CN109063565A - A kind of low resolution face identification method and device - Google Patents

A kind of low resolution face identification method and device Download PDF

Info

Publication number
CN109063565A
CN109063565A CN201810698280.6A CN201810698280A CN109063565A CN 109063565 A CN109063565 A CN 109063565A CN 201810698280 A CN201810698280 A CN 201810698280A CN 109063565 A CN109063565 A CN 109063565A
Authority
CN
China
Prior art keywords
resolution face
resolution
low
face
feature
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
CN201810698280.6A
Other languages
Chinese (zh)
Other versions
CN109063565B (en
Inventor
葛仕明
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.)
Institute of Information Engineering of CAS
Original Assignee
Institute of Information Engineering of CAS
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 Institute of Information Engineering of CAS filed Critical Institute of Information Engineering of CAS
Priority to CN201810698280.6A priority Critical patent/CN109063565B/en
Publication of CN109063565A publication Critical patent/CN109063565A/en
Application granted granted Critical
Publication of CN109063565B publication Critical patent/CN109063565B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/19—Recognition using electronic means
    • G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
    • G06V30/194—References adjustable by an adaptive method, e.g. learning

Landscapes

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

Abstract

本发明提供一种低分辨率人脸识别方法及装置，针对低分辨率下的人脸识别问题，尤其是自然环境下的低分辨率人脸识别问题，采用双流深度学习网络结构，通过选择性知识蒸馏训练高效的低分辨率人脸识别网络，实现低分辨率人脸识别，具有接近高分辨率人脸识别模型的精度，速度、内存上则具有极大的优势。

The present invention provides a low-resolution face recognition method and device. Aiming at low-resolution face recognition problems, especially low-resolution face recognition problems in natural environments, a dual-stream deep learning network structure is adopted, through selective Knowledge distillation trains an efficient low-resolution face recognition network to achieve low-resolution face recognition, which has the accuracy close to the high-resolution face recognition model, and has great advantages in speed and memory.

