[go: up one dir, main page]

CN111126240A - A three-channel feature fusion face recognition method - Google Patents

A three-channel feature fusion face recognition method Download PDF

Info

Publication number
CN111126240A
CN111126240A CN201911317291.6A CN201911317291A CN111126240A CN 111126240 A CN111126240 A CN 111126240A CN 201911317291 A CN201911317291 A CN 201911317291A CN 111126240 A CN111126240 A CN 111126240A
Authority
CN
China
Prior art keywords
image
layer
channel
face
pooling
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911317291.6A
Other languages
Chinese (zh)
Other versions
CN111126240B (en
Inventor
李云红
聂梦瑄
李传真
穆兴
刘旭东
何琛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian Polytechnic University
Original Assignee
Xian Polytechnic University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xian Polytechnic University filed Critical Xian Polytechnic University
Priority to CN201911317291.6A priority Critical patent/CN111126240B/en
Publication of CN111126240A publication Critical patent/CN111126240A/en
Application granted granted Critical
Publication of CN111126240B publication Critical patent/CN111126240B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G06V40/171—Local features and components; Facial parts ; Occluding parts, e.g. glasses; Geometrical relationships
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Human Computer Interaction (AREA)
  • Evolutionary Biology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)
  • Collating Specific Patterns (AREA)

Abstract

本发明公开一种三通道特征融合人脸识别方法，具体步骤如下：步骤1，采集不同的人脸图像组成数据集；对数据集做预处理得到预处理图像集；步骤2，建立基于三通道特征融合的BP神经网络模型，包括粗采样通道20、LBP通道19、细采样通道18；步骤3，用预处理图像集对基于三通道特征融合的BP神经网络模型进行训练；步骤4，输入待识别的图像，利用训练好的基于三通道特征融合的BP神经网络模型进行特征相似度对比，输出最高相似度及其图像；步骤5，设定阈值，根据步骤4中输出的相似度与阈值作比较，判别与待识别的图像中是否为同一个人。本发明解决了传统人脸识别方法无法对人脸全方位特征进行提取的问题，提高了人脸的识别准确率。

Figure 201911317291

The invention discloses a three-channel feature fusion face recognition method. The specific steps are as follows: step 1, collecting different face images to form a data set; preprocessing the data set to obtain a preprocessed image set; step 2, establishing a three-channel based image set The BP neural network model of feature fusion includes coarse sampling channel 20, LBP channel 19, and fine sampling channel 18; step 3, use the preprocessing image set to train the BP neural network model based on three-channel feature fusion; step 4, input the For the identified images, use the trained BP neural network model based on three-channel feature fusion to compare the feature similarity, and output the highest similarity and its image; Step 5, set the threshold, according to the similarity and the threshold output in step 4. Compare and judge whether it is the same person in the image to be recognized. The invention solves the problem that the traditional face recognition method cannot extract all-round features of the face, and improves the recognition accuracy of the face.

