CN112232147A - Method, device and system for face model hyper-parameter adaptive acquisition - Google Patents
Method, device and system for face model hyper-parameter adaptive acquisition Download PDFInfo
- Publication number
- CN112232147A CN112232147A CN202011041737.XA CN202011041737A CN112232147A CN 112232147 A CN112232147 A CN 112232147A CN 202011041737 A CN202011041737 A CN 202011041737A CN 112232147 A CN112232147 A CN 112232147A
- Authority
- CN
- China
- Prior art keywords
- loss
- characteristic diagram
- convolution
- obtaining
- hyper
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 52
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 32
- 238000010586 diagram Methods 0.000 claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 27
- 230000006870 function Effects 0.000 claims description 30
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 abstract description 15
- 238000013461 design Methods 0.000 abstract description 6
- 238000002474 experimental method Methods 0.000 abstract description 5
- 210000004027 cell Anatomy 0.000 description 17
- 238000013528 artificial neural network Methods 0.000 description 15
- 238000005070 sampling Methods 0.000 description 6
- 210000003128 head Anatomy 0.000 description 5
- 230000004931 aggregating effect Effects 0.000 description 4
- 238000001514 detection method Methods 0.000 description 4
- 210000000887 face Anatomy 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000009792 diffusion process Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000004927 fusion Effects 0.000 description 3
- 238000005286 illumination Methods 0.000 description 3
- 230000004044 response Effects 0.000 description 3
- 241000499037 Vanilla Species 0.000 description 2
- 235000009499 Vanilla fragrans Nutrition 0.000 description 2
- 235000012036 Vanilla tahitensis Nutrition 0.000 description 2
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000003333 near-infrared imaging Methods 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 239000013598 vector Substances 0.000 description 2
- 244000030973 Vanilla pompona Species 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000001815 facial effect Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
-
- 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
- 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Human Computer Interaction (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)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Collating Specific Patterns (AREA)
Abstract
The invention discloses a method for face model hyper-parameter adaptive acquisition, which comprises the following steps: step S1: obtaining an output characteristic diagram according to the central difference convolution function; step S2: calculating and obtaining the loss of the output characteristic diagram on the training set; step S3: and obtaining the self-adaptive central difference convolution function hyperparameters according to the loss of the output characteristic graph on the training set. The super-parameters are obtained through self-learning in the training process, so that the influence of subjectivity of designers caused by manual design of the super-parameters is avoided, the face quality and the performance of a face recognition or face anti-counterfeiting model are improved, the deployment and use cost of the model are reduced, and the experiment time is saved. The invention also discloses a device and a system for the face model hyper-parameter adaptive acquisition.
Description
Technical Field
The application relates to the technical field of deep learning, in particular to a method, a device and a system for face model hyper-parameter adaptive acquisition.
Background
The quality of human face, the anti-counterfeiting of human face or the recognition of human face are all very important components in the human face recognition system. In the process of deep neural network modeling on face quality, face anti-counterfeiting or face recognition, the setting of the hyper-parameters has great influence on the final result of the model.
In the process of implementing the embodiments of the present disclosure, it is found that at least the following problems exist in the related art: most of the existing face quality, face anti-counterfeiting or face recognition methods rely on common convolution kernels and neural networks manually designed by experts, detailed fine-grained information cannot be described, and when the environment changes, for example, for different illuminations, the manually designed neural networks are used, so that the face quality, face anti-counterfeiting or face recognition methods are difficult to achieve a very good performance.
Disclosure of Invention
The following presents a simplified summary in order to provide a basic understanding of some aspects of the disclosed embodiments. The foregoing summary is not an extensive overview and is intended to neither identify key/critical elements nor delineate the scope of such embodiments, but is intended to be a prelude to the more detailed description that is presented later.
The embodiment of the disclosure provides a method, a device and a system for face model hyper-parameter adaptive acquisition, so as to solve the aforementioned technical problems to a certain extent.