Description

A kind of low resolution face identification method and device
Technical field
The invention belongs to computer vision and deep learning field more particularly to a kind of people under the conditions of low resolution Face recognition method and device.
Background technique
Face is as the essential attribute distinguished between individual, in computer vision and multimedia application field quilt daily Frequently identification.In these applications, human face recognition model needs are redeployed in cell phone even intelligent video camera head, are used for The numerous areas such as camera auto-focusing, human-computer interaction, photo management, city safety monitoring, intelligent driving.Currently, recognition of face In practical application under the conditions of open environment, it is often necessary to identify low resolution under the conditions of extremely low calculating and memory source Facial image.In this case, current many high-precision human face recognition models are often complex, it is difficult to actual deployment, And after the reduction of facial image resolution ratio, accuracy of identification can also generate very big decline.A kind of effective scheme is These high performance faceforms are converted into efficient model, and make the model after these conversions in identification low-resolution image When be able to maintain the precision of original complex model.
In recent years, low resolution field of face identification also studied certain methods.Existing method is typically split into two classes, packet Include the method based on enhancing and the method based on insertion.Method based on enhancing first carries out super-resolution to low-resolution face image Rate enhancing is rebuild, and then the model of retraining higher resolution is identified, facial image of the such methods after identification is rebuild Can have better as a result, still super-resolution operates and the high identification of rate respectively can bring biggish calculating cost, reduction The processing speed of model, constrains the actual deployment of model.In contrast, the method based on insertion is directly to low resolution people Face image extracts the feature with resolving ability, then training characteristics identifier progress recognition of face, most important in this method Processing be by the Knowledge delivery of high-resolution human face to low resolution face.Nevertheless, a key problem in technology be how from High-resolution human face correctly transmits desired knowledge rather than the A to Z of to low resolution face.This selective knowledge Transmitting is most important when converting existing high-precision human face recognition model into efficient low resolution human face recognition model One of challenge.
Summary of the invention
In order to overcome the deficiencies of the prior art, the present invention provides a kind of low resolution face identification method and device, the party Method passes through choosing using double-current deep learning network structure (including teacher's flow depth degree learning network and student's flow depth degree learning network) The efficient low resolution recognition of face network of sex knowledge distillation training is selected, the network implementations low resolution recognition of face is passed through.
In order to solve the above technical problems, the present invention is achieved through the following technical solutions:
A kind of low resolution face identification method, step include:
Using high-resolution human face image one deep neural network of training with identity marks, obtains teacher and flow deep learning Network；
Utilize high-resolution human face image and low-resolution face image the training one depth nerve marked with common identity Network obtains student's flow depth degree learning network；
High-resolution human face image is inputted Xiang teacher's flow depth degree learning network, obtains high-resolution human face depth characteristic, from The high-resolution human face depth characteristic rich in identity information is filtered out in high-resolution human face depth characteristic, should be rich in identity information High-resolution human face depth characteristic be to refer to correctly identify high-resolution human according to the high-resolution human face depth characteristic The identity of face image；
The low-resolution face image of junction belt identity marks and high-resolution human face Wei Shendute rich in identity information Sign, is finely adjusted training to student's flow depth degree learning network, obtains the trained student's flow depth that can identify low resolution face Spend learning network；
Low-resolution face image to be identified is input to trained student's flow depth degree learning network, obtains low resolution Rate face depth characteristic, and low resolution recognition of face is realized according to the low resolution face depth characteristic.
Further, it according to figure optimization method, filters out from high-resolution human face depth characteristic rich in identity information The step of high-resolution human face feature, the figure optimization method includes:
According to the high-resolution human face picture construction dense graph with identity marks, the dense graph is by face vertex and connection people Between the class inner edge or class on face vertex side form, one high-resolution human face image of a face vertex representation, a class inner edge by A pair of generic high-resolution human face image forms, and side is by a pair of different classes of high-resolution human face image group between a class At；
Sparse graph is constructed based on dense graph to reduce side complexity, to each generic high-resolution human in the sparse graph Face increases a mass center vertex, and a class inner edge is made of a pair of generic high-resolution human face image, between a class side by One mass center vertex of one high-resolution human face image and different classification composition；
Sparse graph is optimized, to filter out the high-resolution human face feature rich in identity information.
Further, when being optimized to sparse graph, by face vertex and mass center vertex representation at feature vector, base It is expressed as Closing Binary Marker formula in the high-resolution human face depth characteristic that this feature vector will be enriched in identity information, to the Closing Binary Marker Formula optimizes.
Further, algorithm is cut using figure to optimize sparse graph.
Further, after student's flow depth degree learning network receives low-resolution face image, high-resolution is first generated Rate face depth characteristic generates low resolution face depth characteristic using compression.
Further, being finely adjusted trained method to student's flow depth degree learning network is: student is flowed deep learning net Network receives high-resolution human face depth characteristic caused by the low-resolution face image with identity marks and is forced by feature recurrence The high-resolution human face depth characteristic for being rich in identity information, then according to after by high-resolution human face depth characteristic compression Low resolution face depth characteristic classified by identities to the low-resolution face image with identity marks.
Further, the recurrence loss function that the feature returns and the Classification Loss function of identities collectively constitute always Loss function obtain trained student's flow depth degree learning network by solving to total loss function.
Further, when the low resolution face depth characteristic is used for face alignment, to two width for face alignment Low-resolution face image carries out characteristic similarity calculating, if similitude is more than a threshold value, judgement is same people.
Further, when the low resolution face depth characteristic is used for face authentication, then to low resolution face depth Feature is classified, to determine identity.
