CN108108662A - Deep neural network identification model and recognition methods - Google Patents
Deep neural network identification model and recognition methods Download PDFInfo
- Publication number
- CN108108662A CN108108662A CN201711209932.7A CN201711209932A CN108108662A CN 108108662 A CN108108662 A CN 108108662A CN 201711209932 A CN201711209932 A CN 201711209932A CN 108108662 A CN108108662 A CN 108108662A
- Authority
- CN
- China
- Prior art keywords
- matrix
- training
- pedestrian
- weight matrix
- full articulamentum
- 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
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
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present application discloses a kind of deep neural network identification model and recognition methods, and wherein method includes:Fine tuning training layer is trained the full articulamentum according to the original weight matrix and training set of full articulamentum, and obtains target weight matrix when training is completed;The full articulamentum extracts pedestrian's feature of target image according to the target weight matrix;The similarity between pedestrian's feature and each preset data of preset data concentration is obtained, obtains multiple similarities;The multiple similarity is ranked up, obtains the corresponding preset data of maximum similarity;When the maximum similarity is more than predetermined threshold value, the characteristic of pedestrian in the target image is determined according to the preset data.The embodiment of the present application can improve recognition speed and accuracy rate that pedestrian identifies again.
Description
Technical field
This application involves neural network algorithm fields, and in particular to a kind of deep neural network identification model and identification side
Method.
Background technology
Pedestrian identifies again has great application potential, its main purpose in fields such as security protection, criminal investigation and image retrievals
It is to judge whether the corresponding pedestrian of the image gathered in some camera is target person.However, the row in actual monitored environment
The accuracy rate of people's recognition methods again is relatively low, and therefore, how to improve the accuracy rate that pedestrian identifies again is that those skilled in the art wait to solve
Certainly the technical issues of.
The content of the invention
The embodiment of the present application proposes a kind of deep neural network identification model and recognition methods, can improve pedestrian and identify again
Recognition speed and accuracy rate.
In a first aspect, the embodiment of the present application provides a kind of deep neural network identification model, finely tune training layer and with institute
The full articulamentum of fine tuning training layer connection is stated, the full articulamentum is the neural network recognization model that training is completed, wherein:
The fine tuning training layer, for the original weight matrix according to the full articulamentum and training set to the full connection
Layer is trained, and obtains target weight matrix when training is completed;
The full articulamentum, for extracting pedestrian's feature of target image according to the target weight matrix;Described in acquisition
Pedestrian's feature and preset data concentrate the similarity between each preset data, obtain multiple similarities;To the multiple similar
Degree is ranked up, and obtains the corresponding preset data of maximum similarity;When the maximum similarity is more than predetermined threshold value, according to this
Preset data determines the characteristic of pedestrian in the target image.
With reference to first aspect, in the first possible embodiment of first aspect, the fine tuning training layer includes going phase
It closes training layer, train the layer strenuous exercise layer being connected and the relaxation training being connected with the strenuous exercise layer with the decorrelation
Layer, wherein:
Layer is trained in the decorrelation, for will the original weight matrix progress singular value decomposition, obtain the first matrix, the
Two matrixes and the 3rd matrix, using the product between first matrix and second matrix as with reference to weight matrix;Wherein,
The original weight matrix W is the orthogonal matrix of n*m, and the first matrix U is the orthonormal matrix of n*n, and the second matrix S is n*m
Diagonal matrix, the 3rd matrix V be m*m orthonormal matrix;
The strenuous exercise layer, it is described with reference to weight matrix for fixing, according to the training set to the full articulamentum
It is trained, obtains suboptimum weight matrix;
The relaxation training layer, for being carried out according to the suboptimum weight matrix and the training set to the full articulamentum
Training, obtains the target weight matrix.
With reference to first aspect or the first possible embodiment of first aspect, second of possible reality of first aspect
Apply in mode, the deep neural network identification model further include with the pretreatment layer that is connected of fine tuning training layer, for pair
Target detection image carries out scratching figure, obtains pedestrian image;The pedestrian image is subjected to size processing, obtains the target figure
Picture, so that the image size of the target image and the input substantially of the deep neural network identification model are in the same size.
With reference to first aspect or the first possible embodiment of first aspect, the third possible reality of first aspect
It applies in mode, the full articulamentum is specifically used for passing through in Euclidean distance, mahalanobis distance, COS distance or Hamming distance at least
A kind of calculation obtains the similarity between pedestrian's feature and each preset data of preset data concentration, obtains described more
A similarity.
With reference to first aspect or the first possible embodiment of first aspect, the 4th kind of possible reality of first aspect
It applies in mode, the training set includes the corresponding training image of multiple angles, and each angle corresponds to an at least training image.
Second aspect, the embodiment of the present application provide a kind of recognition methods of deep neural network identification model, the side
Method based on the deep neural network identification model described in first aspect, wherein:
Fine tuning training layer instructs the full articulamentum according to the original weight matrix and training set of the full articulamentum
Practice, and target weight matrix is obtained when training is completed;
Full articulamentum extracts pedestrian's feature of target image according to the target weight matrix;Obtain pedestrian's feature and
Preset data concentrates the similarity between each preset data, obtains multiple similarities;The multiple similarity is ranked up,
Obtain the corresponding preset data of maximum similarity;It is true according to the preset data when the maximum similarity is more than predetermined threshold value
The characteristic of pedestrian in the fixed target image.
