[go: up one dir, main page]

CN111339810A - Low-resolution large-angle face recognition method based on Gaussian distribution - Google Patents

Low-resolution large-angle face recognition method based on Gaussian distribution Download PDF

Info

Publication number
CN111339810A
CN111339810A CN201910337099.7A CN201910337099A CN111339810A CN 111339810 A CN111339810 A CN 111339810A CN 201910337099 A CN201910337099 A CN 201910337099A CN 111339810 A CN111339810 A CN 111339810A
Authority
CN
China
Prior art keywords
model
face
training
face recognition
picture
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.)
Pending
Application number
CN201910337099.7A
Other languages
Chinese (zh)
Inventor
施海波
孙勇
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.)
Nanjing Teworth High Tech Co ltd
Original Assignee
Nanjing Teworth High Tech Co ltd
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 Nanjing Teworth High Tech Co ltd filed Critical Nanjing Teworth High Tech Co ltd
Priority to CN201910337099.7A priority Critical patent/CN111339810A/en
Publication of CN111339810A publication Critical patent/CN111339810A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06V40/161—Detection; Localisation; Normalisation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06V40/172—Classification, e.g. identification

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to the technical field of intelligent security face recognition, and discloses a low-resolution large-angle face recognition method based on Gaussian distribution, which comprises the following specific steps of: step 1: reading a face data set used for training a face recognition model; step 2: carrying out face detection and cutting on pictures in the data, and screening out training pictures meeting conditions; step 3: sequentially reading training pictures in batches, and performing data enhancement processing such as random rotation and zooming; step 4: training a face recognition model by using a softmax function based on Gaussian distribution as a loss function; step 5; judging whether the face recognition accuracy meets the design requirement or not; if the training data are obtained, the final trained face recognition model is output, otherwise, the training data are continuously read to train the model. The low-resolution large-angle face recognition method based on Gaussian distribution improves the accuracy of the face recognition model on face angles and resolution, and enhances the robustness of the model.