In some embodiments, a method for hyper-parametric adaptive acquisition of a face model, comprising: step S1: obtaining an output characteristic diagram according to the central difference convolution function; step S2: calculating and obtaining the loss of the output characteristic diagram on a training set; step S3: and obtaining a self-adaptive central difference convolution function hyperparameter according to the loss of the output characteristic diagram on the training set.
Optionally, the step S2 further includes: step S21: calculating and obtaining the loss of the output characteristic diagram on the Support data set; step S22: correcting the convolution kernel parameters according to the loss of the output characteristic diagram on the Support data set to obtain corrected convolution kernel parameters; step S23: and calculating and obtaining the loss of the output characteristic diagram on the query data set according to the corrected convolution kernel parameters.
Optionally, the loss of the output feature map on the Support data set is obtained according to the output feature map and a Softmax loss function.
Optionally, the modified convolution kernel parameter ω' is obtained according to the following formula:
Optionally, the loss of the output feature map on the query data set is obtained according to the output feature map, the modified convolution kernel parameter and the Softmax loss function.
Optionally, the adaptive central difference convolution function hyper-parameter θ' is obtained according to the following formula:
Optionally, the central differential convolution function Fcdc(x,ω|θ)，
Wherein p is0For the current position, R is the local neighborhood, pnIs the index of local neighborhood, omega is the parameter of convolution kernel, x is the input characteristic diagram, and the hyper-parameter theta is belonged to [0,1 ]]。
In some embodiments, an apparatus for hyper-parametric adaptive acquisition of a face model, comprises: the convolution module is used for obtaining an output characteristic diagram according to the central difference convolution function; the calculation module is used for calculating and obtaining the loss of the output characteristic diagram on the training set; and the output module is used for obtaining the self-adaptive hyper-parameter according to the loss of the output characteristic diagram on the training set.
In some embodiments, a system for hyper-parametric adaptive acquisition of a face model comprises a processor and a memory storing program instructions, the processor being configured to, upon execution of the program instructions, perform the aforementioned method for hyper-parametric adaptive acquisition of a face model.
In some embodiments, a computer-readable storage medium having stored thereon a program for face model hyper-parametric adaptive acquisition, which when executed by the processor, implements the aforementioned method for face model hyper-parametric adaptive acquisition.
The method, the device, the system and the computer readable storage medium for face model hyper-parameter adaptive acquisition provided by the embodiment of the disclosure can realize the following technical effects:
the central difference convolution function hyperparameters are obtained through self-learning in the training process, the neural network for automatically adjusting the central difference degree is automatically adjusted in the process of executing specific tasks, more characteristics on the face are captured through summarizing the inherent detailed mode strength and gradient information, the influence of the subjectivity of designers caused by manual design hyperparameters is avoided, the face quality, the face recognition performance or the face anti-counterfeiting model performance is improved, meanwhile, the deployment and use cost of the model is reduced, and the experiment time is saved.
The foregoing general description and the following description are exemplary and explanatory only and are not restrictive of the application.
Drawings
One or more embodiments are illustrated by way of example in the accompanying drawings, which correspond to the accompanying drawings and not in limitation thereof, in which elements having the same reference numeral designations are shown as like elements and not in limitation thereof, and wherein:
FIG. 1 is a flow chart of a method for face model hyper-parameter adaptive acquisition provided by an embodiment of the present disclosure;
FIG. 2 is a schematic diagram of a central differential convolution provided by an embodiment of the present disclosure;
FIG. 3 is a schematic diagram of a search space of a central differential convolution architecture provided by an embodiment of the present disclosure;
FIG. 4 is a structural diagram of a multi-scale attention fusion module provided by an embodiment of the present disclosure;
FIG. 5 is a schematic diagram of a central differential convolution network incorporating a multi-attention module according to an embodiment of the present disclosure;
fig. 6 is a schematic diagram of the difference between a normal convolution kernel and a central differential convolution kernel in false cue feature capture provided by the embodiment of the present disclosure.
Detailed Description
So that the manner in which the features and elements of the disclosed embodiments can be understood in detail, a more particular description of the disclosed embodiments, briefly summarized above, may be had by reference to the embodiments, some of which are illustrated in the appended drawings. In the following description of the technology, for purposes of explanation, numerous details are set forth in order to provide a thorough understanding of the disclosed embodiments. However, one or more embodiments may be practiced without these details. In other instances, well-known structures and devices may be shown in simplified form in order to simplify the drawing.