With reference to second aspect, in the first possible embodiment of second aspect, the fine tuning training layer is according to
The original weight matrix of full articulamentum is trained the full articulamentum, obtains target weight matrix, including:
Decorrelation training layer will the original weight matrix progress singular value decomposition, obtain the first matrix, the second matrix and
3rd matrix, using the product between first matrix and second matrix as with reference to weight matrix;Wherein, it is described original
Weight matrix W is the orthogonal matrix of n*m, and the first matrix U is the orthonormal matrix of n*n, and the second matrix S is n*m to angular moment
Battle array, the 3rd matrix V are the orthonormal matrix of m*m;
The fixation of strenuous exercise layer is described to refer to weight matrix, and the full articulamentum is trained according to the training set,
Obtain suboptimum weight matrix;
Relaxation training layer is trained the full articulamentum according to the suboptimum weight matrix and the training set, obtains
The target weight matrix.
With reference to the possible embodiment of the first of second aspect or second aspect, second of possible reality of second aspect
It applies in mode, before the full articulamentum extracts pedestrian's feature of target image according to the target weight matrix, the side
Method further includes:
Pretreatment layer carries out target detection image to scratch figure, obtains pedestrian image;The pedestrian image is carried out at size
Reason, obtains the target image, so that the base of the image size of the target image and the deep neural network identification model
This input is in the same size.
With reference to the possible embodiment of the first of second aspect or second aspect, the third possible reality of second aspect
It applies in mode, the full articulamentum obtains the similarity between pedestrian's feature and each preset data of preset data concentration,
Multiple similarities are obtained, including:
The full articulamentum passes through the calculating of at least one of Euclidean distance, mahalanobis distance, COS distance or Hamming distance
Mode obtains the similarity between pedestrian's feature and each preset data of preset data concentration, obtains the multiple similar
Degree.
With reference to the possible embodiment of the first of second aspect or second aspect, the 4th kind of possible reality of second aspect
It applies in mode, the training set includes the corresponding training image of multiple angles, and each angle corresponds to an at least training image.
The third aspect, the embodiment of the present application provide a kind of computer readable storage medium, the computer storage media
Computer program is stored with, the computer program includes program instruction, and described program instruction makes institute when being executed by a processor
State the method that processor performs above-mentioned second aspect.
After employing above-mentioned deep neural network identification model and recognition methods, by finely tuning training layer according to connecting entirely
The original weight matrix and training set for connecing layer are trained full articulamentum, and obtain target weight matrix when training is completed;
Pedestrian's feature of target image is extracted according to target weight matrix by full articulamentum, and obtains pedestrian's feature and preset data collection
In similarity between each preset data obtain multiple similarities, multiple similarities are ranked up to obtain maximum similarity pair
The preset data answered, and when maximum similarity is more than predetermined threshold value, pedestrian in target image is determined according to the preset data
Characteristic.That is, the weight matrix of full articulamentum is optimized by finely tuning training layer, so as to improve depth nerve
The recognition efficiency and accuracy rate of Network Recognition model improve the recognition speed and accuracy rate that people identifies again.
Description of the drawings
It in order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, below will be to embodiment or existing
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is the application
Some embodiments, for those of ordinary skill in the art, without creative efforts, can also basis
These attached drawings obtain other attached drawings.
Wherein:
Fig. 1 is a kind of structure diagram of deep neural network identification model provided by the embodiments of the present application;
Figure 1A is a kind of structure diagram for finely tuning training layer provided by the embodiments of the present application;
Fig. 2 is a kind of flow diagram of the recognition methods of neural network recognization model provided by the embodiments of the present application.
Specific embodiment
Below in conjunction with the attached drawing in the embodiment of the present application, the technical solution in the embodiment of the present application is carried out clear, complete
Site preparation describes, it is clear that described embodiment is some embodiments of the present application, instead of all the embodiments.Based on this Shen
Please in embodiment, the every other implementation that those of ordinary skill in the art are obtained without creative efforts
Example, shall fall in the protection scope of this application.
It should be appreciated that ought use in this specification and in the appended claims, term " comprising " and "comprising" instruction
Described feature, entirety, step, operation, the presence of element and/or component, but it is not precluded from one or more of the other feature, whole
Body, step, operation, element, component and/or its presence or addition gathered.
It is also understood that the term used in this present specification is merely for the sake of the mesh for describing specific embodiment
And be not intended to limit the application.As present specification and it is used in the attached claims, unless on
Other situations are hereafter clearly indicated, otherwise " one " of singulative, "one" and "the" are intended to include plural form.
It will be further appreciated that the term "and/or" used in present specification and the appended claims is
Refer to any combinations and all possible combinations of one or more of the associated item listed, and including these combinations.
As used in this specification and in the appended claims, term " if " can be according to context quilt
Be construed to " when ... " or " once " or " in response to determining " or " in response to detecting ".Similarly, phrase " if it is determined that " or
" if detecting [described condition or event] " can be interpreted to mean according to context " once it is determined that " or " in response to true
It is fixed " or " once detecting [described condition or event] " or " in response to detecting [described condition or event] ".