Description

Low-resolution large-angle face recognition method based on Gaussian distribution
Technical Field
The invention relates to the technical field of intelligent security face recognition, in particular to a low-resolution large-angle face recognition method based on Gaussian distribution.
Background
With the continuous improvement of security protection standards of crowded areas such as airports, subways, shopping centers and the like, machine vision-based intelligent monitoring systems are receiving more and more attention. In order to obtain a wider visual field of most monitoring videos, the face collected generally is smaller and not at a direct viewing angle, and compared with a clear and high-resolution front face picture, the discrimination and the information content of a small-scale large-angle face picture are greatly reduced, so that a monitoring video face recognition system needs to be specially optimized for the situations of small scale and large angle.
At present, algorithms related to face recognition have been studied for many years, but algorithms specific to monitoring videos are rare, and only some algorithms only consider the problems of pose and resolution respectively, but do not consider the two problems together, and do not pay attention to the real-time problem of the algorithms. Surveillance video face recognition is more challenging than still picture face recognition.
The existing face recognition model training methods can be generally divided into two types: metric learning-based methods and boundary-based classification methods. The metric learning method calculates a Loss function directly on the features extracted from the human face, so that the same human feature is closer to the features extracted from the human face, and different human features are farther from the features extracted from the human face, such as deep id2 using contrast Loss (contrast Loss) and FaceNet using triple Loss (triple Loss). The metric learning based on the comparison loss function and the triple loss function accords with the cognitive rules of people, and a good effect is achieved in practical application, but the metric learning has two fatal problems, namely, the model can be fitted for a long time, and the model depends on the sampling mode of training data, so that the ideal sampling mode not only can improve the final performance of the algorithm, but also can accelerate the training speed.
The boundary classification method is not like a metric learning method which directly calculates a loss function in a characteristic layer, but the face recognition is still used as a classification task for training, and the limitation of applying a boundary to the characteristic layer is indirectly realized through the modification of a softmax formula, so that the finally obtained characteristics of the network have higher discrimination. The idea is proposed from Sphereface, and the AM-softmax/CosFace and ArcFace are improved to obtain the best performance on the existing public data set at present, and the training speed is higher. However, the method does not consider the problem of unbalanced distribution of sample data, so that the algorithm performs well on a high-definition front face picture and has low recognition accuracy on a low-resolution side face picture, and therefore improvement is needed for the problem.
Disclosure of Invention
Technical problem to be solved
Aiming at the defects of the prior art, the invention provides a low-resolution large-angle face recognition method based on Gaussian distribution, which improves the accuracy of a face recognition model on face angles and resolution, enhances the robustness of the model and the like, and solves the problem that the recognition accuracy of the existing face recognition model training method on a low-resolution side face picture is low.
(II) technical scheme
In order to achieve the purpose, the invention provides the following technical scheme: a low-resolution large-angle face recognition method based on Gaussian distribution comprises the following specific steps:
step 1: reading a face data set used for training a face recognition model;
step 2: carrying out face detection and cutting on pictures in the data, and screening out training pictures meeting conditions;
step 3: sequentially reading training pictures in batches, and performing data enhancement processing such as random rotation and zooming;
step 4: training a face recognition model by using a softmax function based on Gaussian distribution as a loss function;
step 5; judging whether the face recognition accuracy meets the design requirement or not; if the training data reaches the preset training data, outputting a final trained face recognition model, otherwise, continuously reading the training data to train the model;
step 6: outputting a face recognition model which is finally trained;
step 7: reading a face picture to be recognized;
step 8: carrying out face detection on the read face picture, and intercepting a face range;
step 9: extracting a face picture characteristic vector by using a face recognition model which is trained before;
step 10: calculating cosine distances between the extracted feature vectors to serve as face similarity measurement;
step 11: and finally, outputting the similarity between the face pictures and judging whether the face pictures are the same person or not.
Preferably, the specific processing procedure of Step3 is divided into random rotation of the training picture, random scaling of the training picture, and modified softmax loss function based on gaussian distribution.