Figure 201911317291

Description

Three-channel feature fusion face recognition method
Technical Field
The invention belongs to the technical field of face recognition, and discloses a three-channel feature fusion face recognition method.
Background
With the rapid development of society, people have great demands on automatic identity recognition devices in every aspect of daily life. The current identification technology mainly comprises: password authentication, fingerprint identification, face identification, iris identification, gait identification and the like. Because the face recognition has the advantages of non-contact property, high safety, convenience, rapidness and the like, the face recognition is gradually accepted by the public.
Especially in recent years, the computing speed is greatly improved as the computer hardware is continuously updated. Further, the convolutional neural network after several wave breaks is paid attention to people again, and more people are put into research and improvement on the face recognition algorithm. The traditional face recognition algorithm tests the recognition accuracy under a rigid condition and can obtain a good recognition effect. Under the non-rigid condition, the recognition accuracy is greatly reduced due to the influence of illumination, human face posture change, shielding and algorithm defects. The multi-task convolutional neural network appearing in recent years can accurately detect the position of a human face, mark key points and make full preparation for extracting the characteristics of the human face; the LBP operator can effectively reduce the influence of illumination change on the input face image. But how to effectively extract the face features is crucial to improving the face recognition accuracy. The method only preprocesses the input image, does not change the structure of the network, and can not effectively improve the identification accuracy; the two-channel model changes the structure of the neural network model. Although the improved model can effectively classify the input facial expressions, the classification accuracy is greatly influenced by non-rigid factors such as illumination change, shielding and the like of the input images; a traditional LBP operator and a Deep Belief Network (DBN) are combined with a feature extraction module, an image processed by the LBP is input into the DBN, and high identification accuracy is achieved on different data sets. However, most of the global feature information in the original image is lost in the image processed by the LBP operator, and the overall features of the input image cannot be effectively embodied.
Disclosure of Invention
The invention aims to provide a three-channel feature fusion face recognition method, which solves the problem that the traditional face recognition method in the prior art cannot extract all-round features of a face, and improves the face recognition accuracy.
The technical scheme adopted by the invention is that,
a three-channel feature fusion face recognition method comprises the following specific steps:
step 1, collecting different face images to form a data set; preprocessing each face image in the data set to obtain a preprocessed image, wherein the preprocessed image is the face image which is subjected to face correction after irrelevant background information is removed; forming all the preprocessed images into a preprocessed image set;
step 2, establishing a BP neural network model based on three-channel feature fusion, wherein the BP neural network model based on the three-channel feature fusion comprises three parallel feature extraction channels which are respectively a coarse sampling channel, an LBP channel and a fine sampling channel;
step 3, training a BP neural network model based on three-channel feature fusion by utilizing a preprocessed image set;
step 4, inputting an image to be recognized, performing feature similarity comparison on the image to be recognized and images in a training set by using a trained BP neural network model based on three-channel feature fusion, and outputting the image with the highest similarity and the similarity of the image;
and 5, setting a threshold, comparing the similarity output in the step 4 with the threshold, further judging whether the image output in the step 4 and the image to be identified are the same person, and outputting a result.
The present invention is also characterized in that,
the pretreatment operation in the step 1 comprises the following specific steps:
step 1.1, inputting a face image;
step 1.2, carrying out face cutting on the face image, removing redundant information such as background and the like, and obtaining a face image without a background;
and step 1.3, carrying out eye key point marking on the face image without the background, connecting the two eye key points, setting an included angle between a connecting line of the two eye key points and the horizontal direction as a, and rotating the face image anticlockwise by the angle a to obtain a preprocessed image.
In step 1.3, the coordinates of the key points are located by Euclidean distance, as shown in formula (1), wherein LiRepresenting the euclidean distance at which the key points are located,
Figure BDA0002326179930000031
representing predicted face keypoint locations, yiRepresenting the positions of key points of a real face;
Figure BDA0002326179930000032
the determination of the face key points is as in formula (2):
Figure BDA0002326179930000033
smaller Y indicates predicted keypoint location
Figure BDA0002326179930000034
With the true keypoint location yiThe smaller the error of (2), the smallest Y
Figure BDA0002326179930000035
The value is the location of the marked keypoint; where Y represents the location information of the final keypoint, N represents the number of training samples, biIndicating a sample label.
The angle a can be expressed as:
Figure BDA0002326179930000041
wherein (x)1，y1) Key point No. 1, the coordinates of the center of the left eye, (x)2，y2) Key point No. 2 is the coordinate of the center of the right eye.
The BP neural network model based on three-channel feature fusion comprises three parallel feature extraction channels, the three feature extraction channels are respectively a coarse sampling channel, an LBP channel and a fine sampling channel, the output ends of the coarse sampling channel, the LBP channel and the fine sampling channel are all connected with a hidden layer, and the output end of the hidden layer is sequentially connected with a dimensionality reduction layer, a first full-connection layer, a second full-connection layer and a loss function layer.
The coarse sampling channel consists of three convolution layers and three pooling layers, and sequentially comprises the following steps: a first convolution layer, a first pooling layer, a second convolution layer, a second pooling layer, a third convolution layer and a third pooling layer;
the first convolution layer, the second convolution layer and the third convolution layer adopt convolution kernels with the size of 5 multiplied by 5;
the first, second and third pooling layers all employ Max _ pooling, the pooling size is 2 × 2, the pooling step size is 2, and the padding mode is set to SAME.
The LBP channel feature extraction method comprises the following steps:
step 2.1, dividing the preprocessed image into a plurality of sub-images with the same size;
step 2.2, for each sub-image, converting each pixel point information into a pixel brightness value, and setting the pixel brightness value at the middle position as gcThe surrounding eight neighboring pixels have brightness values of gi(i ═ 0, 1.., 7), the luminance information is subjected to binarization processing using formula (8) and formula (9):
Figure BDA0002326179930000051
Figure BDA0002326179930000052
wherein x represents gi-gcDifference of (A), BiRepresenting a formula for converting the resulting binary number into a binary value, s (x) the resulting pixel value;
step 2.3, the binarized value on the right side of the pixel at the middle position is taken as an initial position, the obtained binarized value is written into an eight-bit binary number in a counterclockwise rotation mode, the binary number is converted into a decimal number, and the decimal number is the LBP value corresponding to the brightness of the pixel at the central point; and performing the operation on each pixel point in the input face image to finally obtain the local characteristic value of each pixel point of the input image.
The fine sampling channel consists of three convolution layers and three pooling layers, and sequentially comprises the following steps: a fourth convolution layer, a fourth pooling layer, a fifth convolution layer, a fifth pooling layer, a sixth convolution layer and a sixth pooling layer;
the fourth convolution layer, the fifth convolution layer and the sixth convolution layer are formed by stacking 1 x 3 convolution kernels and 3 x 1 convolution kernels;
the fourth, fifth and sixth pooling layers all employ Max _ pooling, the pooling size is 2 × 2, the pooling step size is 2, and the padding mode is set to SAME.
The dimensionality reduction layer adopts PCA dimensionality reduction operation to convert the feature fusion into one-dimensional feature vector information, and the method comprises the following specific steps:
step 3.1, performing feature fusion on the feature information of the human faces of the three channels to convert the feature information into a matrix form, and setting the matrix form of the fused feature image as X(n,m)；
Step 3.2, carrying out zero-mean processing on the matrix X and solving a covariance matrix H:
Figure BDA0002326179930000061
step 3.3, solving the eigenvalue of the matrix H and calculating the eigenvector corresponding to the eigenvalue;
step 3.4, taking the eigenvectors corresponding to the first k larger eigenvalues of the eigenvectors, and arranging the eigenvectors into a Q matrix according to rows;
and 3.5, obtaining the one-dimensional feature vector after dimension reduction by using the U-QX.
The loss function layer adopts a triple loss function as a loss function, and the formula is as follows:
Figure BDA0002326179930000062
Figure BDA0002326179930000063
Figure BDA0002326179930000064
a in the formula represents a sample arbitrarily selected from a training set; p represents a randomly selected sample of the same kind as a, and is called positive sample (positive); n represents a randomly selected sample of a different class, called negative sample (negative); training a parameter-sharing network aiming at each sample in the triple loss function to obtain the characteristic expressions of the three samples of a, p and n, which are respectively recorded as:
Figure BDA0002326179930000065
the triple loss function is to minimize the distance between (a, p) features in two classes and maximize the distance between (a, n) features between two classes through the learning of a training set; the effect of the '+' sign in the formula is: if 2]If the value of the inner is greater than 0, then the value is taken as the loss; when the loss is less than 0, the loss is 0.
The step 5 specifically comprises the following steps: and setting a threshold, if the similarity of the image output in the step 4 is greater than the threshold, determining that the image to be recognized and the human image in the image output in the step 4 are the same person, otherwise, determining that the training set does not have the same human face information as the image to be recognized, wherein the setting range of the threshold is 0.73-0.78.
The invention has the advantages that
(1) Firstly, MTCNN is used for carrying out face detection and face key point marking on an input image, background irrelevant information of the input image is removed, and on the basis, a rotary correction method is used for carrying out posture correction on a face, so that the recognition accuracy rate can be improved.
(2) The three channels are used for face feature extraction, so that local features, edge features and contour features of a face and fine features of face organs can be obtained, and the global feature information of the face image can be accurately reflected.
(3) Since the feature acquisition of the input image by using the three feature extraction channels leads to feature dimension explosion and is not beneficial to subsequent operation, the extracted features are subjected to dimension reduction operation by using a Principal Component Analysis (PCA).
(4) The invention selects triple loss as a loss function, and finally trains out the characteristic with cohesion. The triple loss function can distinguish the same and different features, reduce the inter-class distance as much as possible and enlarge the inter-class distance. Cohesion plays an important role in face recognition, and a very good model can be trained by using a small amount of data.
Drawings
FIG. 1 is a structural diagram of a BP neural network model volume based on three-channel feature fusion in a three-channel feature fusion face recognition method of the present invention;
FIG. 2 is a flow chart of a preprocessing of a three-channel feature fusion face recognition method of the present invention;
FIG. 3 is an LBP mapping chart in the three-channel feature fusion face recognition method of the present invention;
FIG. 4 is a graph of features extracted using different methods; wherein (a) is an original graph, (b) is a characteristic graph extracted by a Gabor method, (c) is a characteristic graph extracted by a Haar method, (d) is a characteristic graph extracted by an LBP method, and (e) is a characteristic graph extracted by the method;
FIG. 5 is another feature map extracted using a different method; wherein (a) is an original graph, (b) is a characteristic graph extracted by a Gabor method, (c) is a characteristic graph extracted by a Haar method, (d) is a characteristic graph extracted by an LBP method, and (e) is a characteristic graph extracted by the method;
fig. 6 is feature numbers extracted from Lena pictures by different methods, wherein (a) is a feature map extracted by a Gabor method, (b) is a feature map extracted by a Haar method, (c) is a feature map extracted by an LBP method, and (d) is a feature map extracted by the method of the present invention.
In the figure, 1, a first convolutional layer, 2, a first pooling layer, 3, a second convolutional layer, 4, a second pooling layer, 5, a third convolutional layer, 6, a third pooling layer, 7, a fourth convolutional layer, 8, a fourth pooling layer, 9, a fifth convolutional layer, 10, a fifth pooling layer, 11, a sixth convolutional layer, 12, a sixth pooling layer, 13, feature fusion, 14, a dimensionality reduction layer, 15, a first full-connection layer, 16, a second full-connection layer, 17, a loss function layer, 18, a fine sampling channel, 19, an LBP channel, and 20, a coarse sampling channel.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention discloses a three-channel feature fusion face recognition method, which comprises the following specific steps:
step 1, collecting different face images to form a data set; preprocessing each face image in the data set to obtain a preprocessed image, wherein the preprocessed image is the face image which is subjected to face correction after irrelevant background information is removed; forming all the preprocessed images into a preprocessed image set;
step 2, establishing a BP neural network model based on three-channel feature fusion, as shown in FIG. 1, wherein the BP neural network model based on three-channel feature fusion comprises three parallel feature extraction channels, and the three feature extraction channels are respectively a rough sampling channel 20, an LBP channel 19 and a fine sampling channel 18;
step 3, training a BP neural network model based on three-channel feature fusion by utilizing a preprocessed image set;
step 4, inputting an image to be recognized, performing feature similarity comparison on the image to be recognized and images in a training set by using a trained BP neural network model based on three-channel feature fusion, and outputting the image with the highest similarity and the similarity of the image;
and 5, setting a threshold, comparing the similarity output in the step 4 with the threshold, further judging whether the image output in the step 4 and the image to be identified are the same person, and outputting a result.
In step 1, as shown in fig. 2, the pretreatment operation specifically includes the following steps:
step 1.1, inputting a face image;
step 1.2, carrying out face cutting on the face image, removing redundant information such as background and the like, and obtaining a face image without a background;
and step 1.3, carrying out eye key point marking on the face image without the background, connecting the two eye key points, setting an included angle between a connecting line of the two eye key points and the horizontal direction as a, and rotating the face image anticlockwise by the angle a to obtain a preprocessed image.
In step 1.3, the coordinates of the key points are located by Euclidean distance, as shown in formula (1), wherein LiRepresenting the euclidean distance at which the key points are located,
Figure BDA0002326179930000091
representing predicted face keypoint locations, yiRepresenting the positions of key points of a real face;
Figure BDA0002326179930000092
the determination of the face key points is as in formula (2):
Figure BDA0002326179930000101
smaller Y indicates predicted keypoint location
Figure BDA0002326179930000102
With the true keypoint location yiThe smaller the error of (2), the smallest Y
Figure BDA0002326179930000103
The value is the location of the marked keypoint; where Y represents the location information of the final keypoint, N represents the number of training samples, biIndicating a sample label.
The angle a can be expressed as:
Figure BDA0002326179930000104
wherein (x)1，y1) Key point No. 1, the coordinates of the center of the left eye, (x)2，y2) Is No. 2 key point, right eyeThe coordinates of the center.
The BP neural network model based on three-channel feature fusion comprises three parallel feature extraction channels, the three feature extraction channels are respectively a coarse sampling channel 20, an LBP channel 19 and a fine sampling channel 18, the output ends of the coarse sampling channel 20, the LBP channel 19 and the fine sampling channel 18 are all connected with a hidden layer 13, and the output end of the hidden layer 13 is sequentially connected with a dimensionality reduction layer 14, a first full connection layer 15, a second full connection layer 16 and a loss function layer 17.