The embodiment of the present application proposes a kind of deep neural network identification model and recognition methods, the god completed based on training
Through Network Recognition model.That is, the application is to be finely adjusted training to neural network recognization model, pedestrian can be improved and known again
Other recognition speed and accuracy rate.Below in conjunction with specific embodiment, and referring to the drawings, the application is further described.
Fig. 1 is refer to, Fig. 1 is a kind of structure diagram for deep neural network identification model that the application provides.Such as Fig. 1
Shown, above-mentioned deep neural network identification model 100 includes fine tuning training layer 102 and full articulamentum 104, and finely tunes training layer
102 be the last layer of full articulamentum 104, that is, the output layer for finely tuning training layer 102 connects the input layer of full articulamentum 104, Quan Lian
Connect the neural network recognization model that layer 104 is completed for training.
In this application, the training set of the training set of deep neural network identification model 100 and neural network recognization model
Unanimously, comprising the corresponding training image of multiple angles, and an each angle correspondence at least training image, training image can be
The image or same camera of different camera acquisitions gather the image under the different angle of same people.Work as training set
During multiple test images corresponding for multiple angles, can be trained for different angles, improve different angle under pedestrian again
The accuracy rate of identification, convenient for subsequent tracing and monitoring.The application includes for the training set in neural network recognization model 100
The quantity of training image is not restricted, and quantity is more, and frequency of training is more, and the accuracy rate of neural network recognization model 100 is higher.
Optionally, full articulamentum 104 is used to carry out angle recognition to training image, obtains the angle of training image;According to
Angle carries out feature recognition to training image, obtains pedestrian's feature of the corresponding pedestrian's object of training image.That is, first really
Determine the angle of training image, then feature recognition is carried out for angle, the accuracy rate of Attribute Recognition can be improved.
For example, the gender attribute of pedestrian includes man, two kinds of female, and angular nature is divided into front, side and the back side 3
Kind, then for the above situation, can be exported by 6 generic attributes as a result, the determined property of the man and 2 class of female before being directed to again obtain
To final output as a result, so as to improve the accuracy rate of pedestrian's feature recognition.
In this application, the convergency value by trained loss is arranged to first threshold, and the threshold value of frequency of training is arranged to
Second threshold.
It should be noted that deep neural network identification model 100, neural network recognization model involved by the application with
The training method of other neural network models is consistent, i.e. a cycle of training has been propagated by single forward operation and reversed gradient
Into.That is, carry out forward operation according to the connection mode in deep neural network identification model 100;In depth nerve net
Loss between output attribute and desired output attribute that network identification model 100 obtains is more than first threshold, and frequency of training is small
When second threshold, reversed gradient propagation is carried out according to loss, i.e., each layer power is corrected in a manner that loss gradient declines
Value, is adjusted each layer weights.By information forward-propagating in cycles and reversed gradient communication process is lost, to depth
Neural network recognization model 100 is trained, and can reduce the loss of the output of deep neural network identification model 100, improves identification
Accuracy rate.
Optionally, training layer 102 is finely tuned to be used for according to formula:Loss=(yp-y)2Calculate the loss loss.
Wherein, ypIt is expected output attribute, y is output attribute.
In this application, training layer 102 is finely tuned to be used for according to the original weight matrix of the full articulamentum and training set pair
The full articulamentum is trained, and obtains target weight matrix when training is completed.
The weight matrix of full articulamentum 104 determines the accuracy rate of full articulamentum 104 and neural network recognization model, when
When being trained to full articulamentum 104, deep neural network identification model can be improved on the basis of neural network recognization model
100 accuracy rate.
Optionally, as shown in Figure 1A, finely tuning training layer 102 includes decorrelation training layer 1021, is trained with the decorrelation
The strenuous exercise layer 1022 of 1021 connection of layer and the relaxation training layer 1023 being connected with the strenuous exercise layer 1022.
Wherein, decorrelation training layer 1021 is used to the original weight matrix carrying out singular value decomposition (singular
Value decomposition, SVD), the first matrix, the second matrix and the 3rd matrix are obtained, by first matrix and described
Product between second matrix is used as with reference to weight matrix.
Singular value decomposition is a kind of data analysing method, and for finding out " pattern " that is implied in mass data, it can be with
Used in pattern-recognition, data set is mapped in lower dimensional space by data compression etc..Specifically calculation is:It is assuming that original
Weight matrix W is m*n rank complex matrix, then
W=USV'
Wherein, U is the orthonormal matrix of n*n, and S is the diagonal matrix of n*m, and V is the orthonormal matrix of m*m, then joins
Examine weight matrix W'=US.
Strenuous exercise layer 1022, it is described with reference to weight matrix for fixing, according to the training set to the full articulamentum
104 are trained, and obtain suboptimum weight matrix.
After fine tuning training layer 102 are with the addition of, the weights of full articulamentum 104 are changed, each time strenuous exercise layer
1022 when being trained, and obtains multiple weights models, multiple weights models are ranked up, and best initial weights model is selected to carry out down
The relaxation training of one step.
Relaxation training layer 1023, for according to the suboptimum weight matrix and the training set to the full articulamentum 104
It is trained, obtains the target weight matrix.