Preferably, the random rotation of the training picture is performed by counting the angle range of the face of the monitoring image, and the angle distribution of the face of the monitoring image is found to be within a range of ± 30 degrees, so that in order to enable the model to cover all the angle ranges of the face, a random rotation operation is added during the training of the face recognition model, and the output face rotation angle is as follows:
θout＝30λθ-θin
wherein: thetaoutThe angle of rotation of the face is output; lambda [ alpha ]θIs a random number between-1 and 1; thetainThe face angle is the input face rotation angle, and after the face rotation, the face angle is uniformly distributed in the angle range of-30 degrees to 30 degrees.
Preferably, the random scaling of the training picture is performed by counting the distribution of the face resolution of the monitoring image, and it is found that the face resolution of the monitoring image is within a range of 16x 16-128 x128, so that in order to make the model cover all the face resolution ranges, a random scaling operation is added during the training of the face recognition model, since the resolution of the model training picture is 128x128, the training picture is firstly reduced by using a bicubic interpolation method, and then is enlarged to 128x128 by using the bicubic interpolation, and the resolution of the output face picture is:
sout＝λs
wherein: soutInterpolating the face reduction resolution for bicubic; lambda [ alpha ]sIs a random integer between 16 and 128. After random scaling, the face resolution is uniformly distributed between 16x16 and 128x 128.
Preferably, the modified softmax loss function based on the gaussian distribution is for the input face data x, pdata(x) Representing the true probability distribution, p, of a samplemodel(x | θ) represents the probability of the face recognition model output, and therefore the KL divergence of the modelComprises the following steps:
Figure BDA0002039497140000041
wherein: x is a face sample picture; x is a face data set; dKL(pdata||pmodel) The KL divergence between the probability distribution of the real data and the probability distribution of the model; p is a radical ofdata(x) Is the true data probability; p is a radical ofmodel(x) Calculating the probability of the face picture by the model; h (p) is a model-independent constant, DKL(pdata||pmodel) The smaller the model prediction probability is closer to the real picture probability, and after a constant H (p) irrelevant to the model is removed, the cross entropy form generally used by a face recognition loss function is obtained.
The method comprises the following steps of taking a softmax function as a model prediction probability, introducing a boundary mechanism, namely a cos-softmax loss function with high accuracy rate of face recognition of a single picture at present, taking a regularization term as an inter-class distance, introducing an inter-class distance constraint mechanism to improve the cos-softmax loss function, and preventing the robustness of a face recognition model area from being influenced by too small inter-class distance:
Lpro-cosSoftmax＝Lcos-Softmax+λLconv
Figure BDA0002039497140000042
Lconv＝-|conv(W，W)||F
Figure RE-GDA0002487547310000051
Figure RE-GDA0002487547310000052
Figure RE-GDA0002487547310000053
wherein: l ispro-cosSoftmaxFor improved cos-a softmax loss function comprising cos-softmax loss function Lpro-cosSoftmaxSum covariance loss function Lconv(ii) a N is the size of the training batch, s is a scaling factor, the weight of the loss function is adjusted, m is the boundary between classes, and the training intensity is adjusted; wjIs the jth column of the weight of the last fully-connected layer of the network, i.e. the representative vector of the jth class in the training sample, Wj *Is WjThe result of the normalization; x is the number ofiRepresenting the feature vector extracted from the ith picture in the batch,
Figure RE-GDA0002487547310000054
is xiThe result of the normalization; cos (theta)j，i) It represents the j-th type normalized feature W in the training samplej *A normalized feature of the ith sample picture
Figure RE-GDA0002487547310000055
The included angle between the two eigenvectors, that is, the distance between the two eigenvectors, which is used for classifying the softmax loss, that is, the softmax part of the loss function; λ represents the weight coefficient of the regularization term, | conv (W, W) | luminanceFThe covariance matrix is calculated for W, and then the Frobenius norm is calculated for the covariance matrix, namely the square sum of the absolute values of matrix elements is squared again to represent the inter-class distance.
In order to consider the imbalance of the training samples, namely the number of high-quality samples in the training samples is large, and the number of low-quality samples is small, so that the amplitude of the extracted feature high-quality samples is large, and the amplitude of the low-quality samples is small; based on this observation, the sample sampling probability is introduced into the model probability distribution using the feature amplitude as the sample sampling distribution:
pmodel(x|θ)＝pmodel-cls*pmodel-samp
Figure BDA0002039497140000055
Figure RE-GDA0002487547310000062
Figure BDA0002039497140000057
wherein: p is a radical ofmodel(x | θ) represents the probability distribution of the face recognition model θ output for the input picture x; including therein a model classification probability pmodel-clsAnd model sampling probability pmodel-samp(ii) a Model classification probability pmodel-clsThe classification probability of the cos-softmax loss function is obtained; model sampling probability pmodel-sampFitting Gaussian distribution; mu is a characteristic mean value; sigma is a characteristic variance; p is to bemodel(x | θ) is substituted into the cross entropy formula, i.e. the modified softmax loss function based on the gaussian distribution.