It is obvious that the drawings in the following description are only examples or embodiments of the present application, and that it is also possible for a person skilled in the art to apply the present application to other similar contexts on the basis of these drawings without inventive effort. Moreover, it should be appreciated that in the development of any such actual implementation, as in any engineering or design project, numerous implementation-specific decisions must be made to achieve the developers' specific goals, such as compliance with system-related and business-related constraints, which may vary from one implementation to another.
Reference in the specification to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the specification. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Those of ordinary skill in the art will explicitly and implicitly appreciate that the embodiments described herein may be combined with other embodiments without conflict.
Unless defined otherwise, technical or scientific terms referred to herein shall have the ordinary meaning as understood by those of ordinary skill in the art to which this application belongs. Reference to "a," "an," "the," and similar words throughout this application are not to be construed as limiting in number, and may refer to the singular or the plural. The present application is directed to the use of the terms "including," "comprising," "having," and any variations thereof, which are intended to cover non-exclusive inclusions; for example, a process, method, system, article, or apparatus that comprises a list of steps or modules (elements) is not limited to the listed steps or elements, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus. Reference to "connected," "coupled," and the like in this application is not intended to be limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The term "plurality" as referred to herein means two or more. "and/or" describes an association relationship of associated objects, meaning that three relationships may exist, for example, "A and/or B" may mean: a exists alone, A and B exist simultaneously, and B exists alone. The character "/" generally indicates that the former and latter associated objects are in an "or" relationship. Reference herein to the terms "first," "second," "third," and the like, are merely to distinguish similar objects and do not denote a particular ordering for the objects.
The technical terms referred to in the present application are explained below.
The human face quality refers to the quality comprehensive evaluation of a detected target human face, and quantitative indexes are obtained, wherein the quantitative indexes mainly comprise human face fuzzy degree, human face illumination brightness, human face posture and the like; the method makes a pre-judgment on whether business applications such as face analysis can be directly carried out or not, improves the accuracy of face analysis, and provides guarantee for avoiding misjudgment.
Face recognition refers to a biometric technology that performs identification based on facial feature information of a person. A series of related technologies, also commonly called face recognition and face recognition, are used to collect images or video streams containing faces by using a camera or a video camera, automatically detect and track the faces in the images, and then perform face recognition on the detected faces.
For example, in a face recognition system, some forged faces are manufactured, for example, videos, photos or masks attack the face recognition system, and the face anti-counterfeiting recognition is to discriminate the forged face features. Generally, in a face recognition system, a user needs to perform face anti-counterfeiting recognition detection, which may also be called live body detection.
The center difference is a finite difference method. The finite difference method optimizes the approximation of the differential operator in the plaque center node under consideration and provides a numerical solution for the differential equation. The center difference scheme is one of schemes for solving the integral convection diffusion equation and calculating the transmission characteristics Φ on the e and w planes. The advantage of this method is that it is easy to understand and implement, at least for simple material relations, and converges faster than some other finite difference methods, such as forward and backward difference methods. To the right of the convection-diffusion equation, the diffusion term is substantially highlighted and can be approximated using the center difference. Thus, to simplify the solution and analysis, linear interpolation can be used logically to compute the cell face value on the left hand side of the equation, which is simply a convection term.
The inventor finds in research that in the prior art, a forged face attack is resisted through near infrared information, but in some devices, systems or scenes, a near infrared device cannot be used, and the forged face attack cannot be effectively resisted, for example: the attack of counterfeiting the face is resisted by judging whether the face can be detected in the near-infrared imaging and the difference between the face and eyes in the visible light imaging and the near-infrared imaging. However, the prior art does not protect well against other modalities than near infrared, such as visible light modalities or depth modalities.