A kind of low resolution face identification device distills module, spy including the feature based on student's flow depth degree learning network Levy compression module and feature recognition module；
Feature distillation module is used to obtain the high-resolution rich in identity information transmitted from teacher's flow depth degree learning network Face depth characteristic；
Feature Compression module is used for Feature Dimension Reduction, obtains the low resolution face depth characteristic with distinguishing ability；
Feature recognition module is used for low resolution recognition of face, is that a feature comparative device or one have Softmax layers of son Network.
Further, student's flow depth degree learning network connects entirely including 10 convolutional layers, 3 maximum pond layers and 3 Connect layer, convolutional layer include several 3 × 3 convolutional layer and between several 3 × 3 convolutional layer or 3 × 3 convolutional layer with Several 1 × 1 convolutional layer for compressive features figure number of active lanes between maximum pond layer, 1 × 1 convolutional layer can slow down The non-thread sexuality for gradually increasing, enhancing student's flow depth degree learning network of characteristic pattern number of active lanes reduces reasoning memory, maximum For pond layer for reducing characteristic dimension, 3 full articulamentums are Image Planes, identity layer and Softmax layers.
The beneficial effects of the present invention are: for the recognition of face problem under low resolution, especially under natural environment Low resolution recognition of face problem, recognition methods of the invention and device have the essence close to high-resolution human face identification model It spends, then there is great advantage in speed, memory.To the face under high definition case, face identification method of the invention and Device also has good processing capacity.
Detailed description of the invention
Fig. 1 is a kind of flow chart of low resolution face identification method of the present invention；
Fig. 2 is student's deep learning network structure of the invention
Fig. 3 A to Fig. 3 C is the figure after dense graph, sparse graph and optimization
Fig. 4 is student's flow depth degree learning network training schematic diagram of the invention
Specific embodiment
To be clearer and more comprehensible above scheme and beneficial effect of the invention, hereafter by embodiment, and attached drawing is cooperated to make Detailed description are as follows.
The present embodiment provides a kind of low resolution face identification method and the device of realization this method, which includes spy Sign distillation module (CNN-D), Feature Compression module (CNN-C) and feature recognition module (CNN-R), the stream of this method cognitive phase As shown in the lower half portion Fig. 1, step includes: journey
1) low-resolution face image is received.
2) higher-dimension depth characteristic (the i.e. high-resolution human face depth spy that module extracts low resolution face is distilled by feature Sign).
Feature distillation module is made of several layers of before student's flow depth degree learning network, and module receives low resolution face figure Picture exports higher-dimension depth characteristic, and this feature, which is approached, correctly proposes high-resolution human face image by teacher's flow depth degree learning network The feature taken.
Since feature distills module, the student's flow depth degree learning network for obtaining training is in processing low-resolution face image When have the ability that teacher's flow depth degree learning network correctly handles high-resolution human face image, i.e. student's flow depth degree learning network Practise the correct knowledge of teacher's flow depth degree learning network.And for the knowledge of teacher's flow depth degree learning network missing or mistake, it learns Raw flow depth degree learning network also has correct processing capacity, to obtain higher recognition of face precision.
3) Feature Dimension Reduction is carried out by Feature Compression module, obtaining low-dimensional depth characteristic, (i.e. low resolution face depth is special Sign).
Since the depth characteristic dimension that high-resolution human face identifies is higher, cause to occupy in actual deployment in biggish Deposit and reduce the efficiency of recognition of face.For this problem, the height of Feature Compression module of the invention to feature distillation module output It ties up depth characteristic and carries out dimensionality reduction, obtain the low-dimensional depth characteristic with discriminating power, be conducive to promote face alignment, certification and inspection The speed of rope, to effectively support actual deployment.
Feature Compression module is made of the several full articulamentums in centre in student's flow depth degree learning network, and feature distills module The higher-dimension depth characteristic of output carries out dimensionality reduction by Feature Compression module, obtains low-dimensional depth characteristic.
4) by feature recognition module, judge the classification of the face, realize the identification of face.
According to recognition of face be for carry out the face alignment of 1:1 (face verification) or for 1:N (N > 1) face authentication (face identification) is configured.If it is face alignment is carried out, feature recognition module is one A feature comparative device measures the distance between two low-dimensional depth characteristics and judges that they belong to same class or inhomogeneity； If it is face authentication is carried out, feature recognition module is the sub-network for having Softmax layers, and low-dimensional depth characteristic is classified At certain one kind.
Student's flow depth degree learning network is to pass through the side of selection sex knowledge distillation based on teacher's flow depth degree learning network Method training obtains, and as shown in the uppermost portion of fig. 1, is described as follows:
Teacher's flow depth degree learning networkUsually more complicated trained deep neural network in advance, It includes more network parameter Wt, WtBy training from identity label sets LtHigh-resolution teacher's human face data collection Dt On drawn a large amount of knowledge, DtIt is sightless generally for student's flow depth degree learning network；Receive high-resolution Rate facial imageAs input, higher-dimension depth characteristic is obtained by reasoning, higher-dimension depth characteristic is selected according to figure optimization It selects, judges whether each higher-dimension depth characteristic belongs to most rich information knowledge and (be rich in the high-resolution human face depth of identity information Feature), to pass to student's flow depth degree learning network.
Student's flow depth degree learning networkIt is a simple deep neural network, receives low resolution face ImageAs input, it includes less network parameter Ws, WsBy from student's human face data collectionStudy obtains, wherein | Ds| it is the quantity of high-resolution human face image, student's human face data It concentrates to every panel height resolution ratio facial imageInclude N corresponding low-resolution face image, jth width low resolution face Image is expressed asHigh-resolution human face imageWith corresponding low-resolution face imageIdentity having the same Mark li∈ { 1,2 ..., L }, wherein L is identity marks sum.
Be made of feature distillation module, Feature Compression module and feature recognition module, returned using feature and The study of two tasks of identities is trained, and it is deep to generate higher-dimension in the last layer (referred to as Image Planes) of feature distillation module Feature is spent, the higher-dimension depth characteristic selected from teacher's flow depth degree learning network is approached by feature recurrence；Then by spy It levies compression module and generates low-dimensional depth characteristic；Last low-dimensional depth characteristic enters feature recognition module, is realized by identities Classify to the low-resolution face image of student's stream.The combination learning with two tasks of identities is returned by feature, Make student's flow depth degree learning networkOn the one hand it can be come fromMost rich information knowledge, another party Face can also recover the knowledge that teacher's flow network does not provide.