The fixation of suboptimum weight matrix is eliminated, full articulamentum 104 is instructed according to suboptimum weight matrix and training set
Practice, the training of full articulamentum 104 each time can also generate multiple weights models, above-mentioned weights model is ranked up, and select
Best initial weights model carries out relaxation training next time, and training is completed when full articulamentum 104 is restrained, obtains target weight square
Battle array.
In this application, full articulamentum 104 is used to extract pedestrian's feature of target image according to the target weight matrix;
The similarity between pedestrian's feature and each preset data of preset data concentration is obtained, obtains multiple similarities;To described
Multiple similarities are ranked up, and obtain the corresponding preset data of maximum similarity;It is more than predetermined threshold value in the maximum similarity
When, the characteristic of pedestrian in the target image is determined according to the preset data.
Wherein, predetermined threshold value is the minimum similarity determined in image required by pedestrian, i.e., more than predetermined threshold value when can be true
Determine characteristic of the preset data for pedestrian in target image.
The application is not construed as limiting for the calculation of similarity, can by Euclidean distance, mahalanobis distance, COS distance or
Hamming distance etc..
It is appreciated that full articulamentum 104 plays feature extraction and feature in entire depth neural network recognization model 100
The effect of identification obtains the similarity between pedestrian's feature in target image and each preset data of data concentration, in basis
Whether maximum similarity is more than the characteristic that predetermined threshold value determines pedestrian in target image, so as to complete identifying again for pedestrian
Journey.
Optionally, the deep neural network identification model 100 further includes the pre- place being connected with the fine tuning training layer 102
Layer is managed, for carrying out scratching figure to target detection image, obtains pedestrian image;The pedestrian image is subjected to size processing, is obtained
The target image, so that the image size of the target image and the basic input of the deep neural network identification model are big
It is small consistent.
Wherein, target detection image is any training image or the image of any camera acquisition.It is appreciated that when extraction
It after pedestrian image in target detection image, will be deleted with the incoherent part of pedestrian image, improve pedestrian's knowledge
Other accuracy rate.
Required by having due to each neural network model to the basic input size of input picture, i.e., it can only be equal to basic
The image of input size is handled, then pedestrian image is carried out size processing, so that the figure of the target image after processing
As the input substantially of size and deep neural network identification model is in the same size, that is to say, that neural network recognization model it is basic
It inputs in the same size.
Optionally, when the image size of the pedestrian image is less than the basic input size, according to described substantially defeated
Enter size and pre-set image to be filled the pedestrian image;When the image size of the pedestrian image is more than depth god
During basic input size through Network Recognition model, the pedestrian image is zoomed in and out according to the basic input size.
Wherein, pre-set image can be the image or relatively low image of resolution ratio etc. of same color, herein no longer
It repeats.
In deep neural network identification model as shown in Figure 1, by finely tuning training layer 102 according to the original of full articulamentum
Weight matrix and training set are trained full articulamentum, and obtain target weight matrix when training is completed;By connecting entirely
Layer 104 extracts pedestrian's feature of target image according to target weight matrix, and obtain pedestrian's feature and preset data concentrate it is each
Similarity between preset data obtains multiple similarities, multiple similarities is ranked up to obtain maximum similarity corresponding pre-
If data, and when maximum similarity is more than predetermined threshold value, the characteristic of pedestrian in target image is determined according to the preset data
According to.That is, it is optimized by the weight matrix for finely tuning the full articulamentum 104 of 102 pairs of layer of training, so as to improve depth nerve
The recognition efficiency and accuracy rate of Network Recognition model improve the recognition speed and accuracy rate that people identifies again.
Fig. 2 is referred to, Fig. 2 is a kind of recognition methods of deep neural network identification model provided by the embodiments of the present application
Flow diagram, as shown in Fig. 2, this method is applied to deep neural network identification model as shown in Figure 1, wherein:
201:Fine tuning training layer according to the original weight matrix of the full articulamentum and training set to the full articulamentum into
Row training, and obtain target weight matrix when training is completed.
202:Full articulamentum extracts pedestrian's feature of target image according to the target weight matrix;It is special to obtain the pedestrian
Preset data of seeking peace concentrates the similarity between each preset data, obtains multiple similarities;The multiple similarity is carried out
Sequence, obtains the corresponding preset data of maximum similarity;When the maximum similarity is more than predetermined threshold value, according to the present count
According to the characteristic for determining pedestrian in the target image.
Optionally, the fine tuning training layer carries out the full articulamentum according to the original weight matrix of the full articulamentum
Training, obtains target weight matrix, including:
Decorrelation training layer will the original weight matrix progress singular value decomposition, obtain the first matrix, the second matrix and
3rd matrix, using the product between first matrix and second matrix as with reference to weight matrix;Wherein, it is described original
Weight matrix W is the orthogonal matrix of n*m, and the first matrix U is the orthonormal matrix of n*n, and the second matrix S is n*m to angular moment
Battle array, the 3rd matrix V are the orthonormal matrix of m*m;
The fixation of strenuous exercise layer is described to refer to weight matrix, and the full articulamentum is trained according to the training set,
Obtain suboptimum weight matrix;
Relaxation training layer is trained the full articulamentum according to the suboptimum weight matrix and the training set, obtains
The target weight matrix.