In the prior art, the performance of face recognition is also improved by designing some loss functions, for example: step S1: training a recognition network through a face recognition training set; step S2: adopting a trained recognition network as a feature extraction module, and training an uncertainty module through the same training set; step S3: aggregating the input video feature set by using the learned uncertainty as the importance degree of the features to obtain aggregated features; step S4: and comparing the aggregated features by using the mutual likelihood fraction to complete the final recognition. A common convolution kernel is used.
In the prior art, the human face is divided into a plurality of regional sets so as to extract finer features, but a neural network which is still designed manually is used. For example: the method comprises the steps of carrying out face detection on an obtained picture, obtaining a face frame image, obtaining face key point coordinates according to feature point detection positioning, intercepting to obtain a face feature region set, respectively inputting the face feature region set into corresponding feature neural networks for feature extraction, fusing extracted feature vectors to obtain a face overall feature vector, and calculating by adopting cosine similarity to obtain a face recognition result. The neural networks used are still manually designed and it is difficult to achieve more extreme performance.
The following description is made of the English fields referred to in the drawings of the present application, and in FIG. 2: inputting a characteristic inputfeaturemap, outputting a characteristic outputfeaturemap, sampling, aggregating and expanding; in fig. 3: stem node stem, head node head, Low-dimensional node group Low-level Cell, medium-dimensional node group Mid-level Cell, High-dimensional node group High-level Cell, operation space and jumper operation Skip-Connect; in fig. 4: combining Concat, Spatial focusing Spatial Attenttion, large core Largekernel, middle core Midkkernel, small core Lowkernel, High-level feature, middle-level feature Mid-level feature, and Low-level feature; in fig. 6: the Print paper attacks Print attach, naked Convolution or original Convolution Vanilla Convolition, consistent consistency, Inconsistent Inconstant.
Fig. 1 is a flowchart of a method for face model hyper-parameter adaptive acquisition according to an embodiment of the present disclosure. As shown in fig. 1, an embodiment of the present disclosure provides a method for obtaining a face model by hyper-parameter adaptive method, including: step S1: obtaining an output characteristic diagram according to a Central Difference Convolution (CDC) function; step S2: calculating and obtaining the loss of the output characteristic diagram on the training set; step S3: and obtaining the self-adaptive central difference convolution function hyperparameters according to the loss of the output characteristic graph on the training set. Therefore, the center difference convolution function hyperparameter is obtained through self-learning in the training process, the neural network for automatically adjusting the center difference degree is automatically adjusted in the process of executing a specific task, more characteristics on the face are captured through summarizing the inherent detailed mode strength and gradient information, the influence of the subjectivity of designers caused by manual design hyperparameter is avoided, the face quality, the face recognition performance or the face anti-counterfeiting model performance is improved, meanwhile, the deployment and use cost of the model is reduced, and the experiment time is saved.
The central differential convolution in the embodiments of the present disclosure includes: sampling and summarizing, wherein the sampling is similar to the steps in the prior convolution, and the detailed description is omitted. Fig. 2 is a schematic diagram of a central differential convolution provided by an embodiment of the present disclosure. As shown in fig. 2, the center difference convolution includes: sampling and aggregating aggregation, wherein input characteristic inputfeaturemap expansion and expansion are subjected to central differential convolution centraldifference Conv to obtain output characteristic outputfeaturemap, the central differential convolution is more prone to aggregating convolution sampling gradients facing to the center, and the proportion of the sampling of the aggregation surface to the central convolution can be automatically adjusted by self-adaptive central differential convolution in the training process. In this way, convolutional neural network side representation and generalization capability can be enhanced.
In some embodiments, step S2 further includes: step S21: calculating and obtaining the loss of the output characteristic diagram on the Support data set; step S22: correcting the convolution kernel parameters according to the loss of the output characteristic diagram on the Support data set to obtain corrected convolution kernel parameters; step S23: and calculating and obtaining the loss of the output characteristic diagram on the query data set according to the corrected convolution kernel parameters. The Support data is derived from a training set, and the data on the training set can be divided into a Support data set and a query data set according to different rules according to different tasks. For example, in a face anti-counterfeiting task, the problem of inter-domain differences (domainshift) in new and old attack modes needs to be learned, the Support data set may be a data set including a plurality of old attack modes, and the query data set may be a data set including a new attack mode.