Preferably, Step1, Step2, Step3, Step4, Step5 and Step6 constitute a model training section, and Step7, Step8, Step9, Step10 and Step11 constitute a face recognition section.
(III) advantageous effects
Compared with the prior art, the invention provides a low-resolution large-angle face recognition method based on Gaussian distribution, which has the following beneficial effects:
1. the low-resolution large-angle face recognition method based on Gaussian distribution is characterized in that the method comprises the following steps of 3: the training pictures are sequentially read in batches, and data enhancement processing such as random rotation and scaling is carried out, so that the accuracy of the face identification model on face angles and resolution is improved.
2. According to the low-resolution large-angle face recognition method based on Gaussian distribution, a softmax function improved by Gaussian distribution is introduced through Step4, the distribution rule of training data is considered, weighting is carried out by using Gaussian distribution probability, the face recognition accuracy of pictures with less distribution is improved, and the robustness of a model is enhanced.
Drawings
Fig. 1 is a flowchart of a low-resolution large-angle face recognition method based on gaussian distribution according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, a low-resolution large-angle face recognition method based on gaussian distribution includes a model training portion and a face recognition portion, where the model training portion is: reading model training data, performing preprocessing work such as data cutting and cleaning, then reading the processed data in batches, performing operations such as random rotation and scaling, and training a face recognition model by using an improved softmax function based on Gaussian distribution as a loss function; the face recognition part is as follows: inputting a human face picture to be recognized, performing human face detection on the picture and intercepting a human face range, extracting human face characteristic vectors by using a human face recognition model trained in S1, calculating cosine distances among the characteristic vectors, and finally outputting similarity among human faces, wherein the specific steps are as follows:
step 1: reading a face data set used for training a face recognition model;
step 2: carrying out face detection and cutting on pictures in the data, and screening out training pictures meeting conditions;
step 3: sequentially reading training pictures in batches, and performing data enhancement processing such as random rotation and zooming;
step 4: training a face recognition model by using a softmax function based on Gaussian distribution as a loss function;
step 5; judging whether the face recognition accuracy meets the design requirement or not; if the training data reaches the preset training data, outputting a final trained face recognition model, otherwise, continuously reading the training data to train the model;
step 6: outputting a face recognition model which is finally trained;
step 7: reading a face picture to be recognized;
step 8: carrying out face detection on the read face picture, and intercepting a face range;
step 9: extracting a face picture characteristic vector by using a face recognition model which is trained before;
step 10: calculating cosine distances between the extracted feature vectors to serve as face similarity measurement;
step 11: and finally, outputting the similarity between the face pictures and judging whether the face pictures are the same person or not.
The specific processing procedure of Step3 is divided into random rotation of the training picture, random scaling of the training picture and modified softmax loss function based on Gaussian distribution.
The random rotation of the training picture is realized by counting the angle range of the face of the monitoring image, and the angle distribution of the face of the monitoring image is found to be within the range of +/-30 degrees, so that in order to enable the model to cover all the angle ranges of the face, the random rotation operation is added during the training of the face recognition model, and the output face rotation angle is as follows:
θout＝30λθ-θin
wherein: thetaoutThe angle of rotation of the face is output; lambda [ alpha ]θIs a random number between-1 and 1; thetainThe human face rotation angle is input, after the human face rotation angle is rotated, the human face angles are uniformly distributed in an angle range of-30 degrees to 30 degrees, and the human face recognition model can be more robust to the human face large angle by using the human face picture after the angle enhancement to carry out human face recognition model training.
The random zooming of the training picture is realized by counting the face resolution distribution of the monitoring image, and the face resolution of the monitoring image is found to be in the range of 16x 16-128 x128, so that in order to enable the model to cover all the face resolution ranges, random zooming operation is added during the training of the face recognition model, because the resolution of the model training picture is 128x128, the training picture is firstly reduced by a bicubic interpolation method, then the bicubic interpolation is used for amplifying the training picture to 128x128, and the resolution of the output face picture is as follows:
sout＝λs
wherein: soutInterpolating a person for bicubic interpolationResolution of face reduction; lambda [ alpha ]sIs a random integer between 16 and 128. After random scaling, the face resolution is uniformly distributed between 16x 16-128 x128, and the face recognition model training is carried out by using the face picture after resolution enhancement, so that the face recognition model has stronger robustness to low-resolution faces and higher recognition accuracy.