Optionally, before the full articulamentum extracts pedestrian's feature of target image according to the target weight matrix,
The method further includes:
Pretreatment layer carries out target detection image to scratch figure, obtains pedestrian image;The pedestrian image is carried out at size
Reason, obtains the target image, so that the base of the image size of the target image and the deep neural network identification model
This input is in the same size.
Optionally, the full articulamentum obtains the phase between pedestrian's feature and each preset data of preset data concentration
Like degree, multiple similarities are obtained, including:
The full articulamentum passes through the calculating of at least one of Euclidean distance, mahalanobis distance, COS distance or Hamming distance
Mode obtains the similarity between pedestrian's feature and each preset data of preset data concentration, obtains the multiple similar
Degree.
Optionally, the training set includes the corresponding training image of multiple angles, and each angle corresponds to an at least training
Image.
In the recognition methods of deep neural network identification model as shown in Figure 2, by finely tuning training layer according to full connection
The original weight matrix and training set of layer are trained full articulamentum, and obtain target weight matrix when training is completed;It is logical
It crosses full articulamentum and pedestrian's feature of target image is extracted according to target weight matrix, and obtain pedestrian's feature and preset data concentration
Similarity between each preset data obtains multiple similarities, and multiple similarities are ranked up to obtain maximum similarity correspondence
Preset data, and when maximum similarity is more than predetermined threshold value, the spy of pedestrian in target image is determined according to the preset data
Levy data.That is, the weight matrix of full articulamentum is optimized by finely tuning training layer, so as to improve depth nerve net
The recognition efficiency and accuracy rate of network identification model improve the recognition speed and accuracy rate that people identifies again.
A kind of computer readable storage medium, above computer readable storage medium are provided in another embodiment of the invention
Matter is stored with computer program, and above computer program includes program instruction, and above procedure instruction makes when being executed by a processor
Above-mentioned processor performs the realization method described in the recognition methods of deep neural network identification model.
Those of ordinary skill in the art may realize that each exemplary lists described with reference to the embodiments described herein
Member and algorithm steps can be realized with the combination of electronic hardware, computer software or the two, in order to clearly demonstrate hardware
With the interchangeability of software, each exemplary composition and step are generally described according to function in the above description.This
A little functions are performed actually with hardware or software mode, specific application and design constraint depending on technical solution.Specially
Industry technical staff can realize described function to each specific application using distinct methods, but this realization is not
It is considered as beyond the scope of this invention.
It is apparent to those skilled in the art that for convenience of description and succinctly, the end of foregoing description
End and the specific work process of unit, may be referred to the corresponding process in preceding method embodiment, details are not described herein.
In several embodiments provided herein, it should be understood that disclosed terminal and method can pass through it
Its mode is realized.For example, the apparatus embodiments described above are merely exemplary, for example, the division of said units, only
Only a kind of division of logic function can have other dividing mode in actual implementation, such as multiple units or component can be tied
It closes or is desirably integrated into another system or some features can be ignored or does not perform.In addition, shown or discussed phase
Coupling, direct-coupling or communication connection between mutually can be INDIRECT COUPLING or the communication by some interfaces, device or unit
Connection or electricity, the connection of mechanical or other forms.
The above-mentioned unit illustrated as separating component may or may not be physically separate, be shown as unit
The component shown may or may not be physical location, you can be located at a place or can also be distributed to multiple
In network element.Some or all of unit therein can be selected to realize the embodiment of the present invention according to the actual needs
Purpose.
In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, it can also
It is that unit is individually physically present or two or more units integrate in a unit.It is above-mentioned integrated
The form that hardware had both may be employed in unit is realized, can also be realized in the form of SFU software functional unit.
If above-mentioned integrated unit is realized in the form of SFU software functional unit and is independent production marketing or use
When, it can be stored in a computer read/write memory medium.Based on such understanding, technical scheme is substantially
The part to contribute in other words to the prior art or all or part of the technical solution can be in the form of software products
It embodies, which is stored in a storage medium, is used including some instructions so that a computer
Equipment (can be personal computer, server or the network equipment etc.) performs the complete of each embodiment above method of the present invention
Portion or part steps.And foregoing storage medium includes:USB flash disk, mobile hard disk, read-only memory (Read-Only memory,
ROM), random access memory (random access memory, RAM), magnetic disc or CD etc. are various can store program
The medium of code.
It should be noted that in attached drawing or specification text, the realization method that does not illustrate or describe is affiliated technology
Form known to a person of ordinary skill in the art, is not described in detail in field.In addition, the above-mentioned definition to each element and method is simultaneously
Various concrete structures, shape or the mode mentioned in embodiment are not limited only to, those of ordinary skill in the art can carry out letter to it
It singly changes or replaces.
Particular embodiments described above has carried out the purpose, technical solution and advantageous effect of the application further in detail
Describe in detail bright, it should be understood that the foregoing is merely the specific embodiment of the application, be not limited to the application, it is all
Within spirit herein and principle, any modification, equivalent substitution, improvement and etc. done should be included in the protection of the application
Within the scope of.