In some embodiments, the loss of the output profile over the Support dataset is obtained from the output profile and a Softmax loss function. Loss of output feature maps on support datasets
Where S is the Support dataset, τiFor a set of tasks in a dataset, Softmax is a loss function, yiTo output a characteristic map, FcdcAs a central differential convolution function, xiFor inputting a characteristic diagram, omega is a convolution kernel parameter, and a hyper-parameter theta is in an element of 0,1]。τiThe task set in the data set is represented, and the task is also a data structure and belongs to a subset of a Support data set and a query data set.
In some embodiments, the modified convolution kernel parameter ω ═ ω -
Wherein, omega is a convolution kernel parameter,
is an outputLoss of feature maps on the Support dataset.
In some embodiments, the loss of the output feature map on the query data set is obtained according to the output feature map, the modified convolution kernel parameters and the Softmax loss function. Loss of output profile on query dataset
Wherein q is a query data set.
In some embodiments, the adaptive center difference convolution function hyperparameters are obtained according to the following formula
Wherein the over-parameter theta belongs to [0,1 ]]，
Is the loss of the output profile on the query dataset.
In some embodiments, the central differential convolution function Fcdc(x,ω|θ)，
Wherein p is0For the current position, R is the local neighborhood, pnIs the index of local neighborhood, omega is the parameter of convolution kernel, x is the input characteristic diagram, and the hyper-parameter theta is belonged to [0,1 ]]. For a central difference convolution function FcdcThe input variable is the input characteristic diagram x, and the value of the convolution kernel parameter omega in the hyper-parameter set theta is FcdcThe output of (2) is an output signature. The formula in the form can make the application of the central differential convolution in the deep neural network model more efficient. FcdcCan be derived from the following formula, the detailed derivation process is not described herein,
wherein, when theta is 0, FcdcBeing a common convolution kernel, e.g. vanilla convolution, F is the value of 1 when θcdcOnly differential information is utilized. In the prior art, a designer sets the hyper-parameter theta according to experience, subjective content of the designer is inevitably introduced, in the embodiment of the disclosure, theta is ideal data obtained by learning in a training process, so that the performance of the obtained model is improved on specific human face quality, human face anti-counterfeiting or human face recognition tasks, the deployment and use cost of the model can be reduced, and the experiment time can be saved.
The embodiment of the disclosure provides a neural network method for adaptive center differential convolution, which is used for searching a network backbone consisting of low-level, middle-level and high-level units for face anti-counterfeiting, face quality or face recognition tasks. Wherein one output of the latest input cell is taken as the input of the current cell. Inspired by hierarchically organizing dedicated neurons in the human visual system, multi-level cells with different structures are searched. The self-adaptive central differential convolution searches the neural network method taking the self-adaptive central differential convolution as a basic unit aiming at different tasks such as human face quality, human face recognition or human face anti-counterfeiting, so that the searching is more flexible, and the freedom degree of the architecture is improved. On a specially designed adaptive central differential convolution Search space, Neural Architecture Search (NAS) can find more contents.