As shown in Fig. 2, student's flow depth degree learning network is mainly made of 3 × 3 convolutional layer, 1 × 1 convolutional layer is placed Between 3 × 3 convolutional layer or between maximum pond layer and 3 × 3 convolutional layer, slow down the gradually increasing of characteristic pattern number of active lanes Add, enhance the non-thread sexuality of student's flow depth degree learning network, and reduces reasoning memory；In addition, global average pond layer by with In carrying out prediction processing, maximum pond layer is for reducing characteristic dimension；The Image Planes of 1 × 1 × D are for receiving from teacher's flow depth degree The knowledge selected in learning network, D are characteristic dimension of teacher's flow depth degree network to high-resolution human face image zooming-out.Finally, Student's flow depth degree learning network includes 10 convolutional layers, 3 maximum pond layers and 3 full articulamentums.To obtained network ginseng Quantity is 0.79M, only the 0.57% of teacher's flow depth degree learning network parameter amount, therefore student's flow depth degree learning network obtains Greatly compress.
As shown in Figure 3A to 3C, in order to select the knowledge of teacher's flow depth degree learning network, the present invention is to high-resolution human face Image selects with passing through the feature progress distinctive after teacher's flow depth degree learning network, and benchmark is selected to flow deep learning for teacher The face characteristic that network can correctly identify the identity of high-resolution human face image is used as most rich information knowledge.Due to for instructing The facial image identity of the identity and training teacher's flow depth degree learning network of practicing the facial image of student's flow depth degree learning network is logical Often inconsistent, i.e. teacher's flow depth degree learning network cannot having to explicitly know whether high-resolution human face image identifies correctly.For this purpose, We are bright to use figure optimization method, by considering training teacher's flow depth degree learning network to the spy of high-resolution human face image zooming-out Distinctive between sign implicitly identifies.
Specifically, it is constructed by the face characteristic dissmilarity of the similar different identity of face characteristic of same identity marks Figure.In the figure of building, every panel height resolution ratio facial image is a vertex (referred to as face vertex), and two vertex form one Side.If including come structure figures (referred to as dense graph) by the way of being fully connected | Ds| a vertex andSide, causes the quantity on side excessive, causes figure optimization complicated, as shown in Figure 3A.For this purpose, this hair Bright building sparse graph is handled, as shown in Figure 3B, in sparse graph, to every group of face vertex with same identity marks Increase a mass center, the vertex in figure is made of face vertex and mass center vertex, while by structure between the vertex of same identity marks At internal edges and vertex be different from constituted between identity marks culminating point external edge composition, to be wrapped in sparse graph Containing (| Ds|+L) a vertex andWherein KcIt is the face vertex number that identity marks are c, hasIt can be seen that the quantity on side is greatly reduced compared to dense graph in sparse graph, to reduce answering for figure optimization Miscellaneous degree.Optimization Solution hereafter is carried out to sparse graph, the figure after being optimized, as shown in Figure 3 C.
In order to optimize to sparse graph, the present invention at feature vector form, divides on face vertex and mass center vertex representation It is not as follows:
Wherein,It is the sub-network of several layers of compositions before teacher's flow depth degree learning network, for high-resolution human Face imageExtract face characteristic；Indicator function δ (li=c) it indicates to work as liOtherwise it is 0 that value, which is 1, when=c.As it can be seen that each face Vertex is portrayed by specific appearance characteristics, and mass center vertex is then indicated by average appearance.
Given sparse graph, the selection of most rich information knowledge can be expressed as Closing Binary Marker formula below:
Meet αi∈ { 0,1 }, i=1,2 ..., | Ds| (2)
Wherein, α is binary set, i-th of element αi=1 indicates from face characteristic fiIt is selected as most rich information knowledge； Function d () is used to measure the similitude between two feature vectors using COS distance.As it can be seen that first item is inclined in formula (2) To the face characteristic in selection with the average face characteristic of other identity marks with less information, λ < 0 is negative weight for putting down Weighing two influences that Section 2 is made to tend to the face characteristic that selection has high similarity with the face of same identity marks. Particularly, using nonnegative function d () and negative weight λ, first item is made to tend to select less face characteristic, Section 2 tendency In selecting more face characteristics, by this two containing, so as to select most rich information knowledge.The present invention is using figure (Graph Cut) algorithm is cut (referring to Y.Boykov and V.Kolmogorov, " An experimental comparison of mincut/max-flow algorithms for energy minimization in computer vision,” IEEE Transactions on Pattern Analysis and Machine Intelligence(PAMI),vol.26, No.9, pp.1124-1137,2004) it is solved.After solution formula (2), the present invention can select limited quantity Face characteristic, corresponding face have similitude between similitude and low class in high class.It is handled by selection, due to teacher's flow depth degree Exception knowledge caused by learning network mistake identifies can effectively be abandoned by the mode that similar characteristics cluster, much useful to know Knowledge is distilled from teacher's flow depth degree learning network and noisy knowledge is then suppressed, so that these knowledge can be used preferably In the training of instruction of papil flow depth degree learning network.
As shown in figure 4, using the most rich information knowledge chosen as supervision, in conjunction with low resolution face identity marks Information, student's flow depth degree learning network handle two problems using multi-task learning simultaneously: 1) being selected by feature regression approach Face characteristic, 2) the face clue of missing is recovered by low resolution face.
Total loss function is formed by loss and Classification Loss two parts are returned, and is passed through and is solved following problems:
Wherein,It is the parameter of feature distillation sub-network,It is the most rich information knowledge common identity with selection Corresponding low-resolution face image is marked, β > 0 returns loss and Classification Loss defines respectively for balancing two parts loss Are as follows:
Formula (3) is solved by backpropagation algorithm, obtains student's flow depth degree learning network model, and the model is available In progress face authentication and recognition of face.
It is above to implement to be merely illustrative of the technical solution of the present invention rather than be limited, the ordinary skill people of this field Member can be with modification or equivalent replacement of the technical solution of the present invention are made, without departing from the spirit and scope of the present invention, this hair Bright protection scope should be subject to described in claims.