Claims (10)
1. a kind of deep neural network identification model, which is characterized in that connect including fine tuning training layer and with the fine tuning training layer
The full articulamentum connect, the neural network recognization model that the full articulamentum is completed for training, wherein:
The fine tuning training layer, for the original weight matrix according to the full articulamentum and training set to the full articulamentum into
Row training, and obtain target weight matrix when training is completed;
The full articulamentum, for extracting pedestrian's feature of target image according to the target weight matrix;Obtain the pedestrian
Feature and preset data concentrate the similarity between each preset data, obtain multiple similarities;To the multiple similarity into
Row sequence, obtains the corresponding preset data of maximum similarity;It is default according to this when the maximum similarity is more than predetermined threshold value
Data determine the characteristic of pedestrian in the target image.
2. deep neural network identification model according to claim 1, which is characterized in that the fine tuning training layer includes going
Correlation training layer trains the layer strenuous exercise layer being connected and the relaxation training being connected with the strenuous exercise layer with the decorrelation
Layer, wherein:
Layer is trained in the decorrelation, for the original weight matrix to be carried out singular value decomposition, obtains the first matrix, the second square
Battle array and the 3rd matrix, using the product between first matrix and second matrix as with reference to weight matrix;Wherein, it is described
Original weight matrix W is the orthogonal matrix of n*m, and the first matrix U is the orthonormal matrix of n*n, and the second matrix S is pair of n*m
Angular moment battle array, the 3rd matrix V are the orthonormal matrix of m*m;
The strenuous exercise layer, it is described with reference to weight matrix for fixing, the full articulamentum is carried out according to the training set
Training, obtains suboptimum weight matrix;
The relaxation training layer, for being instructed according to the suboptimum weight matrix and the training set to the full articulamentum
Practice, obtain the target weight matrix.
3. deep neural network identification model according to claim 1 or 2, which is characterized in that the deep neural network
Identification model further includes the pretreatment layer being connected with the fine tuning training layer, for carrying out scratching figure to target detection image, obtains
Pedestrian image;The pedestrian image is subjected to size processing, the target image is obtained, so that the image of the target image is big
It is small in the same size with the input substantially of the deep neural network identification model.
4. deep neural network identification model according to claim 1 or 2, which is characterized in that the full articulamentum is specific
For obtaining the pedestrian by least one of Euclidean distance, mahalanobis distance, COS distance or Hamming distance calculation
Feature and preset data concentrate the similarity between each preset data, obtain the multiple similarity.
5. deep neural network identification model according to claim 1 or 2, which is characterized in that the training set includes more
The corresponding training image of a angle, each angle correspond to an at least training image.
6. a kind of recognition methods of deep neural network identification model, which is characterized in that the method is based on the claims
1-5 any one of them deep neural network identification models, the described method includes:
Fine tuning training layer is trained the full articulamentum according to the original weight matrix and training set of the full articulamentum, and
Target weight matrix is obtained when training is completed;
Full articulamentum extracts pedestrian's feature of target image according to the target weight matrix;It obtains pedestrian's feature and presets
Similarity in data set between each preset data obtains multiple similarities;The multiple similarity is ranked up, is obtained
The corresponding preset data of maximum similarity;When the maximum similarity is more than predetermined threshold value, institute is determined according to the preset data
State the characteristic of pedestrian in target image.
7. according to the method described in claim 6, it is characterized in that, the fine tuning training layer is according to the original of the full articulamentum
Weight matrix is trained the full articulamentum, obtains target weight matrix, including:
Decorrelation trains layer that the original weight matrix is carried out singular value decomposition, obtains the first matrix, the second matrix and the 3rd
Matrix, using the product between first matrix and second matrix as with reference to weight matrix;Wherein, the original weight
Matrix W is the orthogonal matrix of n*m, and the first matrix U is the orthonormal matrix of n*n, and the second matrix S is the diagonal matrix of n*m, the
Three matrix Vs are the orthonormal matrix of m*m;
The fixation of strenuous exercise layer is described with reference to weight matrix, and the full articulamentum is trained according to the training set, is obtained
Suboptimum weight matrix;
Relaxation training layer is trained the full articulamentum according to the suboptimum weight matrix and the training set, obtains described
Target weight matrix.
8. the method according to claim 6 or 7, which is characterized in that in the full articulamentum according to the target weight square
Before pedestrian's feature of battle array extraction target image, the method further includes:
Pretreatment layer carries out target detection image to scratch figure, obtains pedestrian image;The pedestrian image is subjected to size processing, is obtained
To the target image, so that the basic input of the image size of the target image and the deep neural network identification model
It is in the same size.
9. the method according to claim 6 or 7, which is characterized in that the full articulamentum obtains pedestrian's feature and pre-
If the similarity in data set between each preset data obtains multiple similarities, including:
The full articulamentum passes through at least one of Euclidean distance, mahalanobis distance, COS distance or Hamming distance calculation
The similarity between pedestrian's feature and each preset data of preset data concentration is obtained, obtains the multiple similarity.