The improved softmax loss function based on Gaussian distribution is used for input face data x, pdata(x) Representing the true probability distribution, p, of a samplemodel(x | θ) represents the probability of the face recognition model output, so the KL divergence of the model is:
Figure BDA0002039497140000081
wherein: x is a face sample picture; x is a face data set; dKL(pdata||pmodel) The KL divergence between the probability distribution of the real data and the probability distribution of the model; p is a radical ofdata(x) Is the true data probability; p is a radical ofmodel(x) Calculating the probability of the face picture by the model; h (p) is a model-independent constant, DKL(pdata||pmodel) The smaller the model prediction probability is closer to the real picture probability, and after a constant H (p) irrelevant to the model is removed, the cross entropy form generally used by a face recognition loss function is obtained.
The method comprises the following steps of taking a softmax function as a model prediction probability, introducing a boundary mechanism, namely a cos-softmax loss function with high accuracy rate of face recognition of a single picture at present, taking a regularization term as an inter-class distance, introducing an inter-class distance constraint mechanism to improve the cos-softmax loss function, and preventing the robustness of a face recognition model area from being influenced by too small inter-class distance:
Lpro-cosSoftmax＝Lpro-cosSoftmax+λLconv
Figure RE-GDA0002487547310000101
Lconv＝-||conv(w,w)||F
Figure RE-GDA0002487547310000102
Figure RE-GDA0002487547310000103
Figure RE-GDA0002487547310000104
wherein: l ispro-cosSoftmaxFor improved cos-softmax loss function, comprising a cos-softmax loss function Lpro-cosSoftmaxSum covariance loss function Lconv(ii) a N is the size of the training batch, s is a scaling factor, the weight of the loss function is adjusted, m is the boundary between classes, and the training intensity is adjusted; wjIs the jth column of the weight of the last fully-connected layer of the network, i.e. the representative vector of the jth class in the training sample, Wj *Is WjA result of the normalization; x is the number ofiRepresenting the feature vector extracted from the ith picture in the batch,
Figure RE-GDA0002487547310000105
is xiThe result of the normalization; cos (theta)j，i) It represents the j-th type normalized feature W in the training samplej *A normalized feature of the ith sample picture
Figure RE-GDA0002487547310000106
The included angle between the two eigenvectors, that is, the distance between the two eigenvectors, which is used for classifying the softmax loss, that is, the softmax part of the loss function; λ represents the weight coefficient of the regularization term, | conv (W, W) | luminanceFThe covariance matrix is calculated for W, then the Frobenius norm is calculated for the covariance matrix, namely the square sum of the absolute values of matrix elements is squared again, the inter-class distance is represented, after the constraint of the inter-class distance is increased, the robustness of the model can be increased, and the accuracy of the face recognition of the model is improved.
In order to consider the imbalance of the training samples, namely the number of high-quality samples in the training samples is large, and the number of low-quality samples is small, so that the amplitude of the extracted feature high-quality samples is large, and the amplitude of the low-quality samples is small; based on this observation, the sample sampling probability is introduced into the model probability distribution using the feature amplitude as the sample sampling distribution:
pmodel(x|θ)＝pmodel-cls*pmodel-samp
Figure RE-GDA0002487547310000111
Figure RE-GDA0002487547310000112
Figure RE-GDA0002487547310000113
wherein: p is a radical ofmodel(x | θ) represents the probability distribution of the face recognition model θ output for the input picture x; including therein a model classification probability pmodel-clsAnd model sampling probability pmodel-samp(ii) a Model classification probability pmodel-clsThe classification probability of the cos-softmax loss function is obtained; model sampling probability pmodel-sampFitting Gaussian distribution; mu is a characteristic mean value; sigma is a characteristic variance; p is to bemodelAnd (x | theta) is substituted into a cross entropy formula, namely, an improved softmax loss function based on Gaussian distribution, and after the Gaussian distribution is introduced, the model can give larger weight to the samples with less distribution, so that the model pays more attention to the low-resolution large-angle face picture, and the recognition rate of the trained model to the large-angle low-resolution face picture with more monitoring videos is greatly enhanced.
To sum up, the low-resolution and large-angle face recognition method based on the gaussian distribution is implemented by the steps of 3: the method comprises the steps of sequentially reading training pictures in batches, carrying out data enhancement processing such as random rotation and scaling, improving the accuracy of a face recognition model on face angles and resolution, introducing a softmax function improved by Gaussian distribution through Step4, considering the distribution rule of training data, weighting by using Gaussian distribution probability, improving the accuracy of face recognition of the pictures with less distribution, and enhancing the robustness of the model.
It is to be noted that the term "comprises," "comprising," or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (6)