Claims (10)

1.一种低分辨率人脸识别方法，其步骤包括：1. A low-resolution face recognition method, the steps comprising: 利用带身份标记的高分辨率人脸图像训练一深度神经网络，获得教师流深度学习网络；Using high-resolution face images with identity tags to train a deep neural network to obtain a teacher-flow deep learning network; 利用带相同身份标记的高分辨率人脸图像和低分辨率人脸图像训练一深度神经网络，获得学生流深度学习网络；Using high-resolution face images and low-resolution face images with the same identity tag to train a deep neural network to obtain a student flow deep learning network; 向教师流深度学习网络输入高分辨率人脸图像，得到高分辨率人脸深度特征，从高分辨率人脸深度特征中筛选出富含身份信息的高分辨率人脸深度特征，该富含身份信息的高分辨率人脸深度特征是指能够根据该高分辨率人脸深度特征正确鉴别出高分辨率人脸图像的身份；Input the high-resolution face image to the teacher flow deep learning network to obtain the high-resolution face depth features, and select the high-resolution face depth features rich in identity information from the high-resolution face depth features. The high-resolution face depth feature of identity information refers to the identity of the high-resolution face image that can be correctly identified according to the high-resolution face depth feature; 结合带身份标记的低分辨率人脸图像和富含身份信息的高分辨率人脸维深度特征，对学生流深度学习网络进行微调训练，得到能识别低分辨率人脸的训练好的学生流深度学习网络；将待识别的低分辨率人脸图像输入至训练好的学生流深度学习网络，得到低分辨率人脸深度特征，并根据该低分辨率人脸深度特征实现低分辨率人脸识别。Combining low-resolution face images with identity marks and high-resolution face-dimensional depth features rich in identity information, the student flow deep learning network is fine-tuned and trained to obtain a trained student flow that can recognize low-resolution faces Deep learning network; input the low-resolution face image to be recognized to the trained student flow deep learning network to obtain low-resolution face depth features, and realize low-resolution face recognition based on the low-resolution face depth features identify. 2.如权利要求1所述的方法，其特征在于，根据图优化方法，从高分辨率人脸深度特征中筛选出富含身份信息的高分辨率人脸特征，该图优化方法的步骤包括：2. The method according to claim 1, wherein, according to the graph optimization method, the high-resolution face feature rich in identity information is screened out from the high-resolution face depth feature, and the steps of the graph optimization method include : 根据带身份标记的高分辨率人脸图像构建密集图，该密集图由人脸顶点和连接人脸顶点的类内边或类间边组成，一个人脸顶点表示一个高分辨率人脸图像，一条类内边由一对同类别的高分辨率人脸图像组成，一条类间边由一对不同类别的高分辨率人脸图像组成；Construct a dense graph based on high-resolution face images with identity labels. The dense graph consists of face vertices and intra-class edges or inter-class edges connecting face vertices. A face vertex represents a high-resolution face image. An intra-class edge consists of a pair of high-resolution face images of the same class, and an inter-class edge consists of a pair of high-resolution face images of different classes; 基于密集图构建稀疏图来降低边复杂度，该稀疏图中对每个同类别的高分辨率人脸增加一个质心顶点，一条类内边由一对同类别的高分辨率人脸图像组成，一条类间边由一个高分辨率人脸图像及与之不同类别的一个质心顶点组成；Construct a sparse graph based on a dense graph to reduce edge complexity. In this sparse graph, a centroid vertex is added to each high-resolution face of the same category, and an intra-class edge is composed of a pair of high-resolution face images of the same category. An inter-class edge consists of a high-resolution face image and a centroid vertex of a different class; 对稀疏图进行优化求解，筛选出富含身份信息的高分辨率人脸特征。Optimize and solve the sparse graph, and screen out high-resolution face features rich in identity information. 3.如权利要求2所述的方法，其特征在于，对稀疏图进行优化求解时，将人脸顶点和质心顶点表示成特征向量，基于该特征向量将富含身份信息的高分辨率人脸深度特征表示成二值标记式，对该二值标记式进行优化求解。3. The method according to claim 2, wherein when the sparse graph is optimized and solved, the face vertices and the centroid vertices are represented as feature vectors, based on the feature vectors, the high-resolution face rich in identity information Depth features are represented as binary markers, and the binary markers are optimized and solved. 4.如权利要求3所述的方法，其特征在于，采用图割算法对稀疏图进行优化求解。4. The method according to claim 3, wherein a graph cut algorithm is used to optimize and solve the sparse graph. 5.如权利要求1所述的方法，其特征在于，所述学生流深度学习网络接收到低分辨率人脸图像后，先产生高分辨率人脸深度特征，再经过压缩产生低分辨率人脸深度特征。