, outlier removal, facial projection, hole filling and normal map estimation, and obtaining the normal map of the low-quality three-dimensional face.

[0014] 3. The low-quality three-dimensional face feature extraction and recognition method according to item 1 is characterized in that: the soft threshold denoising module (STM) step in step 2 includes: inserting the designed soft threshold denoising module after the first block of the backbone network to perform data denoising from the feature layer.

[0015] 4. The low-quality three-dimensional face feature extraction and recognition method according to item 1 is characterized in that: the step of introducing the Inception module in step three includes: adding the Inception module at the end of the backbone network to enhance the feature extraction capability.

[0016] 5. The low-quality three-dimensional facial feature extraction and recognition method according to item 1 is characterized in that: the step of using the attention multi-scale feature fusion module in step 4 includes: fusing features of multiple scales in the backbone network in an attention-weighted manner to help the model resist interference factors such as noise, deformation and scale changes in the input low-quality facial normal map, and obtain a more robust and stable feature representation.

[0017] 6. The low-quality three-dimensional facial feature extraction and recognition method according to item 1 is characterized in that the loss function of the network model in step five includes a joint gradient loss function composed of softmax and ArcFace.

[0018] 7. The method for extracting and recognizing low-quality three-dimensional facial features according to item 6, wherein the calculation formula of the ArcFace loss function is as follows:

[0019]

[0020] In the formula, N represents the total number of samples, C represents the number of categories, and θ yi represents the angle between the feature vector of the i-th sample and its true category, 6j represents the angle between the feature vector of the j-th category and the sample. S is a hyperparameter that can control the distribution of feature vectors in the feature space, and m is the angle interval, which represents the penalty term for the edge angle.

[0021] 8. The low-quality 3D face feature extraction and recognition method according to claim 6, wherein the calculation formula of the joint gradient loss function is as follows:

[0022] L = λL.+(1-λ)L.

[0023]

[0024] In the formula, λ is the weight parameter, i represents the number of iterations, and L s and L a They are softmax and Arcface loss functions respectively.

[0025] (III) Beneficial effects

[0026] Compared with the prior art, this application provides a highly discriminative feature extraction and recognition method based on low-quality three-dimensional faces, which has the following beneficial effects:

[0027] 1. A low-quality 3D face feature extraction and recognition method of the present application, the soft threshold denoising module introduced can perform data denoising at the feature level, obtain a cleaner feature representation, and reduce the impact of noise on the features;

[0028] 2. A low-quality 3D face feature extraction and recognition method of the present application, the introduced Inception module uses different convolution kernel branches to obtain feature outputs of multiple receptive fields and obtain more diverse feature representations;

[0029] 3. The attention multi-scale feature fusion module proposed in this application uses the attention mechanism to weightedly fuse the extracted multi-scale features to obtain a discriminative representation of low-quality facial features; BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is an overall framework diagram of a high-discriminative feature extraction and recognition method based on low-quality three-dimensional faces according to a specific implementation method of the present application.

[0031] Figure 2 This is the network structure diagram of the soft threshold denoising module.

[0032] Figure 3 This is the Inception module diagram.

[0033] Figure 4 Schematic diagram of the attention multi-scale feature fusion module. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0035] The present application provides a high-discriminative feature extraction and recognition method based on low-quality three-dimensional faces, which is used for accurate recognition of low-quality three-dimensional faces.

[0036] The present application provides a high-discriminative feature extraction and recognition method based on low-quality three-dimensional faces, introduces the Inception model, uses multiple convolution kernels and pooling operations of different sizes in parallel, extracts image features of different receptive fields at different scales, and helps the network model to simultaneously capture the feature information of low-quality three-dimensional faces at different scales, thereby improving the model's perception and expression capabilities for low-quality three-dimensional face data.

[0037] The present application provides an attention multi-scale feature fusion method, which integrates the attention mechanism in the multi-scale feature fusion process, so that the proposed model can more accurately focus on the key areas and features of low-quality three-dimensional faces, improve the feature's semantic modeling ability, and enhance the robustness and generalization ability of the proposed model.

[0038] The embodiments of the present invention are further described below in conjunction with the accompanying drawings:

[0039] like Figure 1 The overall framework diagram of a high-discriminative feature extraction and recognition method based on low-quality three-dimensional faces of the present application is shown as follows:

[0040] First, the low-quality 3D face data is preprocessed, which is conventional in the art, including linear interpolation, nose tip correction, outlier removal, face projection, hole filling and normal map estimation, and the normal map of the 3D face is obtained as the input of the network.

[0041] Secondly, a soft threshold denoising module is inserted into the backbone network to filter noise for extracting facial features. Where c is the number of channels, h and w are the height and width of the feature map, respectively. The feature map will be transformed through a connected structure consisting of a 3×3 convolution layer, a batch normalization layer, a ReLU activation layer, and a 3×3 convolution layer to obtain a feature map Y. Then take the absolute value and global pooling of the transformed feature map Y to obtain the initial threshold S of the soft threshold module. The initial value of S passes through a fully connected layer and a Sigmoid layer to obtain a scaling vector M ranging from 0 to 1. Apply M to vector S to obtain the final threshold S for each channel b Finally, the threshold is used to filter the noise in the feature map Y to obtain the output O of the soft threshold denoising module.

[0042] Then, add two Inception modules at the end of the backbone network. Figure 3As shown in the figure, by using multiple convolution kernels of different sizes and pooling operations in parallel, image features of different receptive fields are extracted from different scales and connected. This parallel operation allows the network to capture feature information at different scales at the same time, improving the network's perception and expression capabilities. Specifically: the number of input and output channels of each branch of the Inception1 module are (32, 32), (48, 64), (8, 16), (16, 16), respectively; the number of input and output channels of each branch of the Inception2 module are (64, 64), (96, 128), (16, 32), (32, 32), respectively.

[0043] Again, these features are input into the attention multi-scale feature fusion module, such as Figure 4 Specifically, features are extracted from the four convolutional blocks (Block1, Block2, Inception1, Inception2) in the backbone network as the input of the attention multi-scale feature fusion. These feature maps correspond to the information captured by different receptive fields. First, the feature maps F∈R extracted from different layers are H×W×C Attention weights are added separately to emphasize the contribution of pixels at different positions. The two-dimensional feature F is obtained by global channel average pooling avg ∈R H×W , represents the global feature information, where H represents the height of the feature map F, W represents the width of F, and C represents the number of channels of F. At the same time, in order to obtain the most discriminative information, the global channel maximum pooling is used to obtain the two-dimensional feature F max ∈R H×W , which represents the most discriminative feature in F. The global feature information and the most discriminative feature information are fused, and finally a feature map F with unchanged height and width and 1 channel number is obtained. fusion ∈R H×W Then, the sigmoid activation function is used to activate the feature map F fusion ∈R H×W Activate and get the feature map F sigmoid As the attention weight map of the feature map F.

[0044] In this way, each channel of the input feature map F is calculated according to the above steps to obtain the feature after adding attention That is, multiply each channel in F by F sigmoid Finally, the feature maps after adding attention are downsampled to the same size and concatenated. The convolution kernel size, step size and padding used in each maximum pooling layer are (33, 16, 16), (17, 8, 8), (9, 4, 4), and (3, 2, 1), respectively.

[0045] The obtained attention multi-scale fusion features are combined with the identity labels and the training network model based on the softmnax loss function is used until convergence.

[0046] Finally, the low-quality 3D face normal map to be tested is input into the trained network to recognize the test face image.

[0047] This application describes a method for extracting and recognizing high-discriminative features based on low-quality three-dimensional faces. In the training phase, the low-quality three-dimensional faces are first preprocessed to obtain the normal map of the low-quality three-dimensional faces; then the low-quality three-dimensional faces are input into the designed network model to perform attention multi-scale feature fusion; finally, the fused high-discriminative features are input into the deep network model together with the identity labels for training until convergence. In the testing phase, any low-quality three-dimensional face is given a series of conventional preprocessing and input into the trained deep learning model to obtain the predicted identity information.

[0048] Although the embodiments of the present application have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present application, and that the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A high-discriminative feature extraction and recognition method based on low-quality three-dimensional faces, characterized in that The following steps are involved:

1. Preprocess low-quality 3D faces to obtain normal maps; Second, introduce the soft threshold denoising module to filter noise; 3. Introduce the Inception module to enhance feature extraction capabilities; Fourth, use the attention multi-scale feature fusion module to extract multi-scale features and identify key features; 5. Network model training; 6. Network model testing.

2. The low-quality 3D face feature extraction and recognition method according to claim 1, wherein: The steps of preprocessing the low-quality three-dimensional face in step one include: linear interpolation, nose tip correction, outlier removal, face projection, hole filling and normal map estimation, and obtaining the normal map of the low-quality three-dimensional face.

3. The low-quality 3D face feature extraction and recognition method according to claim 1, wherein: The step of the soft threshold denoising module in step 2 includes: inserting the designed soft threshold denoising module after the first block of the backbone network to perform data denoising from the feature layer.

4. The low-quality 3D face feature extraction and recognition method according to claim 1, wherein: The step of introducing the Inception module in step three includes: adding the Inception module at the end of the backbone network to enhance the feature extraction capability.

5. The low-quality 3D face feature extraction and recognition method according to claim 1, wherein: The step of using the attention multi-scale feature fusion module in step 4 includes: fusing features of multiple scales in the backbone network in an attention-weighted manner to help the model resist interference factors such as noise, deformation, and scale changes in the input low-quality face normal map, and obtain a more robust and stable feature representation.

6. The low-quality 3D face feature extraction and recognition method according to claim 1, wherein: The loss function of the network model in step 5 includes softmax and ArcFace together forming a joint gradient loss function.

7. The low-quality three-dimensional face feature extraction and recognition method according to claim 6, characterized in that: The calculation formula of the ArcFace loss function is as follows: Among them, N represents the total number of samples, C represents the number of categories, and θ yi Represents the angle between the feature vector of the i-th sample and its true category, θ j Represents the angle between the feature vector of the jth category and the sample. S is a hyperparameter that controls the distribution of the feature vector in the feature space, and m is the angle interval, which represents the penalty term for the edge angle.

8. The low-quality three-dimensional face feature extraction and recognition method according to claim 6, characterized in that: The calculation formula of the joint gradient loss function is as follows: In the formula, λ is the weight parameter, i represents the number of iterations, and L s and L a They are softmax and Arcface loss functions respectively.

Images