1. A low-resolution large-angle face recognition method based on Gaussian distribution is characterized by comprising the following specific steps:
step 1: reading a face data set used for training a face recognition model;
step 2: carrying out face detection and cutting on pictures in the data, and screening out training pictures meeting conditions;
step 3: sequentially reading training pictures in batches, and performing data enhancement processing such as random rotation and zooming;
step 4: training a face recognition model by using a softmax function based on Gaussian distribution as a loss function;
step 5; judging whether the face recognition accuracy meets the design requirement or not; if the training data reaches the preset training data, outputting a final trained face recognition model, otherwise, continuously reading the training data to train the model;
step 6: outputting a face recognition model which is finally trained;
step 7: reading a face picture to be recognized;
step 8: carrying out face detection on the read face picture, and intercepting a face range;
step 9: extracting a face picture characteristic vector by using a face recognition model which is trained before;
step 10: calculating cosine distances between the extracted feature vectors to serve as face similarity measurement;
step 11: and finally, outputting the similarity between the face pictures and judging whether the face pictures are the same person or not.
2. The low-resolution large-angle face recognition method based on the Gaussian distribution as claimed in claim 1, characterized in that: the specific processing process of Step3 comprises random rotation of the training picture, random scaling of the training picture and an improved softmax loss function based on Gaussian distribution.
3. The low-resolution large-angle face recognition method based on the Gaussian distribution as claimed in claim 2, characterized in that: the random rotation of the training picture is realized by counting the angle range of the face of the monitoring image, and the angle distribution of the face of the monitoring image is found to be within the range of +/-30 degrees, so that in order to enable the model to cover all the angle ranges of the face, random rotation operation is added during the training of the face recognition model, and the output face rotation angle is as follows:
θout＝30λθ-θin
wherein: thetaoutThe angle of rotation of the face is output; lambda [ alpha ]θIs a random number between-1 and 1; thetainThe face angle is the input face rotation angle, and after the face rotation, the face angle is uniformly distributed in the angle range of-30 degrees to 30 degrees.
4. The low-resolution large-angle face recognition method based on the Gaussian distribution as claimed in claim 2, characterized in that: the random scaling of the training picture is realized by counting the distribution of the face resolution of the monitoring image, and the face resolution of the monitoring image is found to be in the range of 16x 16-128 x128, so that in order to enable the model to cover all the face resolution ranges, the random scaling operation is added when the face recognition model is trained, because the resolution of the model training picture is 128x128, the training picture is firstly reduced by using a bicubic interpolation method, then is enlarged to 128x128 by using the bicubic interpolation, and the resolution of the output face picture is as follows:
sout＝λs
wherein: soutInterpolating the face reduction resolution for bicubic; lambda [ alpha ]sIs a random integer between 16 and 128. After random scaling, the face resolution is uniformly distributed between 16x16 and 128x 128.
5. The low-resolution large-angle face recognition method based on the Gaussian distribution as claimed in claim 2, characterized in that: the improved softmax loss function based on the Gaussian distribution is used for input face data x, pdata(x) Representing the true probability distribution, p, of a samplemodel(x|θ)pdata(x) Representing the probability of the face recognition model output, the KL divergence of the model is therefore:
Figure RE-FDA0002487547300000021
wherein: x is a face sample picture; x is a face data set; dKL(pdata||pmodel) The KL divergence between the probability distribution of the real data and the probability distribution of the model is obtained; p is a radical ofdata(x) Is the true data probability; p is a radical ofmodel(x) Calculating the probability of the face picture by the model; h (p) is a model-independent constant, DKL(pdata||pmodel) The smaller the model prediction probability is, the closer the model prediction probability is to the real picture probability, and after a constant H (p) irrelevant to the model is removed, the model is a cross entropy form generally used by a face recognition loss function.