5. The method according to claim 1, wherein after the student flow deep learning network receives the low-resolution face image, it first generates high-resolution face depth features, and then compresses to generate low-resolution face images. Face depth features. 6.如权利要求5所述的方法，其特征在于，对学生流深度学习网络进行微调训练的方法是：将学生流深度学习网络接收带身份标记的低分辨率人脸图像所产生的高分辨率人脸深度特征通过特征回归逼近所述富含身份信息的高分辨率人脸深度特征，再依据由该高分辨率人脸深度特征压缩后的低分辨率人脸深度特征通过身份分类对该带身份标记的低分辨率人脸图像进行分类。6. The method according to claim 5, wherein the method for fine-tuning the student flow deep learning network is: the student flow deep learning network receives the high-resolution facial image produced by the low-resolution face image with identity mark The high-resolution face depth features are approached by feature regression to the high-resolution face depth features rich in identity information, and then the low-resolution face depth features compressed by the high-resolution face depth features are classified by identity. Identity-labeled low-resolution face images for classification. 7.如权利要求6所述的方法，其特征在于，所述特征回归的回归损失函数和身份分类的分类损失函数共同组成总的损失函数，通过对总的损失函数进行求解，获得训练好的学生流深度学习网络。7. The method according to claim 6, wherein the regression loss function of the feature regression and the classification loss function of the identity classification jointly form a total loss function, and by solving the total loss function, the trained Student Stream Deep Learning Networks. 8.如权利要求1所述的方法，其特征在于，所述低分辨率人脸深度特征用于人脸比对时，对用于人脸比对的两幅低分辨率人脸图像进行特征相似性计算，如果相似性超过一阈值，则判定是同一人；所述低分辨率人脸深度特征用于人脸认证时，则对低分辨率人脸深度特征进行分类，以确定身份。8. The method according to claim 1, wherein when the low-resolution face depth feature is used for face comparison, two low-resolution face images for face comparison are characterized Similarity calculation, if the similarity exceeds a threshold, it is determined that they are the same person; when the low-resolution face depth features are used for face authentication, the low-resolution face depth features are classified to determine the identity. 9.一种低分辨率人脸识别装置，包括基于学生流深度学习网络的特征蒸馏模块、特征压缩模块和特征识别模块；9. A low-resolution face recognition device, comprising a feature distillation module, a feature compression module, and a feature recognition module based on a student flow deep learning network; 所述特征蒸馏模块用于获取从教师流深度学习网络传递的富含身份信息的高分辨率人脸深度特征；The feature distillation module is used to obtain the high-resolution face depth features rich in identity information delivered from the teacher's deep learning network; 所述特征压缩模块用于特征降维，得到具有鉴别能力的低分辨率人脸深度特征；The feature compression module is used for feature dimensionality reduction to obtain low-resolution human face depth features with discriminative ability; 所述特征识别模块用于低分辨率人脸识别，为一特征比对器或一含Softmax层的子网络。The feature recognition module is used for low-resolution face recognition, and is a feature comparer or a sub-network containing a Softmax layer. 10.如权利要求9所述的装置，其特征在于，所述学生流深度学习网络包括10个卷积层、3个最大池化层和3个全连接层，该卷积层包括若干3×3的卷积层和位于该若干3×3的卷积层之间或者3×3的卷积层与最大池化层之间的用于压缩特征图通道数目的若干1×1的卷积层，该最大池化层用于降低特征维度，该3个全连接层为镜像层、身份层及Softmax层。10. The device according to claim 9, wherein the student flow deep learning network includes 10 convolutional layers, 3 maximum pooling layers, and 3 fully connected layers, and the convolutional layers include several 3× 3 convolutional layers and several 1×1 convolutional layers between the 3×3 convolutional layers or between the 3×3 convolutional layers and the maximum pooling layer for compressing the number of feature map channels , the maximum pooling layer is used to reduce the feature dimension, and the three fully connected layers are mirror layer, identity layer and Softmax layer.
CN201810698280.6A 2018-06-29 2018-06-29 Low-resolution face recognition method and device Active CN109063565B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810698280.6A CN109063565B (en) 2018-06-29 2018-06-29 Low-resolution face recognition method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810698280.6A CN109063565B (en) 2018-06-29 2018-06-29 Low-resolution face recognition method and device