Fig. 3 is a schematic diagram of a search space of a central differential convolution architecture provided by an embodiment of the present disclosure. As shown in a network in fig. 3(a), a target is to search three levels of units of a Low-dimensional node group Low-level Cell, a medium-dimensional node group Mid-level Cell and a High-dimensional node group High-level Cell, one network consists of the three levels of units which are spliced in a laminated manner, and further comprises an input and an output, Stem nodes Stem0 and Stem1, Head nodes Head0 and Head1, wherein the hyper-parameters are obtained by learning in a training process, and artificial intelligence is avoidedThe impact of the subjectivity of the person involved in the design. As shown in the cell of FIG. 3(B), each cell includes six nodes, an input node, four intermediate nodes B1, B2, B3, and B4, and an output node output, and each cell can be represented as a directed acyclic graph of N nodes, βiIs a loss function of the initial learnable weight of the intermediate node. As shown in the operation space operationspace of fig. 3(c), an edge between two nodes node1 and node2 except for the output node represents one possible operation, with eight candidate operations including: the operations CDC, CDC _2_1.2, CDC _2_1.4, CDC _2_1.6, CDC _2_1.8 and CDC _2_2.0 related to the no operation none, the Skip-Connect operation and the six central differential convolutions, wherein CDC _2_ r represents that the two-stack central differential convolution is used to increase the number of channels by the proportion of r and then decrease the number of channels to the original number of channels. The size of the search space is 3 x 8(1+2+3+4)＝3*108。
In some embodiments, multiple levels of attention modules are added to the network backbone, integrating multiple levels of features. Fig. 4 is a structural diagram of a multi-scale attention fusion module provided in an embodiment of the present disclosure. As shown in FIG. 4, adjusting different levels of features of the Resize Low-dimensional feature Low-levelfeatures, medium-dimensional feature Mid-levelfeatureres and High-levelfeaturees through Spatial focusing and receiving domain-dependent Kernel sizes Large core Kernel, Medium core Mid Kernel and Small core Kernel, and then combining them with Concat, wherein MaxCaro is a maximum pooling layer, AvgPool is an average pooling layer, Convlayer is a convolutional layer, and sigmoid is an activation function of a neural network. For example, high-level or mid-level hierarchy features use small kernel sizes, and low-level hierarchies use large kernel sizes, reducing the impact of garbage. The powerful network structure searched in the self-adaptive central differential convolution search space is fused with the multi-scale attention fusion module, so that the performance can be further improved.
Fig. 5 is a schematic diagram of a central differential convolution network incorporating a multi-attention module according to an embodiment of the present disclosure. As shown in fig. 5, the central differential convolution network fused with a multiple Attention Module (MAFM for short) includes a central differential convolution search backbone network and a multiple Attention Module, and a maximum pool layer is behind each unit. The unit difference of different levels is large, and the Mid-dimensional node group Mid-level Cell has deeper layers. Inputting an RGB image, searching three-level units of a Low-dimensional node group Low-level Cell, a medium-dimensional node group Mid-level Cell and a High-dimensional node group High-level Cell through Stem nodes Stem0 and Stem1, fusing MAFM, and outputting a Depth Map Depth Map through Head nodes Head0 and Head 1.
The neural network with the self-adaptive central difference unit is searched in different tasks of face recognition, the trunk of the neural network changes according to the specific task, and the self-adaptive central difference convolution can learn different difference weights according to the specific task. The neural network of the self-adaptive central difference unit is applied to a face recognition task or a face quality task, and the searched network structure can be greatly changed. For example, a human face quality task generally requires that the model reasoning time is short, and a lightweight self-adaptive central difference network can be searched by taking a network with small calculation amount as a template.
Fig. 6 is a schematic diagram of the difference between a normal convolution kernel and a central differential convolution kernel in false cue feature capture provided by the embodiment of the present disclosure. As shown in fig. 6, the naked Convolution or the original Convolution may also be referred to as a Difference between the original Convolution (Vanilla Convolution) and the Central differential Convolution kernel (Central Difference Convolution) in the feature response to the inter-domain Difference, the response of the normal Convolution kernel and the Central differential Convolution kernel to the printing paper attack printack is different, for example, the feature response of the illumination and the camera to forge a human face, the normal Convolution kernel cannot capture a clear forged feature, which is an Inconsistent incongruent, but the Central differential Convolution kernel can extract a consistent consistant forged feature, which is a grid, for example.
The embodiment of the present disclosure provides a device for obtaining hyper-parameter self-adaption of a face model, including: the convolution module is used for obtaining an output characteristic diagram according to the central difference convolution function; the calculation module is used for calculating and obtaining the loss of the output characteristic diagram on the training set; and the output module is used for obtaining the self-adaptive hyper-parameter according to the loss of the output characteristic diagram on the training set.
The above modules may be functional modules or program modules, and may be implemented by software or hardware. For a module implemented by hardware, the modules may be located in the same processor; or the modules can be respectively positioned in different processors in any combination.