10. the method according to claim 6 or 7, which is characterized in that the training set includes the corresponding training of multiple angles
Image, each angle correspond to an at least training image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711209932.7A CN108108662B (en) | 2017-11-24 | 2017-11-24 | Deep neural network recognition model and recognition method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711209932.7A CN108108662B (en) | 2017-11-24 | 2017-11-24 | Deep neural network recognition model and recognition method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108108662A true CN108108662A (en) | 2018-06-01 |
CN108108662B CN108108662B (en) | 2021-05-25 |
Family
ID=62207702
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711209932.7A Active CN108108662B (en) | 2017-11-24 | 2017-11-24 | Deep neural network recognition model and recognition method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108108662B (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108769598A (en) * | 2018-06-08 | 2018-11-06 | 复旦大学 | Across the camera video method for concentration identified again based on pedestrian |
CN110008997A (en) * | 2019-03-06 | 2019-07-12 | 平安科技(深圳)有限公司 | Image texture similarity recognition method, device and computer readable storage medium |
CN110493598A (en) * | 2019-08-12 | 2019-11-22 | 北京中科寒武纪科技有限公司 | Method for processing video frequency and relevant apparatus |
CN111191601A (en) * | 2019-12-31 | 2020-05-22 | 深圳云天励飞技术有限公司 | Method, device, server and storage medium for identifying peer users |
CN111191675A (en) * | 2019-12-03 | 2020-05-22 | 深圳市华尊科技股份有限公司 | Pedestrian attribute recognition model implementation method and related device |
CN111191602A (en) * | 2019-12-31 | 2020-05-22 | 深圳云天励飞技术有限公司 | Pedestrian similarity obtaining method and device, terminal equipment and readable storage medium |
CN111242217A (en) * | 2020-01-13 | 2020-06-05 | 支付宝实验室(新加坡)有限公司 | Training method and device of image recognition model, electronic equipment and storage medium |
CN111860629A (en) * | 2020-06-30 | 2020-10-30 | 北京滴普科技有限公司 | Jewelry classification system, method, device and storage medium |
CN112307227A (en) * | 2020-11-24 | 2021-02-02 | 国家电网有限公司大数据中心 | Data classification method |
CN112560720A (en) * | 2020-12-21 | 2021-03-26 | 奥比中光科技集团股份有限公司 | Pedestrian identification method and system |
WO2021068142A1 (en) * | 2019-10-09 | 2021-04-15 | 深圳大学 | Training method and detection method for automatically identifying recaptured image of original document |
CN113205451A (en) * | 2021-03-30 | 2021-08-03 | 北京达佳互联信息技术有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140372112A1 (en) * | 2013-06-18 | 2014-12-18 | Microsoft Corporation | Restructuring deep neural network acoustic models |
CN104239892A (en) * | 2014-08-25 | 2014-12-24 | 西安电子科技大学 | SAR image mixed model fitting method based on KSVD training dictionary |
CN104850531A (en) * | 2014-02-19 | 2015-08-19 | 日本电气株式会社 | Method and device for establishing mathematical model |
CN105979210A (en) * | 2016-06-06 | 2016-09-28 | 深圳市深网视界科技有限公司 | Pedestrian identification system based on multi-ball multi-gun camera array |
CN106295584A (en) * | 2016-08-16 | 2017-01-04 | 深圳云天励飞技术有限公司 | Depth migration study is in the recognition methods of crowd's attribute |
CN106355248A (en) * | 2016-08-26 | 2017-01-25 | 深圳先进技术研究院 | Deep convolution neural network training method and device |
CN106503687A (en) * | 2016-11-09 | 2017-03-15 | 合肥工业大学 | The monitor video system for identifying figures of fusion face multi-angle feature and its method |
CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
CN106803063A (en) * | 2016-12-21 | 2017-06-06 | 华中科技大学 | A kind of metric learning method that pedestrian recognizes again |
CN107301380A (en) * | 2017-06-01 | 2017-10-27 | 华南理工大学 | One kind is used for pedestrian in video monitoring scene and knows method for distinguishing again |
CN107341541A (en) * | 2016-04-29 | 2017-11-10 | 北京中科寒武纪科技有限公司 | A kind of apparatus and method for performing full articulamentum neural metwork training |
-
2017
- 2017-11-24 CN CN201711209932.7A patent/CN108108662B/en active Active
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140372112A1 (en) * | 2013-06-18 | 2014-12-18 | Microsoft Corporation | Restructuring deep neural network acoustic models |
CN104850531A (en) * | 2014-02-19 | 2015-08-19 | 日本电气株式会社 | Method and device for establishing mathematical model |
CN104239892A (en) * | 2014-08-25 | 2014-12-24 | 西安电子科技大学 | SAR image mixed model fitting method based on KSVD training dictionary |
CN107341541A (en) * | 2016-04-29 | 2017-11-10 | 北京中科寒武纪科技有限公司 | A kind of apparatus and method for performing full articulamentum neural metwork training |
CN105979210A (en) * | 2016-06-06 | 2016-09-28 | 深圳市深网视界科技有限公司 | Pedestrian identification system based on multi-ball multi-gun camera array |
CN106295584A (en) * | 2016-08-16 | 2017-01-04 | 深圳云天励飞技术有限公司 | Depth migration study is in the recognition methods of crowd's attribute |
CN106355248A (en) * | 2016-08-26 | 2017-01-25 | 深圳先进技术研究院 | Deep convolution neural network training method and device |
CN106503687A (en) * | 2016-11-09 | 2017-03-15 | 合肥工业大学 | The monitor video system for identifying figures of fusion face multi-angle feature and its method |
CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
CN106803063A (en) * | 2016-12-21 | 2017-06-06 | 华中科技大学 | A kind of metric learning method that pedestrian recognizes again |
CN107301380A (en) * | 2017-06-01 | 2017-10-27 | 华南理工大学 | One kind is used for pedestrian in video monitoring scene and knows method for distinguishing again |
Non-Patent Citations (2)
Title |
---|
YIFAN S.等: "SVDNet for Pedestrian Retrieval", 《百度学术》 * |
中国支付清算协会: "《中国支付清算行业运行报告 2015》", 31 July 2015, 北京:中国金融出版社 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108769598A (en) * | 2018-06-08 | 2018-11-06 | 复旦大学 | Across the camera video method for concentration identified again based on pedestrian |
CN110008997A (en) * | 2019-03-06 | 2019-07-12 | 平安科技(深圳)有限公司 | Image texture similarity recognition method, device and computer readable storage medium |
CN110008997B (en) * | 2019-03-06 | 2023-11-24 | 平安科技(深圳)有限公司 | Image texture similarity recognition method, device and computer readable storage medium |
CN110493598A (en) * | 2019-08-12 | 2019-11-22 | 北京中科寒武纪科技有限公司 | Method for processing video frequency and relevant apparatus |
WO2021068142A1 (en) * | 2019-10-09 | 2021-04-15 | 深圳大学 | Training method and detection method for automatically identifying recaptured image of original document |
CN111191675B (en) * | 2019-12-03 | 2023-10-24 | 深圳市华尊科技股份有限公司 | Pedestrian attribute identification model realization method and related device |
CN111191675A (en) * | 2019-12-03 | 2020-05-22 | 深圳市华尊科技股份有限公司 | Pedestrian attribute recognition model implementation method and related device |
CN111191601A (en) * | 2019-12-31 | 2020-05-22 | 深圳云天励飞技术有限公司 | Method, device, server and storage medium for identifying peer users |
CN111191602A (en) * | 2019-12-31 | 2020-05-22 | 深圳云天励飞技术有限公司 | Pedestrian similarity obtaining method and device, terminal equipment and readable storage medium |
CN111242217A (en) * | 2020-01-13 | 2020-06-05 | 支付宝实验室(新加坡)有限公司 | Training method and device of image recognition model, electronic equipment and storage medium |
CN111860629A (en) * | 2020-06-30 | 2020-10-30 | 北京滴普科技有限公司 | Jewelry classification system, method, device and storage medium |
CN112307227A (en) * | 2020-11-24 | 2021-02-02 | 国家电网有限公司大数据中心 | Data classification method |
CN112307227B (en) * | 2020-11-24 | 2023-08-29 | 国家电网有限公司大数据中心 | Data classification method |
CN112560720A (en) * | 2020-12-21 | 2021-03-26 | 奥比中光科技集团股份有限公司 | Pedestrian identification method and system |
CN113205451A (en) * | 2021-03-30 | 2021-08-03 | 北京达佳互联信息技术有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
CN113205451B (en) * | 2021-03-30 | 2024-05-24 | 北京达佳互联信息技术有限公司 | Image processing method, device, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN108108662B (en) | 2021-05-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108108662A (en) | Deep neural network identification model and recognition methods | |
CN108197532B (en) | The method, apparatus and computer installation of recognition of face | |
CN107247949A (en) | Face identification method, device and electronic equipment based on deep learning | |
CN106575365A (en) | Object recognition trait analysis systems and methods | |
US9183431B2 (en) | Apparatus and method for providing activity recognition based application service | |
CN110088776A (en) | For training the computer equipment of deep neural network | |
CN106716450A (en) | Image-based feature detection using edge vectors | |
CN107992844A (en) | Face identification system and method based on deep learning | |
CN105740808B (en) | Face identification method and device | |
CN109117897A (en) | Image processing method, device and readable storage medium storing program for executing based on convolutional neural networks | |
WO2021120961A1 (en) | Brain addiction structure map evaluation method and apparatus | |
CN111652087A (en) | Car checking method and device, electronic equipment and storage medium | |
CN110503076A (en) | Video classification methods, device, equipment and medium based on artificial intelligence | |
CN108875481A (en) | Method, apparatus, system and storage medium for pedestrian detection | |
CN113822254B (en) | Model training method and related device | |
CN108229262A (en) | A kind of pornographic video detecting method and device | |
CN110287848A (en) | The generation method and device of video | |
CN109101866B (en) | Pedestrian re-identification method and system based on segmentation silhouette | |
CN108629170A (en) | Personal identification method and corresponding device, mobile terminal | |
CN106022287A (en) | Over-age face verification method based on deep learning and dictionary representation | |
CN108875509A (en) | Biopsy method, device and system and storage medium | |
WO2024060684A1 (en) | Model training method, image processing method, device, and storage medium | |
CN108875500A (en) | Pedestrian recognition methods, device, system and storage medium again | |
CN110929243B (en) | Pedestrian identity recognition method based on mobile phone inertial sensor | |
CN114495241A (en) | Image identification method and device, electronic equipment and storage medium |
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 |