Publications (2)

Publication Number Publication Date
CN109063565A true CN109063565A (en) 2018-12-21
CN109063565B CN109063565B (en) 2021-12-10

Family

ID=64818451

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810698280.6A Active CN109063565B (en) 2018-06-29 2018-06-29 Low-resolution face recognition method and device

Country Status (1)

Country Link
CN (1) CN109063565B (en)

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109816636A (en) * 2018-12-28 2019-05-28 汕头大学 A crack detection method based on intelligent terminal
CN109919110A (en) * 2019-03-13 2019-06-21 北京航空航天大学 Video attention area detection method, device and equipment
CN109948555A (en) * 2019-03-21 2019-06-28 于建岗 Human face super-resolution recognition methods based on video flowing
CN109977792A (en) * 2019-03-04 2019-07-05 上海商汤智能科技有限公司 Face characteristic compression method and device
CN110263756A (en) * 2019-06-28 2019-09-20 东北大学 A kind of human face super-resolution reconstructing system based on joint multi-task learning
CN110659665A (en) * 2019-08-02 2020-01-07 深圳力维智联技术有限公司 A method for constructing a model of different-dimensional features, and a method and device for image recognition
CN110674688A (en) * 2019-08-19 2020-01-10 深圳力维智联技术有限公司 Method, system and medium for obtaining face recognition model for video surveillance scene
CN111160533A (en) * 2019-12-31 2020-05-15 中山大学 Neural network acceleration method based on cross-resolution knowledge distillation
CN111507135A (en) * 2019-01-31 2020-08-07 北京奇虎科技有限公司 Face detection method and device, computer equipment and storage medium
CN111553333A (en) * 2020-07-10 2020-08-18 支付宝(杭州)信息技术有限公司 Face image recognition model training method, recognition method, device and electronic equipment
CN111753793A (en) * 2020-06-30 2020-10-09 重庆紫光华山智安科技有限公司 Model training method and device, face screening method and electronic equipment
CN113222813A (en) * 2021-04-19 2021-08-06 西安电子科技大学 Image super-resolution reconstruction method and device, electronic equipment and storage medium
CN113435334A (en) * 2021-06-28 2021-09-24 中国科学院上海微系统与信息技术研究所 Small target face recognition method based on deep learning
CN113705532A (en) * 2021-09-10 2021-11-26 中国人民解放军国防科技大学 Target detection method, device and equipment based on medium-low resolution remote sensing image
CN114090986A (en) * 2020-07-31 2022-02-25 华为技术有限公司 Method for identifying user on public equipment and electronic equipment
CN116452424A (en) * 2023-05-19 2023-07-18 山东大学 Face super-resolution reconstruction method and system based on double generalized distillation

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105787462A (en) * 2016-03-16 2016-07-20 武汉工程大学 Semi-coupling-crucial-dictionary-learning-based extremely-low-resolution face identification method and system
WO2016150240A1 (en) * 2015-03-24 2016-09-29 北京天诚盛业科技有限公司 Identity authentication method and apparatus
CN106033540A (en) * 2016-05-27 2016-10-19 北京大学第医院 A method and system for automatic analysis of vaginal microecological morphology
CN106203256A (en) * 2016-06-24 2016-12-07 山东大学 A kind of low resolution face identification method based on sparse holding canonical correlation analysis
CN107833183A (en) * 2017-11-29 2018-03-23 安徽工业大学 A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring
CN107944379A (en) * 2017-11-20 2018-04-20 中国科学院自动化研究所 White of the eye image super-resolution rebuilding and image enchancing method based on deep learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2016150240A1 (en) * 2015-03-24 2016-09-29 北京天诚盛业科技有限公司 Identity authentication method and apparatus
CN105787462A (en) * 2016-03-16 2016-07-20 武汉工程大学 Semi-coupling-crucial-dictionary-learning-based extremely-low-resolution face identification method and system
CN106033540A (en) * 2016-05-27 2016-10-19 北京大学第医院 A method and system for automatic analysis of vaginal microecological morphology
CN106203256A (en) * 2016-06-24 2016-12-07 山东大学 A kind of low resolution face identification method based on sparse holding canonical correlation analysis
CN107944379A (en) * 2017-11-20 2018-04-20 中国科学院自动化研究所 White of the eye image super-resolution rebuilding and image enchancing method based on deep learning
CN107833183A (en) * 2017-11-29 2018-03-23 安徽工业大学 A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring

Cited By (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109816636A (en) * 2018-12-28 2019-05-28 汕头大学 A crack detection method based on intelligent terminal
CN109816636B (en) * 2018-12-28 2020-11-27 汕头大学 A crack detection method based on intelligent terminal
CN111507135A (en) * 2019-01-31 2020-08-07 北京奇虎科技有限公司 Face detection method and device, computer equipment and storage medium
CN109977792A (en) * 2019-03-04 2019-07-05 上海商汤智能科技有限公司 Face characteristic compression method and device
CN109919110A (en) * 2019-03-13 2019-06-21 北京航空航天大学 Video attention area detection method, device and equipment
CN109919110B (en) * 2019-03-13 2021-06-04 北京航空航天大学 Video attention area detection method, device and equipment
CN109948555B (en) * 2019-03-21 2020-11-06 于建岗 Face super-resolution identification method based on video stream
CN109948555A (en) * 2019-03-21 2019-06-28 于建岗 Human face super-resolution recognition methods based on video flowing
CN110263756A (en) * 2019-06-28 2019-09-20 东北大学 A kind of human face super-resolution reconstructing system based on joint multi-task learning
CN110659665A (en) * 2019-08-02 2020-01-07 深圳力维智联技术有限公司 A method for constructing a model of different-dimensional features, and a method and device for image recognition
CN110659665B (en) * 2019-08-02 2023-09-29 深圳力维智联技术有限公司 Model construction method of different-dimension characteristics and image recognition method and device
CN110674688A (en) * 2019-08-19 2020-01-10 深圳力维智联技术有限公司 Method, system and medium for obtaining face recognition model for video surveillance scene
CN111160533A (en) * 2019-12-31 2020-05-15 中山大学 Neural network acceleration method based on cross-resolution knowledge distillation
CN111160533B (en) * 2019-12-31 2023-04-18 中山大学 Neural network acceleration method based on cross-resolution knowledge distillation
CN111753793B (en) * 2020-06-30 2022-11-22 重庆紫光华山智安科技有限公司 Model training method and device, face screening method and electronic equipment
CN111753793A (en) * 2020-06-30 2020-10-09 重庆紫光华山智安科技有限公司 Model training method and device, face screening method and electronic equipment
CN111553333A (en) * 2020-07-10 2020-08-18 支付宝(杭州)信息技术有限公司 Face image recognition model training method, recognition method, device and electronic equipment
CN111553333B (en) * 2020-07-10 2020-10-16 支付宝(杭州)信息技术有限公司 Face image recognition model training method, recognition method, device and electronic equipment
CN114090986A (en) * 2020-07-31 2022-02-25 华为技术有限公司 Method for identifying user on public equipment and electronic equipment
CN113222813A (en) * 2021-04-19 2021-08-06 西安电子科技大学 Image super-resolution reconstruction method and device, electronic equipment and storage medium
CN113222813B (en) * 2021-04-19 2024-02-09 西安电子科技大学 Image super-resolution reconstruction method and device, electronic equipment and storage medium
CN113435334A (en) * 2021-06-28 2021-09-24 中国科学院上海微系统与信息技术研究所 Small target face recognition method based on deep learning
CN113435334B (en) * 2021-06-28 2024-02-27 中国科学院上海微系统与信息技术研究所 Small target face recognition method based on deep learning
CN113705532A (en) * 2021-09-10 2021-11-26 中国人民解放军国防科技大学 Target detection method, device and equipment based on medium-low resolution remote sensing image
CN113705532B (en) * 2021-09-10 2023-05-23 中国人民解放军国防科技大学 Target detection method, device and equipment based on medium-low resolution remote sensing image
CN116452424A (en) * 2023-05-19 2023-07-18 山东大学 Face super-resolution reconstruction method and system based on double generalized distillation
CN116452424B (en) * 2023-05-19 2023-10-10 山东大学 A face super-resolution reconstruction method and system based on double generalized distillation

Also Published As

Publication number Publication date
CN109063565B (en) 2021-12-10

Similar Documents

Publication Publication Date Title
CN109063565A (en) A kind of low resolution face identification method and device
CN111783576B (en) Pedestrian re-identification method based on improved YOLOv3 network and feature fusion
CN104143079B (en) The method and system of face character identification
CN102799870B (en) Based on the single training image per person method of the consistent LBP of piecemeal and sparse coding
CN111597870B (en) Human body attribute identification method based on attention mechanism and multi-task learning
CN109886072B (en) Face attribute classification system based on bidirectional Ladder structure
CN105426850A (en) Human face identification based related information pushing device and method
CN108875602A (en) Monitor the face identification method based on deep learning under environment
CN111539255A (en) Cross-modal pedestrian re-identification method based on multi-modal image style conversion
CN108629338A (en) A kind of face beauty prediction technique based on LBP and convolutional neural networks
CN116912949B (en) A gait recognition method based on perspective-aware part-wise attention mechanism
WO2019153175A1 (en) Machine learning-based occluded face recognition system and method, and storage medium
CN106778796A (en) Human motion recognition method and system based on hybrid cooperative model training
Yang et al. Recognizing clothes patterns for blind people by confidence margin based feature combination
CN104392233B (en) A kind of image saliency map extracting method based on region
CN116311483A (en) Micro-expression Recognition Method Based on Partial Facial Region Reconstruction and Memory Contrastive Learning
CN104361348A (en) Flower and plant recognition method on intelligent terminal
CN104966052A (en) Attributive characteristic representation-based group behavior identification method
CN116704611A (en) A cross-view gait recognition method based on motion feature mixture and fine-grained multi-stage feature extraction
CN105160290A (en) Mobile boundary sampling behavior identification method based on improved dense locus
CN118116082A (en) Gait recognition method based on space-time feature and multi-scale time sequence feature extraction
CN115100681B (en) A clothing recognition method, system, medium and device
CN110633631B (en) Pedestrian re-identification method based on component power set and multi-scale features
CN115909488A (en) An occluded person re-identification method based on pose guidance and dynamic feature extraction
Yan et al. Person re-identification with vision and language

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