The disclosed embodiment provides a system for human face model hyper-parametric adaptive acquisition, which comprises a processor and a memory storing program instructions, wherein the processor is configured to execute the method for human face model hyper-parametric adaptive acquisition when executing the program instructions.
The embodiment of the present disclosure provides a computer-readable storage medium, on which a program for face model hyper-parameter adaptive acquisition is stored, and when the program for face model hyper-parameter adaptive acquisition is executed by a processor, the method for face model hyper-parameter adaptive acquisition is implemented.
The hyper-parameters are obtained through self-learning in the training process, the neural network of the central difference degree is automatically adjusted in the process of executing specific tasks, more characteristics on the face are captured through summarizing the inherent detailed mode intensity and gradient information, the influence of the subjectivity of designers caused by manual design of the hyper-parameters is avoided, the face quality, the face recognition performance or the face anti-counterfeiting model performance is improved, meanwhile, the deployment and use cost of the model is reduced, and the experiment time is saved.
The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The foregoing description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in other forms, so that those skilled in the art may apply the above-described modifications and variations to the present invention without departing from the spirit of the present invention.
Claims (10)
1. A method for hyper-parametric adaptive acquisition of a face model, comprising:
step S1: obtaining an output characteristic diagram according to the central difference convolution function;
step S2: calculating and obtaining the loss of the output characteristic diagram on a training set;
step S3: and obtaining a self-adaptive central difference convolution function hyperparameter according to the loss of the output characteristic diagram on the training set.
2. The method according to claim 1, wherein the step S2 further comprises:
step S21: calculating and obtaining the loss of the output characteristic diagram on the Support data set;
step S22: correcting the convolution kernel parameters according to the loss of the output characteristic diagram on the Support data set to obtain corrected convolution kernel parameters;
step S23: and calculating and obtaining the loss of the output characteristic diagram on the query data set according to the corrected convolution kernel parameters.
3. The method of claim 2, wherein the loss of the output feature map on the Support data set is obtained from the output feature map and a Softmax loss function.
4. The method of claim 3, wherein the modified convolution kernel parameter ω' is obtained according to the following equation:
5. The method of claim 4, wherein the loss of the output feature map on the query data set is obtained according to the output feature map, the modified convolution kernel parameters and a Softmax loss function.
6. The method of claim 5, wherein the adaptive center difference convolution function hyperparameter θ' is derived from the following equation:
7. The method of claim 1, wherein the central differential convolution function Fcdc(x, ω | θ) is:
wherein p is0For the current position, R is the local neighborhood, pnIs the index of local neighborhood, omega is the parameter of convolution kernel, x is the input characteristic diagram, and the hyper-parameter theta is belonged to [0,1 ]]。
8. An apparatus for hyper-parametric adaptive acquisition of a face model, comprising:
the convolution module is used for obtaining an output characteristic diagram according to the central difference convolution function;
the calculation module is used for calculating and obtaining the loss of the output characteristic diagram on the training set;
and the output module is used for obtaining the self-adaptive central difference convolution function hyperparameters according to the loss of the output characteristic graph on the training set.
9. A system for hyper-parametric adaptive acquisition of a face model, comprising a processor and a memory storing program instructions, characterized in that the processor is configured to perform the method for hyper-parametric adaptive acquisition of a face model according to any of claims 1 to 7 when executing the program instructions.
10. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored thereon a program for hyper-parametric adaptive acquisition of a face model, which when executed by a processor, implements the method for hyper-parametric adaptive acquisition of a face model according to any of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011041737.XA CN112232147A (en) | 2020-09-28 | 2020-09-28 | Method, device and system for face model hyper-parameter adaptive acquisition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011041737.XA CN112232147A (en) | 2020-09-28 | 2020-09-28 | Method, device and system for face model hyper-parameter adaptive acquisition |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112232147A true CN112232147A (en) | 2021-01-15 |
Family
ID=74119452
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011041737.XA Pending CN112232147A (en) | 2020-09-28 | 2020-09-28 | Method, device and system for face model hyper-parameter adaptive acquisition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112232147A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112906780A (en) * | 2021-02-08 | 2021-06-04 | 中国科学院计算技术研究所 | Fruit and vegetable image classification system and method |
| CN113033406A (en) * | 2021-03-26 | 2021-06-25 | 睿云联(厦门)网络通讯技术有限公司 | Face living body detection method and system based on depth separable circle center differential convolution |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103559693A (en) * | 2013-11-18 | 2014-02-05 | 东南大学 | Image local structure self-adaption recovery method based on non-continuity indicator |
| CN110781829A (en) * | 2019-10-28 | 2020-02-11 | 华北电力大学 | Light-weight deep learning intelligent business hall face recognition method |
-
2020
- 2020-09-28 CN CN202011041737.XA patent/CN112232147A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103559693A (en) * | 2013-11-18 | 2014-02-05 | 东南大学 | Image local structure self-adaption recovery method based on non-continuity indicator |
| CN110781829A (en) * | 2019-10-28 | 2020-02-11 | 华北电力大学 | Light-weight deep learning intelligent business hall face recognition method |
Non-Patent Citations (2)
| Title |
|---|
| ZITONG YU，ET AL: "《Searching Central Difference Convolutional Networks for Face Anti-Spoofing》", 《ARXIV:2003.04092V1》, pages 1 - 13 * |
| 吴建鑫: "《模式识别》", 《机械工业出版社》, pages: 285 - 286 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112906780A (en) * | 2021-02-08 | 2021-06-04 | 中国科学院计算技术研究所 | Fruit and vegetable image classification system and method |
| CN113033406A (en) * | 2021-03-26 | 2021-06-25 | 睿云联(厦门)网络通讯技术有限公司 | Face living body detection method and system based on depth separable circle center differential convolution |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Han et al. | CNNs-based RGB-D saliency detection via cross-view transfer and multiview fusion | |
| Lei Ba et al. | Predicting deep zero-shot convolutional neural networks using textual descriptions | |
| Zareian et al. | Learning visual commonsense for robust scene graph generation | |
| Liu et al. | Multi-objective convolutional learning for face labeling | |
| Cai et al. | New graph structured sparsity model for multi-label image annotations | |
| Tahir et al. | An open-ended continual learning for food recognition using class incremental extreme learning machines | |
| Chaabouni et al. | Transfer learning with deep networks for saliency prediction in natural video | |
| Esmaeili et al. | Fast-at: Fast automatic thumbnail generation using deep neural networks | |
| CN111783831A (en) | Complex image accurate classification method based on multi-source multi-label shared subspace learning | |
| CN112395442B (en) | Automatic identification and content filtering method for popular pictures on mobile internet | |
| Zamiri et al. | MVDF-RSC: Multi-view data fusion via robust spectral clustering for geo-tagged image tagging | |
| CN108960260B (en) | Classification model generation method, medical image classification method and medical image classification device | |
| Nie et al. | Perceptual attributes optimization for multivideo summarization | |
| Ning et al. | Conditional generative adversarial networks based on the principle of homologycontinuity for face aging | |
| CN112232147A (en) | Method, device and system for face model hyper-parameter adaptive acquisition | |
| Chen et al. | Engineering deep representations for modeling aesthetic perception | |
| Huang et al. | An evidential combination method with multi-color spaces for remote sensing image scene classification | |
| Liao et al. | FERGCN: facial expression recognition based on graph convolution network | |
| Joodi et al. | Increasing validation accuracy of a face mask detection by new deep learning model-based classification | |
| Anjali et al. | A novel based decision tree for content based image retrieval: an optimal classification approach | |
| Li et al. | Unsupervised domain adaptation via discriminative feature learning and classifier adaptation from center-based distances | |
| Patidar et al. | Performance analysis of artificial neural network and K Nearest neighbors image classification techniques with wavelet features | |
| Mallet et al. | Deepfake Detection Analyzing Hybrid Dataset Utilizing CNN and SVM | |
| Everett et al. | ProtoCaps: A Fast and Non-Iterative Capsule Network Routing Method | |
| Bhad et al. | Content based image retrieval a comparative based analysis for feature extraction approach |
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 |