Wherein the rough sampling channel 20 is used for collecting rough edge and contour features in the preprocessed image;
the LBP channel 19 is used for acquiring a local characteristic map of a local input image of the preprocessed image;
wherein the fine sampling channel 18 is used for acquiring rough edge and contour features in the preprocessed image;
the hidden layer 13 is used for fusing and converting the features extracted by the three sampling channels into a matrix form.
The rough sampling channel 20 is composed of three convolutional layers and three pooling layers, and sequentially comprises: a first convolutional layer 1, a first pooling layer 2, a second convolutional layer 3, a second pooling layer 4, a third convolutional layer 5, and a third pooling layer 6;
the first convolution layer 1, the second convolution layer 3 and the third convolution layer 5 all adopt convolution kernels with the size of 5 multiplied by 5;
the first, second and third pooling layers 2, 4 and 6 all employ Max _ pooling, the pooling size is 2 × 2, the pooling step size is 2, and the padding mode is set to SAME.
The formula for extracting features of an image by using the coarse sampling channel 20 is as follows:
let the input image size be H1×H2Convolution kernel size of F1×F2Output image size W1×W2. The convolutional layer formula is shown in formula (4):
W1×W2＝(H1-F1+1)*(H2-F2+1) (4)
the picture size output by the convolution layer is W1×W2Then inputting the mixture into a pooling layer with a pooling size of F3×F4The step length of the pooling is S; when padding is in Valid mode, the calculation formula of the image size after passing through the pooling layer is shown as formula (5):
Figure BDA0002326179930000111
and when padding is in the Same mode, calculating the size of the image after passing through the pooling layer. FiFor the size of the pooling layer, P is the size of the fill around the image, as shown in equation (6):
Figure BDA0002326179930000112
Figure BDA0002326179930000113
as shown in fig. 3, the LBP channel performs the feature extraction method as follows:
step 2.1, dividing the preprocessed image into a plurality of sub-images with the same size;
step 2.2, for each sub-image, converting each pixel point information into a pixel brightness value, and setting the pixel brightness value at the middle position as gcThe surrounding eight neighboring pixels have brightness values of gi(i ═ 0, 1.., 7), the luminance information is subjected to binarization processing using formula (8) and formula (9):
Figure BDA0002326179930000121
Figure BDA0002326179930000122
wherein x represents gi-gcDifference of (A), BiRepresenting a formula for converting the resulting binary number into a binary value, s (x) the resulting pixel value;
step 2.3, the binarized value on the right side of the pixel at the middle position is taken as an initial position, the obtained binarized value is written into an eight-bit binary number in a counterclockwise rotation mode, the binary number is converted into a decimal number, and the decimal number is the LBP value corresponding to the brightness of the pixel at the central point; and performing the operation on each pixel point in the input face image to finally obtain the local characteristic value of each pixel point of the input image.
The fine sampling channel 18 is composed of three convolutional layers and three pooling layers, and sequentially comprises: a fourth convolutional layer (7), a fourth pooling layer (8), a fifth convolutional layer (9), a fifth pooling layer (10), a sixth convolutional layer (11), and a sixth pooling layer (12);
the fourth convolution layer (7), the fifth convolution layer (9) and the sixth convolution layer (11) are convolution layers formed by stacking 1 x 3 convolution kernels and 3 x 1 convolution kernels;
the fourth pooling layer (8), the fifth pooling layer (10) and the sixth pooling layer (12) all adopt maximum pooling (Max _ pooling), the pooling size is 2 x 2, the pooling step size is 2, and the padding mode is set to SAME.
The feature extraction using the fine sampling channel 18 is the same as the feature extraction using the coarse sampling channel 20.
The dimensionality reduction layer 14 transforms the feature fusion into one-dimensional feature vector information by adopting PCA dimensionality reduction operation, and the method specifically comprises the following steps:
step 3.1, performing feature fusion on the feature information of the human faces of the three channels to convert the feature information into a matrix form, and setting the matrix form of the fused feature image as X(n,m)；
Step 3.2, carrying out zero-mean processing on the matrix X and solving a covariance matrix H:
Figure BDA0002326179930000131
step 3.3, solving the eigenvalue of the matrix H and calculating the eigenvector corresponding to the eigenvalue;
step 3.4, taking the eigenvectors corresponding to the first k larger eigenvalues of the eigenvectors, and arranging the eigenvectors into a Q matrix according to rows;
and 3.5, obtaining the one-dimensional feature vector after dimension reduction by using the U-QX.
The loss function layer (17) adopts a triple loss function as a loss function, and the formula is as follows:
Figure BDA0002326179930000132
Figure BDA0002326179930000133
Figure BDA0002326179930000134
a in the formula represents a sample arbitrarily selected from a training set; p represents a randomly selected sample of the same kind as a, and is called positive sample (positive); n represents a randomly selected sample of a different class, called negative sample (negative); training a parameter-sharing network aiming at each sample in the triple loss function to obtain the characteristic expressions of the three samples of a, p and n, which are respectively recorded as:
Figure BDA0002326179930000135
the triple loss function is to minimize the distance between (a, p) features in two classes and maximize the distance between (a, n) features between two classes through the learning of a training set; the effect of the '+' sign in the formula is: if 2]If the value of the inner is greater than 0, then the value is taken as the loss; when the loss is less than 0, the loss is 0.
Specifically, step 5 is to set a threshold, if the similarity of the image output in step 4 is greater than the threshold, the image to be recognized and the portrait in the image output in step 4 are determined to be the same person, otherwise, the training set is determined to have no facial information the same as that of the image to be recognized, and the set range of the threshold is 0.73-0.78.
In order to verify the invention, five algorithms (CNN, LBP + DBN, rough learning + fine learning, HOG + SVM and the invention algorithm) are adopted in the ORL face database to carry out face recognition simulation experiments. The Gabor, Haar, LBP and three-channel fused profiles were used herein for comparison, as shown in fig. 4. The Gabor and LBP algorithm can only extract the local features of the face image; the Haar algorithm value can extract contour characteristic information of the human face; the extracted local and global features of the human face are fused by the method to form output including the local and global features, so that the identification accuracy can be effectively improved. Fig. 5 shows feature numbers extracted from Lena pictures by four methods, in which the abscissa and ordinate represent the number of features extracted from the length and width of an input image, respectively, and the height represents the feature intensity. By comparing the four feature extraction graphs in fig. 5, the image features after three-channel feature fusion can be obtained, more feature information can be extracted, and the feature intensity is higher than that obtained by other methods, so that the identification accuracy is effectively improved.
TABLE 1 comparison of time consumption and accuracy for identification in ORL database。The ORL face database contains 40 subjects of different ages, gender, and ethnicity. 10 pictures per person for a total of 400 pictures. The original picture size is 92 × 112. These pictures were collected under differential expression, pose and lighting. The original pictures were processed to 64 × 64 in the experiment for the experiment. In the experiment, 200 training samples are selected, and the remaining 200 testing samples are selected.
TABLE 1 comparison of recognition time consumption and accuracy in ORL database
Figure BDA0002326179930000151
As can be clearly seen from table 1, the performance of the algorithm proposed by the present invention is compared with that of the comparison algorithm on the ORL database. Compared with the traditional CNN algorithm, the algorithm provided by the invention almost has no difference with the comparison algorithm in terms of face recognition time consumption, and the average time consumption is only 0.0135 second more compared with four comparison algorithms of CNN, LBP + DBN, rough learning + fine learning and HOG + SVM. But there is a significant improvement in recognition accuracy. The average accuracy is higher by 5.38%.