The method comprises the following steps of taking a softmax function as a model prediction probability, introducing a boundary mechanism, namely a cos-softmax loss function with high accuracy rate of single-picture face recognition at present, taking a regularization term as an inter-class distance, introducing an inter-class distance constraint mechanism to improve the cos-softmax loss function, and preventing the inter-class distance from being too small to influence the robustness of a face recognition model area:
Lpro-cosSoftmax＝Lpro-cosSoftmax+λLconv
Figure RE-FDA0002487547300000031
Lconv＝-||conv(w,w)||F
Figure RE-FDA0002487547300000032
Figure RE-FDA0002487547300000033
Figure RE-FDA0002487547300000034
wherein: l ispro-cosSoftmaxFor improved cos-softmax loss function, comprising a cos-softmax loss function Lcos-SoftmaxSum covariance loss function Lconv(ii) a N is the size of the training batch, s is a scaling factor, the weight of the loss function is adjusted, m is the boundary between classes, and the training intensity is adjusted; wjIs the jth column of the weight of the last fully-connected layer of the network, i.e. the representative vector of the jth class in the training sample, Wj *Is WjThe result of the normalization; x is the number ofiRepresenting the feature vector extracted from the ith picture in the batch,
Figure RE-FDA0002487547300000041
is xiThe result of the normalization; cos (theta)j，i) It represents the j-th normalized feature in the training sample
Figure RE-FDA0002487547300000042
With respect to the ith sample pictureA normalization feature
Figure RE-FDA0002487547300000043
The included angle between the two eigenvectors, that is, the distance between the two eigenvectors, which is used for classifying the softmax loss, that is, the softmax part in the loss function; λ represents the weight coefficient of the regularization term, | conv (W, W) | luminanceFThe method comprises the steps of solving a covariance matrix for W, and solving a Frobenius norm for the covariance matrix, namely, the square sum of absolute values of matrix elements is solved, and then the square is opened, so as to represent the inter-class distance.
In order to consider the imbalance of the training samples, namely the number of high-quality samples in the training samples is large, and the number of low-quality samples is small, so that the amplitude of the extracted feature high-quality samples is large, and the amplitude of the low-quality samples is small; based on this observation, the sample sampling probability is introduced into the model probability distribution using the feature amplitude as the sample sampling distribution:
pmodel(x|θ)＝pmodel-cls*pmodel-samp
Figure RE-FDA0002487547300000044
Figure RE-FDA0002487547300000045
Figure RE-FDA0002487547300000046
wherein: p is a radical ofmodel(x | θ) represents the probability distribution of the face recognition model θ output for the input picture x; including the model classification probability pmodel-clsAnd model sampling probability pmodel-samp(ii) a Model classification probability pmodel-clsThe classification probability of the cos-softmax loss function is obtained; model sampling probability pmodel-sampThe Gaussian distribution is met; mu is a characteristic mean value; sigma is a characteristic variance; p is to bemodelThe cross entropy formula is substituted by (x | theta) to form an improved softmax loss function based on Gaussian distribution.
6. The low-resolution large-angle face recognition method based on the Gaussian distribution as claimed in claim 1, characterized in that: the Step1, the Step2, the Step3, the Step4, the Step5 and the Step6 form a model training part, and the Step7, the Step8, the Step9, the Step10 and the Step11 form a face recognition part.
CN201910337099.7A 2019-04-25 2019-04-25 Low-resolution large-angle face recognition method based on Gaussian distribution Pending CN111339810A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910337099.7A CN111339810A (en) 2019-04-25 2019-04-25 Low-resolution large-angle face recognition method based on Gaussian distribution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910337099.7A CN111339810A (en) 2019-04-25 2019-04-25 Low-resolution large-angle face recognition method based on Gaussian distribution

