A three-dimensional face recognition method

Through the three-dimensional face recognition method and multi-feature fusion cascade classifier, combined with deep learning algorithms and neural network models, the problem of traditional two-dimensional face recognition technology being affected by factors such as lighting and posture is improved, and the robustness of recognition is improved and precision.

CN111582223BActive Publication Date: 2024-11-19MIANYANG YOUSEN TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010424913.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-05-19
Publication Date
2024-11-19
Estimated Expiration
2040-05-19

AI Technical Summary

Technical Problem

Traditional two-dimensional facial recognition technology is greatly affected by factors such as lighting, posture, and expressions, and the amount of information is insufficient, so it cannot effectively overcome the impact of makeup and other appendages on recognition.

Method used

Using a three-dimensional face recognition method, the grayscale features and texture features of the face's area of ​​interest are fused through a multi-feature fusion cascade classifier, and the deep learning algorithm is used for training, combining DBN and BP neural network models to improve classification accuracy .

Benefits of technology

It improves the robustness and accuracy of facial recognition, overcomes the shortcomings of two-dimensional facial recognition technology, enhances the ability to adapt to lighting and view changes, and reduces the impact of factors such as makeup on recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111582223B_ABST
    Figure CN111582223B_ABST
Patent Text Reader

Abstract

The present invention discloses a three-dimensional face recognition method, and relates to the field of face recognition technology. The present invention adopts a multi-feature fusion super cascade classifier, and fuses the grayscale features and texture features of the face interest area to facilitate the subsequent classifier training. The fused face information is learned and trained by a deep learning algorithm, and deep learning after fusion of various face images is realized, and a BP neural network model is applied to realize feedback learning, which not only performs dimensionality reduction processing on the acquired high-dimensional data information to obtain low-dimensional data that is easier to identify and calculate, but also applies the BP neural network to reversely fine-tune the structural parameters, further simplifying the complexity of the acquired data, so that the BP neural network greatly reduces the data processing amount, thereby improving the calculation efficiency. The present invention cascades strong classifiers to improve the ability of the classifier, wherein the number of strong classifications is greater than 5, and the data processing ability is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of face recognition, and more specifically to a three-dimensional face recognition method. Background Art

[0002] Face recognition technology has become a widely used intelligent biometric technology. Traditional face recognition methods are based on the theory of two-dimensional image analysis to represent, analyze, extract features and classify faces. Although face recognition algorithms based on two-dimensional images have achieved fruitful results in more than half a century of research, studies have shown that face recognition methods based on two-dimensional image analysis are greatly affected by imaging conditions such as lighting, posture, and expression. Compared with two-dimensional data, three-dimensional data contains spatial information of the face, which is inherent information of the face itself and has good robustness to changes in external conditions. The collected three-dimensional shape data of the face does not change with changes in lighting and viewing angles, and accessories such as makeup have a great impact on the image but not on the three-dimensional data. Moreover, three-dimensional data is richer in information than two-dimensional images. Three-dimensional data has explicit spatial shape information, while two-dimensional images are essentially just projections of three-dimensional objects in two-dimensional space, and a lot of information that is conducive to recognition is lost during the projection process. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention discloses a three-dimensional face recognition method, which uses a multi-feature fusion cascade classifier to fuse the grayscale features and texture features of the face area of ​​interest to facilitate subsequent classifier training; uses a super classifier to enhance the classification ability, improve the classification accuracy, and overcome the deficiencies in the prior art.

[0004] The present invention adopts the following technical solutions:

[0005] A three-dimensional face recognition method, the method comprising the following steps:

[0006] (S1) Input a face picture from a picture library;

[0007] In this step, before inputting the face picture, the picture is first preprocessed, and then the face matching model is set;

[0008] The image preprocessing methods include filtering and denoising, grayscale normalization, edge detection or size normalization. Through image processing, a relatively pure data image is output, eliminating a large number of irrelevant areas.