Claims (10)

1.一种三通道特征融合人脸识别方法，其特征在于，具体步骤如下：1. a three-channel feature fusion face recognition method, is characterized in that, concrete steps are as follows: 步骤1，采集不同的人脸图像组成数据集；对所述数据集中的每张人脸图像做预处理操作得到预处理图像，所述预处理图像为剔除无关背景信息、进行人脸矫正后的人脸图像；将所有预处理图像组成预处理图像集；Step 1: Collect different face images to form a data set; perform a preprocessing operation on each face image in the data set to obtain a preprocessed image, and the preprocessed image is obtained after removing irrelevant background information and performing face correction. face image; group all preprocessed images into a preprocessed image set; 步骤2，建立基于三通道特征融合的BP神经网络模型，所述基于三通道特征融合的BP神经网络模型包括并行的三个特征提取通道，所述三个特征提取通道分别为粗采样通道(20)、LBP通道(19)、细采样通道(18)；Step 2, establish a BP neural network model based on three-channel feature fusion, the BP neural network model based on three-channel feature fusion includes three parallel feature extraction channels, and the three feature extraction channels are respectively rough sampling channels (20 ), LBP channel (19), fine sampling channel (18); 步骤3，利用所述预处理图像集对基于三通道特征融合的BP神经网络模型进行训练；Step 3, using the preprocessing image set to train the BP neural network model based on three-channel feature fusion; 步骤4，输入待识别的图像，利用训练好的基于三通道特征融合的BP神经网络模型对待识别的图像和训练集中的图像进行特征相似度对比，输出相似度最高的图像和该图像的相似度；Step 4: Input the image to be identified, use the trained BP neural network model based on three-channel feature fusion to compare the feature similarity between the image to be identified and the image in the training set, and output the image with the highest similarity and the similarity of the image. ; 步骤5，设定阈值，根据所述步骤4中输出的相似度与阈值作比较，进而判别步骤4输出图像与待识别的图像中是否为同一个人，输出结果。Step 5: Set a threshold, compare the similarity output in step 4 with the threshold, and then determine whether the image output in step 4 and the image to be recognized are the same person, and output the result. 2.如权利要求1所述的一种三通道特征融合人脸识别方法，其特征在于，步骤1中所述预处理操作的具体步骤为：2. a kind of three-channel feature fusion face recognition method as claimed in claim 1 is characterized in that, the concrete steps of the preprocessing operation described in step 1 are: 步骤1.1，输入人脸图像；Step 1.1, input face image; 步骤1.2，对所述人脸图像进行人脸裁剪，除去背景等冗余信息，得到无背景的人脸图像；Step 1.2, performing face cropping on the face image, removing redundant information such as background, and obtaining a face image without background; 步骤1.3，对所述无背景的人脸图像进行眼部关键点标记，把两个所述眼部关键点连接，其连线与水平方向的夹角设为a，再将所述人脸图像逆时针旋转a得到预处理图像。Step 1.3, mark the eye key points on the face image without background, connect the two eye key points, and set the angle between the connection line and the horizontal direction as a, and then use the face image Rotate a counterclockwise to get the preprocessed image. 3.如权利要求2所述的一种三通道特征融合人脸识别方法，其特征在于，所述步骤1.3中，采用欧式距离定位关键点坐标，如公式(1)所示，其中，Li表示关键点定位的欧氏距离，
Figure FDA0002326179920000021
表示预测人脸关键点位置，yi表示真实人脸关键点位置；3. a kind of three-channel feature fusion face recognition method as claimed in claim 2, is characterized in that, in described step 1.3, adopts Euclidean distance to locate key point coordinates, as shown in formula (1), wherein, L i represents the Euclidean distance for keypoint positioning,
Figure FDA0002326179920000021
Represents the predicted face key point position, and y i represents the real face key point position;
Figure FDA0002326179920000022
Figure FDA0002326179920000022
 人脸关键点的确定如公式(2)：The key points of the face are determined as formula (2):
Figure FDA0002326179920000023
Figure FDA0002326179920000023
 Y越小表明预测关键点位置
Figure FDA0002326179920000024
与真实关键点位置yi的误差越小，使得Y最小的
Figure FDA0002326179920000025
值便是标记出的关键点的位置；其中，Y代表最终关键点的位置信息，N表示训练样本的数量，bi表示样本标签。The smaller Y indicates the predicted key point location
Figure FDA0002326179920000024
The smaller the error with the real key point position yi is, the smaller the value of Y is.
Figure FDA0002326179920000025
The value is the position of the marked key point; among them, Y represents the location information of the final key point, N represents the number of training samples, and b i represents the sample label. 所述夹角a可以表示为：The included angle a can be expressed as:
Figure FDA0002326179920000026
Figure FDA0002326179920000026
 其中(x1，y1)为1号关键点，是左眼中心的坐标，(x2，y2)为2号关键点，是右眼中心的坐标。Among them, (x 1 , y 1 ) is the key point No. 1, which is the coordinate of the center of the left eye, and (x 2 , y 2 ) is the key point No. 2, which is the coordinate of the center of the right eye. 4.如权利要求1所述的一种三通道特征融合人脸识别方法，其特征在于，所述基于三通道特征融合的BP神经网络模型包括并行的三个特征提取通道，所述三个特征提取通道分别为粗采样通道(20)、LBP通道(19)、细采样通道(18)，所述粗采样通道(20)、LBP通道(19)和细采样通道(18)的输出端均与隐藏层(13)连接，所述隐藏层(13)的输出端依次连接有降维层(14)、第一全连接层(15)、第二全连接层(16)和损失函数层(17)。4. a three-channel feature fusion face recognition method as claimed in claim 1, is characterized in that, the described BP neural network model based on three-channel feature fusion comprises three parallel feature extraction channels, and described three features The extraction channels are respectively a coarse sampling channel (20), an LBP channel (19), and a fine sampling channel (18). The hidden layer (13) is connected, and the output end of the hidden layer (13) is sequentially connected with a dimensionality reduction layer (14), a first fully connected layer (15), a second fully connected layer (16) and a loss function layer (17) ). 5.如权利要求4所述的一种三通道特征融合人脸识别方法，其特征在于，所述粗采样通道(20)由三层卷积层和三层池化层组成，依次为：第一卷积层(1)、第一池化层(2)、第二卷积层(3)、第二池化层(4)、第三卷积层(5)和第三池化层(6)；5. a three-channel feature fusion face recognition method as claimed in claim 4, is characterized in that, described rough sampling channel (20) is made up of three layers of convolution layers and three layers of pooling layers, in order: the first A convolutional layer (1), a first pooling layer (2), a second convolutional layer (3), a second pooling layer (4), a third convolutional layer (5), and a third pooling layer ( 6); 所述第一卷积层(1)、第二卷积层(3)和第三卷积层(5)均采用5×5大小的卷积核；The first convolutional layer (1), the second convolutional layer (3) and the third convolutional layer (5) all use a 5×5 convolution kernel; 所述第一池化层(2)、第二池化层(4)和第三池化层(6)均采用最大池化，池化尺寸为2×2，池化步长为2，padding模式设置为SAME。The first pooling layer (2), the second pooling layer (4) and the third pooling layer (6) all use maximum pooling, the pooling size is 2×2, the pooling step size is 2, and the padding Mode is set to SAME. 6.如权利要求4所述的一种三通道特征融合人脸识别方法，其特征在于，LBP通道进行特征提取的方法为：6. a kind of three-channel feature fusion face recognition method as claimed in claim 4, is characterized in that, the method that LBP channel carries out feature extraction is: 步骤2.1，把预处理图像分割成大小相同的若干个子图；Step 2.1, divide the preprocessed image into several sub-images of the same size; 步骤2.2，对于每个子图，将其中每个像素点信息转化为像素亮度值，设中间位置像素亮度值为gc，周围的八近邻像素亮度值为gi(i＝0,1,...,7)，利用公式(8)和公式(9)对亮度信息进行二值化处理：Step 2.2, for each sub-image, convert the information of each pixel into a pixel brightness value, set the pixel brightness value in the middle position to g c , and the eight neighboring pixels around the brightness value g i (i=0, 1, . . . ., 7), using formula (8) and formula (9) to binarize the luminance information:
Figure FDA0002326179920000031
Figure FDA0002326179920000031
 其中x表示gi-gc的差值，Bi表示把所得的二进制数转化为二进制数值的公式，s(x)最终得到的像素值；Where x represents the difference between g i -g c , B i represents the formula for converting the obtained binary number into a binary value, and the pixel value finally obtained by s(x); 步骤2.3，以中间位置像素右侧二值化的值为起始位置，逆时针旋转把所得二值化的值写为八位的二进制数，将二进制数转化为十进制数，十进制数就是中心点像素亮度对应的LBP值；对输入人脸图像中每个像素点进行此操作，最终得到输入图像每个像素点的局部特征值。Step 2.3, take the binarized value on the right side of the pixel at the middle position as the starting position, rotate counterclockwise to write the obtained binarized value as an eight-bit binary number, convert the binary number into a decimal number, and the decimal number is the center point The LBP value corresponding to the pixel brightness; perform this operation on each pixel in the input face image, and finally obtain the local feature value of each pixel in the input image. 7.如权利要求4所述的一种三通道特征融合人脸识别方法，其特征在于，所述细采样通道(18)由三层卷积层和三层池化层组成，依次为：第四卷积层(7)、第四池化层(8)、第五卷积层(9)、第五池化层(10)、第六卷积层(11)和第六池化层(12)；7. A three-channel feature fusion face recognition method as claimed in claim 4, wherein the fine sampling channel (18) is composed of three layers of convolution layers and three layers of pooling layers, in order: the first Four convolutional layers (7), fourth pooling layer (8), fifth convolutional layer (9), fifth pooling layer (10), sixth convolutional layer (11) and sixth pooling layer ( 12); 所述第四卷积层(7)、第五卷积层(9)和第六卷积层(11)均采用1×3卷积核与3×1卷积核堆叠形成的卷积层；The fourth convolution layer (7), the fifth convolution layer (9) and the sixth convolution layer (11) all use convolution layers formed by stacking 1×3 convolution kernels and 3×1 convolution kernels; 所述第四池化层(8)、第五池化层(10)和第六池化层(12)均采用最大池化，池化尺寸为2×2，池化步长为2，padding模式设置为SAME。The fourth pooling layer (8), the fifth pooling layer (10) and the sixth pooling layer (12) all adopt maximum pooling, the pooling size is 2×2, the pooling step size is 2, and the padding Mode is set to SAME. 8.如权利要求4所述的一种三通道特征融合人脸识别方法，其特征在于，所述降维层(14)采用PCA降维运算将特征融合后转化为一维特征向量信息，具体的步骤为：8. a kind of three-channel feature fusion face recognition method as claimed in claim 4, is characterized in that, described dimension reduction layer (14) adopts PCA dimension reduction operation to convert into one-dimensional feature vector information after feature fusion, concrete The steps are: 步骤3.1，将三个通道的人脸的特征信息进行特征融合转化为矩阵形式，设融合后的特征图像的矩阵形式为X(n,m)；Step 3.1, the feature information of the faces of the three channels is carried out feature fusion and transformed into a matrix form, and the matrix form of the feature image after the fusion is assumed to be X (n, m) ; 步骤3.2，对矩阵X做零均值化处理并求出其协方差矩阵H：Step 3.2, perform zero-average processing on the matrix X and obtain its covariance matrix H:
Figure FDA0002326179920000051
Figure FDA0002326179920000051
 步骤3.3，求矩阵H的特征值，并计算特征值对应的特征向量；Step 3.3, find the eigenvalues of the matrix H, and calculate the eigenvectors corresponding to the eigenvalues; 步骤3.4，取特征向量的前k个较大特征值对应的特征向量，并按行排列成Q矩阵；Step 3.4, take the eigenvectors corresponding to the first k larger eigenvalues of the eigenvectors, and arrange them into a Q matrix in rows; 步骤3.5，U＝QX即为降维后的一维特征向量。Step 3.5, U=QX is the one-dimensional feature vector after dimension reduction. 9.如权利要求4所述的一种三通道特征融合人脸识别方法，其特征在于，所述损失函数层(17)采用Triplet loss函数作为损失函数，公式如下：9. a kind of three-channel feature fusion face recognition method as claimed in claim 4, is characterized in that, described loss function layer (17) adopts Triplet loss function as loss function, and formula is as follows:
Figure FDA0002326179920000052
Figure FDA0002326179920000052
Figure FDA0002326179920000053
Figure FDA0002326179920000053
Figure FDA0002326179920000054
Figure FDA0002326179920000054
 公式中a代表从训练集中任意选取的一个样本；p代表随机选取和a同类的一个样本，称为正样本(positive)；n代表随机选取和a不同类的一个样本，称为负样本(negative)；针对Triplet loss函数中每个样本，训练一个参数共享的网络，得到a，p，n三个样本的特征表达分别记为：
Figure FDA0002326179920000055
Triplet loss函数就是要通过训练集的学习，使得(a,p)两个类内特征之间的距离最小化，(a,n)两个类间特征之间的距离最大化；公式中‘+’号的作用是：如果[]内的值大于0时，那么取该值作为损失；小于0时，损失取0。In the formula, a represents a sample randomly selected from the training set; p represents a randomly selected sample of the same class as a, which is called a positive sample; n represents a randomly selected sample of a different class from a, which is called a negative sample. ); for each sample in the Triplet loss function, a network with shared parameters is trained, and the feature expressions of the three samples a, p, and n are respectively recorded as:
Figure FDA0002326179920000055
The Triplet loss function is to learn from the training set to minimize the distance between (a, p) two intra-class features, and (a, n) to maximize the distance between two inter-class features; in the formula '+ The function of the ' number is: if the value in [] is greater than 0, the value is taken as the loss; when it is less than 0, the loss is taken as 0. 10.如权利要求4所述的一种三通道特征融合人脸识别方法，其特征在于，步骤5具体的为：设定阈值，若所述步骤4输出图像的相似度大于阈值，则认定待识别的图像与步骤4输出图像中的人像为同一个人，否则认定训练集中不存在与待识别的图像相同的人脸信息，其中阈值的设定范围在0.73～0.78。10. a kind of three-channel feature fusion face recognition method as claimed in claim 4, it is characterized in that, step 5 is specifically: set threshold value, if the similarity of described step 4 output image is greater than threshold value, it is determined that waiting The recognized image is the same person as the portrait in the output image in step 4, otherwise it is determined that the training set does not have the same face information as the image to be recognized, and the threshold is set in the range of 0.73 to 0.78.
CN201911317291.6A 2019-12-19 2019-12-19 Three-channel feature fusion face recognition method Active CN111126240B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911317291.6A CN111126240B (en) 2019-12-19 2019-12-19 Three-channel feature fusion face recognition method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911317291.6A CN111126240B (en) 2019-12-19 2019-12-19 Three-channel feature fusion face recognition method