Publications (1)

Publication Number Publication Date
CN111339810A true CN111339810A (en) 2020-06-26

Family

ID=71183270

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910337099.7A Pending CN111339810A (en) 2019-04-25 2019-04-25 Low-resolution large-angle face recognition method based on Gaussian distribution

Country Status (1)

Country Link
CN (1) CN111339810A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112434714A (en) * 2020-12-03 2021-03-02 北京小米松果电子有限公司 Multimedia identification method, device, storage medium and electronic equipment
CN113111781A (en) * 2021-04-14 2021-07-13 东南大学 Method for classifying layer supervision for face recognition based on Sigmoid function
CN113449704A (en) * 2021-08-31 2021-09-28 北京的卢深视科技有限公司 Face recognition model training method and device, electronic equipment and storage medium
CN113763348A (en) * 2021-09-02 2021-12-07 北京格灵深瞳信息技术股份有限公司 Image quality determination method and device, electronic equipment and storage medium
CN115862118A (en) * 2023-01-29 2023-03-28 南京开为网络科技有限公司 Face age estimation method and device based on Gaussian distribution assumption and MSE loss
CN116453201A (en) * 2023-06-19 2023-07-18 南昌大学 Face recognition method and system based on adjacent edge loss

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307072A1 (en) * 2015-04-17 2016-10-20 Nec Laboratories America, Inc. Fine-grained Image Classification by Exploring Bipartite-Graph Labels
CN107103281A (en) * 2017-03-10 2017-08-29 中山大学 Face identification method based on aggregation Damage degree metric learning
CN108319943A (en) * 2018-04-25 2018-07-24 北京优创新港科技股份有限公司 A method of human face recognition model performance under the conditions of raising is worn glasses

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160307072A1 (en) * 2015-04-17 2016-10-20 Nec Laboratories America, Inc. Fine-grained Image Classification by Exploring Bipartite-Graph Labels
CN107103281A (en) * 2017-03-10 2017-08-29 中山大学 Face identification method based on aggregation Damage degree metric learning
CN108319943A (en) * 2018-04-25 2018-07-24 北京优创新港科技股份有限公司 A method of human face recognition model performance under the conditions of raising is worn glasses

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112434714A (en) * 2020-12-03 2021-03-02 北京小米松果电子有限公司 Multimedia identification method, device, storage medium and electronic equipment
CN113111781A (en) * 2021-04-14 2021-07-13 东南大学 Method for classifying layer supervision for face recognition based on Sigmoid function
CN113449704A (en) * 2021-08-31 2021-09-28 北京的卢深视科技有限公司 Face recognition model training method and device, electronic equipment and storage medium
CN113449704B (en) * 2021-08-31 2022-03-25 北京的卢深视科技有限公司 Face recognition model training method and device, electronic equipment and storage medium
CN113763348A (en) * 2021-09-02 2021-12-07 北京格灵深瞳信息技术股份有限公司 Image quality determination method and device, electronic equipment and storage medium
CN115862118A (en) * 2023-01-29 2023-03-28 南京开为网络科技有限公司 Face age estimation method and device based on Gaussian distribution assumption and MSE loss
CN115862118B (en) * 2023-01-29 2023-05-23 南京开为网络科技有限公司 Face age estimation method and device based on Gaussian distribution assumption and MAE loss
CN116453201A (en) * 2023-06-19 2023-07-18 南昌大学 Face recognition method and system based on adjacent edge loss
CN116453201B (en) * 2023-06-19 2023-09-01 南昌大学 Face recognition method and system based on adjacent edge loss

Similar Documents

Publication Publication Date Title
CN111339810A (en) Low-resolution large-angle face recognition method based on Gaussian distribution
CN115861135B (en) Image enhancement and recognition method applied to panoramic detection of box body
CN103093215B (en) Human-eye positioning method and device
CN101916383B (en) Vehicle detecting, tracking and identifying system based on multi-camera
CN100550037C (en) Utilize and improve Hausdorff apart from the method for extracting the identification human ear characteristic
CN113177467A (en) Flame identification method, system, device and medium
CN107392968B (en) Image saliency detection method fused with color contrast map and color space distribution map
CN110309781A (en) Remote sensing recognition method for house damage based on multi-scale spectral texture adaptive fusion
Tsai et al. Road sign detection using eigen colour
CN105512683A (en) Target positioning method and device based on convolution neural network
Tao et al. Smoke vehicle detection based on multi-feature fusion and hidden Markov model
CN109063619A (en) A kind of traffic lights detection method and system based on adaptive background suppression filter and combinations of directions histogram of gradients
CN106845457A (en) Method for detecting infrared puniness target based on spectrum residual error with fuzzy clustering
CN112784712B (en) Missing child early warning implementation method and device based on real-time monitoring
CN112200746A (en) Dehazing method and device for foggy traffic scene images
CN103955949A (en) Moving target detection method based on Mean-shift algorithm
Lim et al. Text segmentation in color images using tensor voting
Zhang et al. Contour detection via stacking random forest learning
CN107341456B (en) Weather sunny and cloudy classification method based on single outdoor color image
CN109313705A (en) Fingerprint identification method, device, equipment and storage medium
CN107977608B (en) Method for extracting road area of highway video image
CN113449663B (en) Collaborative intelligent security method and device based on polymorphic fitting
Liu et al. An image edge detection algorithm based on genetic fuzzy
Li et al. Real-time traffic sign detection algorithm based on dynamic threshold segmentation and SVM
Lin et al. Face detection based on skin color segmentation and SVM classification

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20200626

RJ01 Rejection of invention patent application after publication