[0009] The face matching model method is an ICP registration method, which is an operation of geometrically transforming the same face images obtained at different times, different image blocks, and different shooting angles so that the image points of the same name are completely overlapped in position and orientation;

[0010] (S2) applying a multi-feature data fusion model to perform feature fusion on the face images in the image library;

[0011] (S21) extracting image features by different means, wherein the extraction means at least include eigenface method, elastic graph matching method, singular value decomposition method SVD, discrete cosine transform method DCT or wavelet transform method, principal component analysis PCA method or independent component analysis ICA method;

[0012] (S22) Fusion of multiple feature information, constructing a multiple feature information fusion algorithm; the fusion algorithm includes a feature layer fusion algorithm, a matching layer fusion algorithm and a decision layer fusion algorithm, wherein:

[0013] In the feature layer fusion algorithm, the input image data information is the output facial feature expression image data information after facial feature extraction in a variety of different ways;

[0014] In the matching layer fusion algorithm, the input image data information is the matching score output by the feature layer fusion algorithm;

[0015] In the decision layer fusion algorithm, the input image data information is the data information output by the matching layer fusion algorithm, and the decision layer fusion algorithm outputs the final discrimination result;

[0016] (S23) outputting the fused data information;

[0017] When the data image is output, the output image is judged. If the output image meets the user's needs, the fused data image is output. If the output image does not meet the user's needs, the feature layer fusion algorithm is returned to re-fusing the data.

[0018] (S3) applying a deep learning algorithm to learn and train the fused face information; wherein the deep learning algorithm model is a DBN deep learning algorithm model, and the algorithm steps of the DBN deep learning algorithm model are:

[0019] (1) The first layer of RBM is fully trained using the CD-1 algorithm, and the facial image feature vectors are mapped in different feature spaces to save multiple facial image feature information. Assuming that the input facial image sample is x, the first visible layer is denoted as X, then the first hidden layer h 1 The parameters are {W 1 ; α 1};

[0020] (2) After the first layer of RBM input is completed, it means that the weights and offsets of the first layer of RBM input have become fixed. At this time, depending on the state of the hidden neurons, the vector of the second layer of RBM is input;

[0021] (3) The second RBM is then fully trained and superimposed on top of the first one;

[0022] Similarly, in the last RBM layer, each element is calculated using the following logistic function:

[0023]

[0024] Then construct a new vector:

[0025] t l ={t l1 , t l2 , t l3 ,...t lj} T (7)

[0027] Then use this vector as the previous layer h l+1 Input vector, where: l∈{1, 2, ..., c-1};

[0028] (4) Follow steps (1)-(3) and repeat several times, then set the BP neural network model in the last layer of DBN, and use the output feature vector of RBM as the input feature vector of the model. In each layer of the trained RBM network, the user can only check whether the weights in their own layer have achieved the optimal mapping of the feature vector of that layer, rather than the feature vector mapping of the entire DBN. Finally, the back propagation network of the BP neural network will also propagate the error information from top to bottom to each layer of RBM, and the entire DBN network needs to be fine-tuned.

[0029] (5) In order to facilitate the recognition of face images, the entire energy metering face image training set is usually provided with label information that can be recognized by users. By setting it in this way, when training the first layer of RBM, the explicit layer of RBM can have explicit neurons that can be trained together and neurons with classification labels. Finally, h l+1 The parameters of the layer {W l+1 ; α l+1}, the final training result is:

[0030] h l+1 =α l +(W 1 ) T t l (8)

[0031] 6) Complete the entire DBN deep learning training, form a complete DBN deep learning algorithm network, and obtain the parameter set of the DBN network:

[0032] {W 1 , W 2 , ..., W c ; α 1 , α 2 , ..., α c}

[0033] Then the training and integration process of the RBM network training model is equivalent to the initialization of the deep BP network weight parameters, which reduces the training time of the DBN deep learning algorithm model. When performing face image information recognition, the BP algorithm in the DBN deep learning algorithm model is started to perform a local search on the received weight parameter space to save convergence time.

[0034] (S4) applying a super classifier to classify the data; the super classifier is a cascaded strong classifier;

[0035] (S5) Determine whether the recognition requirements are met. If the recognition requirements are met, output the recognition image data; if the recognition requirements are not met, return to step (S3) and perform data recognition again.

[0036] In the present invention, in the step (S2), the steps of the feature layer fusion algorithm are:

[0037] (1) Extract test samples from the face image database;

[0038] (2) Use principal component analysis (PCA) to reduce the dimensionality of big data;

[0039] (3) Normalize the features after dimensionality reduction of big data, and then determine the subspace so that the covariance matrix of each feature is the unit matrix; fuse the face image information through connection and complementation;

[0040] (4) Feature matching: matching facial image information through feature matching templates;

[0041] (5) Standardization score processing; wherein the standardization method is a Min-Max standardization method, a Z-score standardization method, a Mean-MAD standardization method or a Tanh standardization method;

[0042] (6) Output fusion scores, sum up different output information respectively, and obtain the image fusion value.

[0043] In the present invention, in the step (S2), the steps of the matching layer fusion algorithm are:

[0044] (1) Extract test samples from the face image database;

[0045] (2) Use principal component analysis (PCA) to reduce the dimensionality of big data;

[0046] (3) Standardization features; wherein the standardization method is a Min-Max standardization method, a Z-score standardization method, a Mean-MAD standardization method, or a Tanh standardization method;

[0047] (4) Output fusion score;

[0048] (5) Discriminant subspace;

[0049] (6) Feature matching output.

[0050] In the present invention, the number of layers of the DBN deep learning algorithm model is greater than 3.

[0051] In the present invention, the cascade strong classifier is a Haar classifier including an Adaboost algorithm.

[0052] In the present invention, the Haar classifier integrates Haar-like features, integral graph method, Adaboost algorithm and cascade.

[0053] In the present invention, the main steps of the Haar classifier are:

[0054] (1) Extract Haar-like features;

[0055] (2) Accelerate the extraction of Haar-like features using the integral graph method;

[0056] (3) Use the Adaboost algorithm to train a strong classifier to distinguish between faces and non-faces;

[0057] (4) Use screening cascades to cascade strong classifiers together to improve detection accuracy.

[0058] In the present invention, the Haar-like feature is the sum of the pixel values ​​corresponding to the white area in the rectangular feature minus the sum of the pixel values ​​corresponding to the black area.

[0059] In the present invention, the working method of the AdaBoost classifier is:

[0060] (1) Data sample selection: In the collected data, assume that the training set {(x1, y1), (x2, y2), …, (x n ,y n )}, in this training set, assuming X i is the input data training sample vector, yi It is the category mark classified by AdaBoost classifier;

[0061] (2) Initialization: When initializing the data samples, all data training samples are given the same weight D = 1 / n, and then the AdaBoost classifier learning algorithm is used to perform N rounds of learning and training on the training sample set; where: D1(x i )=1 / n, i=1, 2,...,n.

[0062] (3) Training weights: After training the weak classifier into a strong classifier, t Under this condition, the weak learning algorithm is trained to obtain the prediction function h t :X→{-1,+1};

[0063] (4) Output the predicted value and evaluate the error rate of the predicted value. The error calculation formula is:

[0064] ε t =∑D t (x i )[h t (x i )≠y i Formula (1)

[0065] (5) Determine whether the requirements are met. When h t (x i )≠y i When t (x i )=-1D t (x i )exp(-α t y i h t (x i )) Formula (2)

[0066] When h t (x i )=y i hour,

[0067] y i h t (x i )=e a Formula (3)

[0068] Where Z t The following formula is normalized to:

[0069]

[0070] (6) Output the final prediction function and obtain the normalized image data. After N rounds of training, the final prediction function is:

[0071]

[0072] In the present invention, the cascade strong classifier is obtained by adding trained strong classifiers, and the number of the strong classifiers is greater than 5.

[0073] Positive effect

[0074] The present invention adopts a multi-feature fusion super cascade classifier to fuse the grayscale features and texture features of the face area of ​​interest to facilitate the subsequent classifier training;

[0075] The present invention applies a deep learning algorithm to study and train the fused face information, realizes deep learning after fusing various face images, and digs out the deep meaning hidden in the surface image; in the deep learning process, the BP neural network model is applied to realize feedback learning, which not only performs dimensionality reduction processing on the acquired high-dimensional data information to obtain low-dimensional data that is easier to identify and calculate, but also applies the BP neural network to reversely fine-tune the structural parameters, further simplifying the complexity of the acquired data, greatly reducing the data processing amount of the BP neural network, thereby improving the calculation efficiency.

[0076] The present invention cascades strong classifiers to improve the capabilities of the classifiers, wherein the number of strong classifiers is greater than 5, thereby improving the data processing capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0078] Figure 1 A schematic diagram of a process of a three-dimensional face recognition method of the present invention;

[0079] Figure 2 A schematic diagram of a process flow of a three-dimensional face recognition method based on a multi-feature fusion cascade classifier in the present invention;

[0080] Figure 3 A schematic diagram of constructing a multi-feature information fusion algorithm in a three-dimensional face recognition method of the present invention;

[0081] Figure 4 A method for constructing a multi-feature data fusion model in a three-dimensional face recognition method of the present invention;

[0082] Figure 5 A method for constructing a matching layer multi-feature fusion multi-feature data fusion model in a three-dimensional face recognition method of the present invention;

[0083] Figure 6 A method for constructing a multi-feature fusion model of multi-feature data fusion in a feature layer in a three-dimensional face recognition method of the present invention;

[0084] Figure 7 This is a DBN deep learning algorithm model in a three-dimensional face recognition method of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] like Figure 1 As shown, a three-dimensional face recognition method comprises the following steps:

[0087] (S1) Input a face picture from a picture library;

[0088] In this step, before inputting the face picture, the picture is first preprocessed, and then the face matching model is set;

[0089] The image preprocessing methods include filtering and denoising, grayscale normalization, edge detection or size normalization. Through image processing, a relatively pure data image is output, eliminating a large number of irrelevant areas.

[0090] The face matching model method is an ICP registration method, which is an operation of geometrically transforming the same face images obtained at different times, different image blocks, and different shooting angles so that the image points of the same name are completely overlapped in position and orientation;

[0091] In the preprocessing step of this step, an important step in the face recognition system is the preprocessing process of the face image. Due to the influence of various factors, the quality of the input image is not uniform, and the image cannot be directly used in the subsequent face recognition process, which greatly affects the recognition performance. Therefore, image preprocessing has a standardized and unified standard to eliminate the influence of size, posture, brightness, facial occlusions, etc. on recognition performance as much as possible. Different face recognition systems use different preprocessing methods according to the image source and recognition algorithm they use. Commonly used face image preprocessing methods include: filtering denoising, grayscale transformation (grayscale normalization), image binarization, edge detection, size normalization, etc. Only one or several preprocessing methods may be used in the same system. Each step is explained below.

[0092] (1) Filtering and denoising: Noise causes distortion and degradation to the image. Using filtering to remove noise before feature extraction is a necessary step in face recognition systems. There are many filtering methods, including linear filtering, median filtering, and adaptive filtering.

[0093] (2) Grayscale transformation (also called grayscale normalization): Grayscale transformation is a type of image enhancement technology. Grayscale transformation can be used to compensate for uneven illumination in the original image, so that the face image to be identified follows the same or similar grayscale distribution. Only in this way can different images be comparable during feature extraction and recognition. This process is also called grayscale normalization. The grayscale transformation methods commonly used in face recognition systems are: histogram equalization based on image statistical features, histogram normalization, and grayscale mean variance normalization.

[0094] (3) Edge detection: In image processing, local discontinuities in an image are called "edges". Image edge extraction is to retain these important edge information in the original image while reducing the total amount of data for later processing. From a mathematical analysis, edge extraction is to retain the areas where the grayscale changes dramatically in the image; from the perspective of signal processing, the areas where the grayscale changes dramatically are the areas with higher image frequencies, which can be filtered and retained using a high-pass filter. Edge detection of input face images is a preprocessing method used by many face recognition systems when detecting and locating faces. There are many edge detection methods, mainly: differential operator method, Sobel operator method, Laplace operator method, Canny operator method, etc. Each operator has different detection capabilities for edges in different directions and the ability to suppress noise.

[0095] (4) Scale normalization: Scale normalization of face images includes image translation, rotation, scaling, and standard cutting. Translation transformation is probably the simplest type of geometric transformation. Image translation reflects the translation invariance of the face in the image plane. The rotation of the face image is to make the eyes parallel to the horizontal axis, and rotate all the pixels of the image with the center of the eyes as the rotation point. This ensures the consistency of the center position of the face and reflects the rotation invariance of the face in the image plane.

[0096] (S2) applying a multi-feature data fusion model to perform feature fusion on the face images in the image library;

[0097] (S21) extracting image features by different means, wherein the extraction means at least include eigenface method, elastic graph matching method, singular value decomposition method SVD, discrete cosine transform method DCT or wavelet transform method, principal component analysis PCA method or independent component analysis ICA method;

[0098] (S22) Fusion of multiple feature information, constructing a multiple feature information fusion algorithm; the fusion algorithm includes a feature layer fusion algorithm, a matching layer fusion algorithm and a decision layer fusion algorithm, wherein:

[0099] In the feature layer fusion algorithm, the input image data information is the output facial feature expression image data information after facial feature extraction in a variety of different ways;

[0100] In the matching layer fusion algorithm, the input image data information is the matching score output by the feature layer fusion algorithm;

[0101] In the decision layer fusion algorithm, the input image data information is the data information output by the matching layer fusion algorithm, and the decision layer fusion algorithm outputs the final discrimination result;

[0102] (S23) outputting the fused data information;

[0103] When the data image is output, the output image is judged. If the output image meets the user's needs, the fused data image is output. If the output image does not meet the user's needs, the feature layer fusion algorithm is returned to re-fusing the data.

[0104] (S3) applying a deep learning algorithm to learn and train the fused face information; wherein the deep learning algorithm model is a DBN deep learning algorithm model, and the algorithm steps of the DBN deep learning algorithm model are:

[0105] (1) The first layer of RBM is fully trained using the CD-1 algorithm, and the facial image feature vectors are mapped in different feature spaces to save multiple facial image feature information. Assuming that the input facial image sample is x, the first visible layer is denoted as X, then the first hidden layer h 1 The parameters are {W 1 ; α 1};

[0106] (2) After the first layer of RBM input is completed, it means that the weights and offsets of the first layer of RBM input have become fixed. At this time, depending on the state of the hidden neurons, the vector of the second layer of RBM is input;

[0107] (3) The second RBM is then fully trained and superimposed on top of the first one;

[0108] Similarly, in the last RBM layer, each element is calculated using the following logistic function:

[0109]

[0110] Then construct a new vector:

[0111] t l ={t l1 , t l2 , t l3 ,...t lj} T (7)

[0113] Then use this vector as the previous layer h l+1 Input vector, where: l∈{l, 2, ..., cl};

[0114] (4) Follow steps (1)-(3) and repeat several times, then set the BP neural network model in the last layer of DBN, and use the output feature vector of RBM as the input feature vector of the model. In each layer of the trained RBM network, the user can only check whether the weights in their own layer have achieved the optimal mapping of the feature vector of that layer, rather than the feature vector mapping of the entire DBN. Finally, the back propagation network of the BP neural network will also propagate the error information from top to bottom to each layer of RBM, and the entire DBN network needs to be fine-tuned.

[0115] (5) In order to facilitate the recognition of face images, the entire energy metering face image training set is usually provided with label information that can be recognized by users. By setting it in this way, when training the first layer of RBM, the explicit layer of RBM can have explicit neurons that can be trained together and neurons with classification labels. Finally, h l+1The parameters of the layer {W l+1 ; α l+1}, the final training result is:

[0116] h l+1 =α l +(W l ) T t l (8)

[0117] (6) The entire DBN deep learning training is completed, forming a complete DBN deep learning algorithm network, and the parameter set of the DBN network is obtained as follows:

[0118] {W 1 , W 2 , ..., W c ; α 1 , α 2 , ..., α c}

[0119] Then the training and integration process of the RBM network training model is equivalent to the initialization of the deep BP network weight parameters, which reduces the training time of the DBN deep learning algorithm model. When performing face image information recognition, the BP algorithm in the DBN deep learning algorithm model is started to perform a local search on the received weight parameter space to save convergence time.

[0120] (S4) applying a super classifier to classify the data; the super classifier is a cascaded strong classifier;

[0121] (S5) Determine whether the recognition requirements are met. If the recognition requirements are met, output the recognition image data; if the recognition requirements are not met, return to step (S3) and perform data recognition again.

[0122] In a specific embodiment, in the step (S2), the steps of the feature layer fusion algorithm are:

[0123] (1) Extract test samples from the face image database;

[0124] (2) Use principal component analysis (PCA) to reduce the dimensionality of big data;

[0125] (3) Normalize the features after dimensionality reduction of big data, and then determine the subspace so that the covariance matrix of each feature is the unit matrix; fuse the face image information through connection and complementation;

[0126] (4) Feature matching: matching facial image information through feature matching templates;

[0127] (5) Standardization score processing; wherein the standardization method is a Min-Max standardization method, a Z-score standardization method, a Mean-MAD standardization method or a Tanh standardization method;

[0128] (6) Output fusion scores, sum up different output information respectively, and obtain the image fusion value.

[0129] In a specific embodiment, in the step (S2), the matching layer fusion algorithm is similar to the above steps, and the steps are:

[0130] (1) Extract test samples from the face image database;

[0131] (2) Use principal component analysis (PCA) to reduce the dimensionality of big data;

[0132] (3) Standardization features; wherein the standardization method is a Min-Max standardization method, a Z-score standardization method, a Mean-MAD standardization method, or a Tanh standardization method;

[0133] (4) Output fusion score;

[0134] (5) Discriminant subspace;

[0135] (6) Feature matching output.

[0136] In a specific embodiment, the number of layers of the DBN deep learning algorithm model is greater than 3. In a specific embodiment, the cascade strong classifier is a Haar classifier including an Adaboost algorithm.

[0137] In a specific embodiment, the Haar classifier integrates Haar-like features, integral graph method, Adaboost algorithm and cascade.

[0138] In a specific embodiment, the main steps of the Haar classifier are:

[0139] (1) Extract Haar-like features;

[0140] (2) Accelerate the extraction of Haar-like features using the integral graph method;

[0141] (3) Use the Adaboost algorithm to train a strong classifier to distinguish between faces and non-faces;

[0142] (4) Use screening cascades to cascade strong classifiers together to improve detection accuracy.

[0143] In a specific embodiment, the Haar-like feature is the sum of pixel values ​​corresponding to the white area in the rectangular feature minus the sum of pixel values ​​corresponding to the black area.

[0144] In a specific embodiment, the working method of the AdaBoost classifier is:

[0145] (1) Data sample selection: In the collected data, assume that the training set {(x1, y1), (x2, y2), …, (x n ,y n )}, in this training set, assuming X i is the input data training sample vector, y i It is the category mark classified by AdaBoost classifier;

[0146] (2) Initialization: When initializing the data samples, all data training samples are given the same weight D = 1 / n, and then the AdaBoost classifier learning algorithm is used to perform N rounds of learning and training on the training sample set; where: D1(x i )=1 / n, i=1, 2,...,n.

[0147] (3) Training weights: After training the weak classifier into a strong classifier, t Under this condition, the weak learning algorithm is trained to obtain the prediction function h t :X→{-1,+1};

[0148] (4) Output the predicted value and evaluate the error rate of the predicted value. The error calculation formula is:

[0149] ε t =∑D t (x i )[h t (x i )≠y i Formula (1)

[0150] (5) Determine whether the requirements are met. When h t (x i )≠y i hour,

[0151] h t (x i )=-1D t (x i )exp(-α t y i h t (x i )) Formula (2)

[0152] When h t (x i )=y i hour,

[0153] y i ht (x i )=e a Formula (3)

[0154] Where Z t The following formula is normalized to:

[0155]

[0156] (6) Output the final prediction function and obtain the normalized image data. After N rounds of training, the final prediction function is:

[0157]

[0158] In a specific embodiment, the cascade strong classifier is obtained by adding trained strong classifiers, and the number of the strong classifiers is greater than 5.

[0159] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A three-dimensional face recognition method, Features: The method comprises the following steps: (S1) inputting a face picture from a picture library; in this step, before inputting the face picture, firstly perform picture preprocessing, and then set a face matching model; wherein the picture preprocessing method is filtering denoising, grayscale normalization processing, edge detection or size normalization processing; through image processing, output a relatively pure data image, eliminate a large number of irrelevant areas, wherein the face matching model method is an ICP registration method, which is an operation of geometrically transforming the same face images obtained at different times, different image blocks, and different shooting angles so that the same-name image points are completely overlapped in position and orientation; (S2) applying a multi-feature data fusion model to perform feature fusion on the face images in the image library; (S21) extracting image features by different means, wherein the extraction means at least include eigenface method, elastic graph matching method, singular value decomposition method SVD, discrete cosine transform method DCT or wavelet transform method, principal component analysis PCA method or independent component analysis ICA method; (S22) fusing multiple feature information, constructing a multiple feature information fusion algorithm; the fusion algorithm includes a feature layer fusion algorithm, a matching layer fusion algorithm and a decision layer fusion algorithm, wherein: in the feature layer fusion algorithm, the input image data information is the output facial feature expression image data information after facial features are extracted in a variety of different ways; in the matching layer fusion algorithm, the input image data information is the matching score output by the feature layer fusion algorithm; in the decision layer fusion algorithm, the input image data information is the data information output by the matching layer fusion algorithm, and the decision layer fusion algorithm outputs the final discrimination result; (S23) outputting the fused data information; when the data image is output, judging the output image, if the output image meets the user's needs, outputting the fused data image, if the output image does not meet the user's needs, returning to the feature layer fusion algorithm and re-performing data fusion; (S3) applying a deep learning algorithm to learn and train the fused face information; wherein the deep learning algorithm model is a DBN deep learning algorithm model, and the algorithm steps of the DBN deep learning algorithm model are: (1) The first layer of RBM is fully trained using the CD-1 algorithm, and the facial image feature vectors are mapped in different feature spaces to save multiple facial image feature information. Assuming that the input facial image sample is x, the first visible layer is denoted as X, then the first hidden layer h 1 The parameters are {W 1 ; α 1 }; (2) After the first layer of RBM input is completed, it means that the weights and offsets of the first layer of RBM input have become fixed. At this time, depending on the state of the hidden neurons, the vector of the second layer of RBM is input; (3) The second layer of RBM is then fully trained and stacked on top of the first one; and so on, in the last RBM layer, each element is calculated using the following logistic function: ;(6) Then form a new vector: t l ={t l1 , t l2 , t l3 ,...t lj } T (7) Then use this vector as the previous layer h l+1 Input vector, where: l∈{1, 2, ..., c-1}; (4) Follow steps (1)-(3) and repeat several times, then set the BP neural network model in the last layer of the DBN, and use the output feature vector of the RBM as the input feature vector of the model. In each layer of the trained RBM network, the user can only check whether the weights in the layer are optimal for the feature vector mapping of that layer, rather than the feature vector mapping of the entire DBN. Finally, the back propagation network of the BP neural network will also propagate the error information from top to bottom to each layer of the RBM, and the entire DBN network needs to be fine-tuned. (5) In order to facilitate the recognition of face images, label information that can be recognized by users is set in the entire energy metering face image training set. By setting it in this way, when training the first layer of RBM, the explicit layer of RBM can have explicit neurons that can be trained together and neurons with classification labels. Finally, h l+1 The parameters of the layer {W l+1 ; α l+1 }, the final training result is: h l+1 =α l +(W 1 ) T t l (8)6) The entire DBN deep learning training is completed, forming a complete DBN deep learning algorithm network, and the parameter set of the DBN network is obtained as follows: {W 1 , W 2 , ..., W c ; α 1 , α 2 , ..., α c }; Then the training and integration process of the RBM network training model is equivalent to the initialization of the deep BP network weight parameters, which reduces the training time of the DBN deep learning algorithm model. When performing face image information recognition, the BP algorithm in the DBN deep learning algorithm model is started to perform a local search on the received weight parameter space to save convergence time; (S4) applying a super classifier to classify the data; the super classifier is a cascaded strong classifier; (S5) Determine whether the recognition requirements are met? If the recognition requirements are met, output the recognition image data; If the identification requirements are not met, return to step (S3) and perform data identification again; The steps of the feature layer fusion algorithm are as follows: (1) extracting test samples from the face image database; (2) performing big data dimensionality reduction processing using principal component analysis (PCA); (3) normalizing the features after the big data dimensionality reduction processing, and then determining the subspace so that the covariance matrix of each feature is a unit matrix; fusing face image information by connection and complementation; (4) feature matching, performing face image information matching by feature matching templates; (5) standardization score processing; wherein the standardization method is Min-Max standardization method, Z-score standardization method, Mean-MAD standardization method or Tanh standardization method; (6) outputting fusion scores, respectively summing up different output information to obtain image fusion values; The steps of the matching layer fusion algorithm are as follows: (1) extracting test samples from the face image database; (2) performing big data dimension reduction processing using principal component analysis (PCA); (3) standardizing features; wherein the standardization method is a Min-Max standardization method, a Z-score standardization method, a Mean-MAD standardization method, or a Tanh standardization method; (4) outputting a fusion score; (5) discriminating a subspace; (6) outputting feature matching; Haar classifier combines Haar-like features, integral graph method, Adaboost algorithm and cascade; The main steps of the Haar classifier are: (1) extracting Haar-like features; (2) using the integral graph method to accelerate the extraction of Haar-like features; (3) using the Adaboost algorithm to train a strong classifier to distinguish between faces and non-faces; (4) using a screening cascade to cascade the strong classifiers together to improve detection accuracy.

2. A three-dimensional face recognition method according to claim 1, Features: The number of layers of the DBN deep learning algorithm model is greater than 3.

3. A three-dimensional face recognition method according to claim 1, Features: The cascade strong classifier is a Haar classifier including an Adaboost algorithm.

4. A three-dimensional face recognition method according to claim 1, Features: The Haar-like feature is the sum of the pixel values ​​corresponding to the white area in the rectangular feature minus the sum of the pixel values ​​corresponding to the black area.

5. A three-dimensional face recognition method according to claim 4, Features: The working method of the AdaBoost classifier is: (1) Data sample selection: In the collected data, assume that the training set {(x 1 ,y 1 )、(x 2 ,y 2 ),…,(x n ,y n )}, in this training set, assuming X i is the input data training sample vector, y i It is the category mark classified by AdaBoost classifier; (2) Initialization: When initializing the data samples, all data training samples are given the same weight D = 1 / n, and then the AdaBoost classifier learning algorithm is used to perform N rounds of learning and training on the training sample set; where: D t (x i )=1 / n, i=1, 2,...,n; (3) Training weights: After training the weak classifier into a strong classifier, t Under this condition, the weak learning algorithm is trained to obtain the prediction function h t :X→{-1,+1}; (4) Output the predicted value and evaluate the error rate of the predicted value. The error calculation formula is: ε t =∑D t (x i )[h t (x i )≠y i ] Formula (1); (5) Determine whether the requirements are met. When h t (x i )≠y i When t (x i )=-1D t (x i )exp(-α t y i h t (x i )) Formula (2) When h t (x i )=y i When i h t (x i )=e a Formula (3) where D t The following formula is normalized to: ; (6) Output the final prediction function and obtain the normalized image data. After N rounds of training, the final prediction function is: 。 6. A three-dimensional face recognition method according to claim 1, Features: The cascade strong classifier is obtained by adding trained strong classifiers, and the number of the strong classifiers is greater than 5.