Publications (2)

Publication Number Publication Date
CN111126240A true CN111126240A (en) 2020-05-08
CN111126240B CN111126240B (en) 2023-04-07

Family

ID=70500921

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911317291.6A Active CN111126240B (en) 2019-12-19 2019-12-19 Three-channel feature fusion face recognition method

Country Status (1)

Country Link
CN (1) CN111126240B (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111696090A (en) * 2020-06-08 2020-09-22 电子科技大学 Method for evaluating quality of face image in unconstrained environment
CN111723709A (en) * 2020-06-09 2020-09-29 大连海事大学 A fly face recognition method based on deep convolutional neural network
CN112364828A (en) * 2020-11-30 2021-02-12 姜召英 Face recognition method and financial system
CN112800914A (en) * 2021-01-20 2021-05-14 桂林电子科技大学 Face recognition method based on expert prior knowledge and LTP complex illumination
CN113269022A (en) * 2021-03-24 2021-08-17 七台河市公安局 System and method for predicting recessive specific personnel
CN113361495A (en) * 2021-07-27 2021-09-07 平安科技（深圳）有限公司 Face image similarity calculation method, device, equipment and storage medium
CN117437522A (en) * 2023-12-19 2024-01-23 福建拓尔通软件有限公司 Face recognition model training method, face recognition method and device
CN118135622A (en) * 2023-10-23 2024-06-04 来邦科技股份公司 Face recognition-based non-contact body temperature measurement method and system
CN118692155A (en) * 2024-08-26 2024-09-24 浙江大华技术股份有限公司 Image detection method and device

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105069400A (en) * 2015-07-16 2015-11-18 北京工业大学 Face image gender recognition system based on stack type sparse self-coding
WO2018107760A1 (en) * 2016-12-16 2018-06-21 北京大学深圳研究生院 Collaborative deep network model method for pedestrian detection
CN108629338A (en) * 2018-06-14 2018-10-09 五邑大学 A kind of face beauty prediction technique based on LBP and convolutional neural networks
WO2019128367A1 (en) * 2017-12-26 2019-07-04 广州广电运通金融电子股份有限公司 Face verification method and apparatus based on triplet loss, and computer device and storage medium

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105069400A (en) * 2015-07-16 2015-11-18 北京工业大学 Face image gender recognition system based on stack type sparse self-coding
WO2018107760A1 (en) * 2016-12-16 2018-06-21 北京大学深圳研究生院 Collaborative deep network model method for pedestrian detection
WO2019128367A1 (en) * 2017-12-26 2019-07-04 广州广电运通金融电子股份有限公司 Face verification method and apparatus based on triplet loss, and computer device and storage medium
CN108629338A (en) * 2018-06-14 2018-10-09 五邑大学 A kind of face beauty prediction technique based on LBP and convolutional neural networks

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
徐佳等: "基于鱼眼摄像头的一种人脸识别技术", 《信息通信》 *
李佳妮等: "特征匹配融合结合改进卷积神经网络的人脸识别", 《激光与光电子学进展》 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111696090A (en) * 2020-06-08 2020-09-22 电子科技大学 Method for evaluating quality of face image in unconstrained environment
CN111696090B (en) * 2020-06-08 2022-07-29 电子科技大学 A face image quality assessment method in an unconstrained environment
CN111723709A (en) * 2020-06-09 2020-09-29 大连海事大学 A fly face recognition method based on deep convolutional neural network
CN112364828B (en) * 2020-11-30 2022-01-04 天津金城银行股份有限公司 Face recognition method and financial system
CN112364828A (en) * 2020-11-30 2021-02-12 姜召英 Face recognition method and financial system
CN112800914A (en) * 2021-01-20 2021-05-14 桂林电子科技大学 Face recognition method based on expert prior knowledge and LTP complex illumination
CN113269022A (en) * 2021-03-24 2021-08-17 七台河市公安局 System and method for predicting recessive specific personnel
CN113361495A (en) * 2021-07-27 2021-09-07 平安科技（深圳）有限公司 Face image similarity calculation method, device, equipment and storage medium
CN113361495B (en) * 2021-07-27 2024-04-09 平安科技（深圳）有限公司 Method, device, equipment and storage medium for calculating similarity of face images
CN118135622A (en) * 2023-10-23 2024-06-04 来邦科技股份公司 Face recognition-based non-contact body temperature measurement method and system
CN117437522A (en) * 2023-12-19 2024-01-23 福建拓尔通软件有限公司 Face recognition model training method, face recognition method and device
CN117437522B (en) * 2023-12-19 2024-05-03 福建拓尔通软件有限公司 Face recognition model training method, face recognition method and device
CN118692155A (en) * 2024-08-26 2024-09-24 浙江大华技术股份有限公司 Image detection method and device

Also Published As

Publication number Publication date
CN111126240B (en) 2023-04-07

Similar Documents

Publication Publication Date Title
CN111126240B (en) Three-channel feature fusion face recognition method
CN101763503B (en) Face recognition method of attitude robust
CN110348319A (en) A kind of face method for anti-counterfeit merged based on face depth information and edge image
CN102194108B (en) Smile face expression recognition method based on clustering linear discriminant analysis of feature selection
CN100395770C (en) A Hand Feature Fusion Authentication Method Based on Feature Relationship Measurement
CN108009472B (en) Finger back joint print recognition method based on convolutional neural network and Bayes classifier
CN108520226A (en) A Pedestrian Re-Identification Method Based on Body Decomposition and Saliency Detection
CN103839042B (en) Face identification method and face identification system
CN102682287A (en) Pedestrian detection method based on saliency information
CN105574509B (en) A kind of face identification system replay attack detection method and application based on illumination
CN105205480A (en) Complex scene human eye locating method and system
CN104951940A (en) Mobile payment verification method based on palmprint recognition
CN103440035A (en) Gesture recognition system in three-dimensional space and recognition method thereof
CN112101208A (en) Feature series fusion gesture recognition method and device for elderly people
CN110826408A (en) A face recognition method for regional feature extraction
CN112132117A (en) A Converged Identity Authentication System Assisting Coercion Detection
CN113486712A (en) Multi-face recognition method, system and medium based on deep learning
CN111832405A (en) A face recognition method based on HOG and deep residual network
CN111967363A (en) Emotion prediction method based on micro-expression recognition and eye movement tracking
CN110598574A (en) Intelligent face monitoring and identifying method and system
CN113313149B (en) A Dishes Recognition Method Based on Attention Mechanism and Metric Learning
CN111274883A (en) A synthetic sketch face recognition method based on multi-scale HOG features and deep features
CN105760815A (en) Heterogeneous human face verification method based on portrait on second-generation identity card and video portrait
CN107392105A (en) A kind of expression recognition method based on reverse collaboration marking area feature
CN110458064B (en) Combining data-driven and knowledge-driven low-altitude target detection and recognition methods

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant