CN107527068A - Model recognizing method based on CNN and domain adaptive learning - Google Patents

Model recognizing method based on CNN and domain adaptive learning Download PDF

Info

Publication number
CN107527068A
CN107527068A CN201710670157.9A CN201710670157A CN107527068A CN 107527068 A CN107527068 A CN 107527068A CN 201710670157 A CN201710670157 A CN 201710670157A CN 107527068 A CN107527068 A CN 107527068A
Authority
CN
China
Prior art keywords
sample
domain
formula
cnn
training sample
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201710670157.9A
Other languages
Chinese (zh)
Other versions
CN107527068B (en
Inventor
孙伟
赵玉舟
张小瑞
郭强强
杜宏吉
施顺顺
杨翠芳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN201710670157.9A priority Critical patent/CN107527068B/en
Publication of CN107527068A publication Critical patent/CN107527068A/en
Application granted granted Critical
Publication of CN107527068B publication Critical patent/CN107527068B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/08—Detecting or categorising vehicles

Abstract

The present invention relates to the model recognizing method based on CNN and domain adaptive learning, and by adding invariable rotary layer in Alexnet networks, distinguishing diagnostic horizon and designing new object function, foundation is based on CNN network initial models；The characteristic pattern of different field sample convolution layer is extracted respectively using the initial model of foundation, calculate the cosine similarity between sample characteristics figure, the shared convolution kernel of CNN networks or unshared convolution kernel are determined, retains weight and the biasing of shared convolution kernel, updates weight and the biasing of unshared convolution kernel；Based on target domain training sample, the cosine similarity between every layer of characteristic pattern and whole target domain average similarity are calculated, is clustered according to average similarity per class similar features figure；The source domain sample for having similar distribution character to sample in target domain is expanded into the new samples for target domain, whole CNN network models are finely tuned with the new samples of target domain, then vehicle classification is carried out to the test sample in target domain by softmax graders.

Description

Model recognizing method based on CNN and domain adaptive learning
Technical field
The present invention relates to identification technology field, more particularly to a kind of vehicle cab recognition side based on CNN and domain adaptive learning Method.
Background technology
The identification of type of vehicle in traffic video image, as a key technology of Traffic monitoring management, for a long time By the extensive concern of researcher.Because vehicle appearance complexity is various, influenceed by factors such as background, light intensity, angles, Stability is by very big interference in actual applications.Deep learning is theoretical fast-developing in recent years, and priori is relied on different from tradition The feature extraction algorithm of knowledge, deep neural network can adaptively construction feature describes under training data driving, has more High flexibility and applicability.As an important technology for realizing deep learning, convolutional neural networks successfully trained first Deep-neural-network structure, and obtain good effect in fields such as image recognition, speech recognitions.Convolutional neural networks can be by original Beginning data such as image pixel value avoids process of data preprocessing extra in tional identification algorithm directly as input.Its class The weights for being similar to biological neural network share network structure and can produce local receptor field effect in biological vision, reduce network mould Type complexity, weights quantity is reduced, while the resistivity of height is respectively provided with to the vision deformation such as translation, proportional zoom, inclination.
Existing Image Classfication Technology mainly trains the other disaggregated model of target class by the method for supervised learning, i.e., Need there is labeled data to carry out model training for the collection of each disaggregated model is enough, high quality.This mode is applied to letter The fewer scene of single classification task and classification.But with the complication of classification task, for example categorical measure is more, classification is special Industry, particularization etc., collect enough costs for having labeled data for target classification and rise significantly.
The content of the invention
Present invention aims to overcome that vehicle image carry out feature extraction and to target domain vehicle image data sample too Less, the problem of mark cost of data is higher, there is provided a kind of model recognizing method based on CNN and domain adaptive learning, tool Body is realized by following technical scheme：
The model recognizing method based on CNN and domain adaptive learning, comprises the following steps：
Step 1) gathers the vehicle image for including various under natural scene respectively, forms vehicle image data base, A part is source domain sample in described image, and source domain sample includes source domain training and source domain test sample, remainder Being divided into target domain sample object field sample includes target domain training and source domain test sample, right after collection vehicle image Vehicle pictures in the vehicle image data base are pre-processed；
Step 2) builds CNN network models, and the vehicle image data base is imported into Alexnet networks carries out pre-training, Update the weight parameter W of Alexnet network structuresiWith offset parameter Bi, WiRepresent the weight of the i-th layer network, BiRepresent i-th layer The biasing of network, i=1,2 ... m；
Step 3) adds invariable rotary layer FCa in CNN networks, rotates source domain training sample, obtains source domain enhancing Training sample, training sample and first object function are strengthened according to the source domain, the network after invariable rotary layer is added in training Weight parameter WRI={ W1,W2,…,Wm,Wa,WbAnd biasing BRI={ B1,B2,…,Bm,Ba,Bb}；
Step 4) adds Fisher after invariable rotary layer and distinguishes layer FCc formation improvement CNN network models, passes through to input and marks Label and the second object function, training renewal with the addition of Fisher and distinguish the CNN network weight parameters W after diagnostic horizonFD={ W1, W2,…,Wm,Wa,Wc,WdAnd BFD={ B1,B2,…,Bm,Ba,Bc,Bd}；
Step 5) is extracted target domain training sample by the improvement CNN network models and trained with source domain enhancing respectively The characteristic pattern of sample, the vehicle image that source domain is strengthened to training sample and target domain training sample are separately input to by instruction In experienced improvement CNN network models, the feature of the vehicle image of source domain and target domain is extracted respectively, obtains corresponding feature Figure；
Step 6) calculates the cosine between target domain training sample characteristic pattern and source domain enhancing training sample characteristic pattern Similarity, the shared convolution kernel of improved CNN networks convolutional layer or unshared convolution kernel are determined, retain each convolutional layer shared volume The weight of product core and biasing, based on target domain training sample, update the unshared convolution of convolutional layer using stochastic gradient descent method The weight of core and biasing；
Step 7) convolution target domain training sample image, characteristic pattern is obtained, calculates the cosine similarity between characteristic pattern, The average similarity matrix of target domain training sample is obtained, is clustered using analytic hierarchy process (AHP), a similar features are retained per class Figure；
Step 8) finely tunes whole CNN networks by the target domain training sample, ultimately forms vehicle cab recognition network；
Each test sample in the target domain test sample is substituted into the vehicle cab recognition network, car by step 9) Type network judges the vehicle of vehicle according to the output result of vehicle cab recognition network.
The further design of the model recognizing method based on CNN and domain adaptive learning is, in the step 1), A part is source domain sample in described image, including CNN training sample set X={ x1,x2,…,xhAnd CNN networks Model measurement sample set R={ γ1,γ2,…,γσ, xh、γσThe sample in source domain set, h, σ difference table are represented respectively Show the number of sample in source domain set；Remainder is target domain sample, including domain adaptive learning training sample setWith domain adaptive learning test sample set M={ δ1,δ2,…,δτ, tj、δτTarget domain sample is represented respectively, T, τ represents the number of sample in target domain set respectively.
The further design of the model recognizing method based on CNN and domain adaptive learning is, described in step 1) Pretreatment operation unifies picture size, and adds label to image, marks the affiliated vehicle classification of the image.
The further design of the model recognizing method based on CNN and domain adaptive learning is, in the step 2), Alexnet network structures are made up of five convolutional layers sequentially set gradually with three full articulamentums, Alexnet networks The softmax loss functions of fusion tag are exported, according to softmax loss functions to weight parameter WiWith offset parameter BiCarry out Stochastic gradient descent is finely tuned, mistake of the fusion tag between the output result of Alexnet networks and the label actual value Difference.
The further design of the model recognizing method based on CNN and domain adaptive learning is that the step 3) is wrapped Include following steps：
The source domain training sample of vehicle image 3-1) is set as X, rotation change is carried out to each vehicle image with angle φ Change, and be added in former source domain training sample X, obtain source domain enhancing training sample χRI, χRI={ X, TφX } it is one group of figure As the data set of rotation transformation, wherein Represent that it is φ to carry out angle to training samplekRotation behaviour Make；
Full connection rotation 3-2) is added after the full articulamentum of Alexnet network structure layer 7s in CNN network models Turn not change layer FCa, by Om(xu) it is expressed as preceding layer FCm output, Oa(xu) exported for FCa layers, Ob(xu) it is expressed as softmax Classification layer FCb output, (Wa,Ba) and (Wb,Bb) new parameter of FCa layers and FCb layers is represented respectively；
Oa(xu) and Ob(xu) calculation formula respectively such as formula (1), formula (2)：
Oa(xu)=κ (WaOm(xu)+Ba) (1)
Wherein, κ (WaOm(xu)+Ba)=max (0, WaOm(xu)+Ba),
It is ReLU and softmax nonlinear activation functions respectively；
The enhancing training sample of whole CNN network models is set as χRI={ X, TφX }, corresponding label is Represent sample xuCorresponding label；New network parameter WRIAnd BRILearnt by new object function Obtain, new object function such as formula (3)：
Wherein, λ1And λ2It is weighting parameter, λ1And λ2Belong to [0,1], M (χRI,γRI) be softmax layers loss letter Number, expression formula such as formula (4)；R(X,TφX be) a rotational invariance normalized constraint item, expression formula such as formula (5), before rotation and Postrotational training sample is respectively X and TφX,It is weights attenuation term, for preventing over-fitting；
In formula (4),It isAnd logOb(xu) between Inner product, N is the total of initial training sample in X Number, K is for xuThe total degree of ∈ X rotation transformations,
Wherein, Oa(xu) be FCa layers activation primitive output,It is the FCa layers activation based on sample after rotation The average output of function, such as formula (6)；
The first object function such as formula (7)：
The further design of the model recognizing method based on CNN and domain adaptive learning is that the step 4) is wrapped Include following steps：
Fisher 4-1) is added after invariable rotary layer FCa and distinguishes diagnostic horizon FCc, by invariable rotary floor and Fisher areas Diagnostic horizon is divided to combine, CNN training samples are all real border frames of each object class, are defined as It is e-th of object class real border frame；Training sample is χFD={ xv, it is corresponding Output result isBy input results to (χFD,γFD) training；
4-2) random initializtion (Wc,Bc) and (Wd,Bd), counting loss function, distinguishing rule item and the second object function JFD(WFD, BFD), the parameter that upgrades in time WFD={ W1,W2,…,Wm,Wa,Wc,WdAnd BFD={ B1,B2,…,Bm,Ba,Bc,Bd, make Optimize structure with stochastic gradient descent method；To training sample xk∈χFD, Oa(xv) it is expressed as preceding layer FCa output, Oc(xv) be FCd layers export, Od(xv) it is expressed as softmax classification layers FCd output, (Wc,Bc) and (Wd,Bd) FCc layers and FCd are represented respectively The new parameter of layer；
Oc(xv) and Od(xv) calculation formula respectively such as formula (8), formula (9)：
Oc(xv)=κ (WcOa(xv)+Bc) (8)
The second target object function such as formula (10)
Wherein, λ3And λ4It is weighting parameter, λ3And λ4Belong to [0,1], M (χFD,γFD) be softmax layers loss letter Number, M (χFD,γFD) expression formula such as formula (11)；F(χFD) it is the discrimination regularization constraint applied to CNN features, by minimizing class Interior interval SW(χFD) maximize class between be spaced SB(χFD) obtain, SW(χFD)、SB(χFD) expression formula respectively such as formula (12), formula (13)；
In formula (11), | χFD| it is training sample χFDQuantity,
In formula (12), formula (13), neRepresent the quantity of sample in e-th of object class, wherein meRepresented respectively with mAnd χFD Average feature such as formula (14), formula (15)；
Distinguish regularization term F (χFD) such as formula (16),
F(χFD)=tr (SW(χFD))-tr(SB(χFD)) (16)
In formula (16), tr (SW(χFD)) and tr (SB(χFD)) the mark computing of representing matrix, i.e. matrix leading diagonal member
The summation of element；
Second object function is integrated into：
The further design of the model recognizing method based on CNN and domain adaptive learning is, in the step 5), Assuming that the collection of input feature vector figure is combined intoOutput characteristic atlas is combined intoWherein Rk-1And RkIt is set x respectivelykAnd zkThe number of element, k represent the number of plies of convolution；Alexnet networks convolutional layer includes convolution list Member and sub-sample unit, the intermediate features atlas between convolution unit and sub-sample unit are expressed as Each characteristic pattern ykIt is calculated as in convolution unit Represent characteristic pattern xkAnd convolution kernelIt Between convolution,Biasing is represented, F (x)=max (0, x) is activation primitive, and sub-sampling layer uses fixed size after convolution unit Average in every characteristic pattern of Nuclear receptor co repressorPond characteristic pattern corresponding to formation Convolution kernel collectionCharacteristic pattern is accordingly
The further design of the model recognizing method based on CNN and domain adaptive learning is, the step (6) In, vehicle image is strengthened into training sample χRIIt is expressed asThat is source domain enhancing training sample,For mesh Mark field training sample；
The vehicle image for having same alike result in A is divided into η classes by K mean cluster algorithm, i.e.,Each Class set AμIn vehicle image obtain similar characteristic pattern by convolutional layer, be calculated as per a kind of average characteristics figure|Aμ| the number of data is represented, is found and target image t in the η class images dividedjThe feature being consistent Figure, μ classes and target image tjWith maximum cosine average similarity, computational methods are Cos () represents the cosine similarity between two images；
Setting estimation convolution kernel fl kWhether it is each tjShared convolution kernel decision condition formula such as formula (17)；
In formula (17), ε is decision-making value, and the decision rule of convolution (17) is：If The f of this layer is then all thought based on most numerical example in target domain training samplel kIt is shared convolution kernel, retains fl kWeight And biasing；IfThen need to update fl kWeight and biasing；Softmax loss functions represent ForSo that characteristic patternApproach its corresponding Feature MappingSo as to more New unshared convolution kernel collection fl k。
The further design of the model recognizing method based on CNN and domain adaptive learning is, the step (8) In, use the ratio λ of target density and source densityωAs to source domain sample sωAvailability discrimination standard, λωUse condition is general Rate is defined such as formula (18)
p(sω| E) and p (sω| A) s is represented respectivelyωProbability in source domain and target domain, conditional probability are modeled such as formula (19), formula (20),
ot(sω) and os(sω) it is illustrated respectively in the target domain detector of previous migration and the s of source domain detectorωIt is defeated Go out value, if sωCorrectly classified by cross-cutting, then p (sω| E) ∈ (0.8,1) and p (sω| A) ∈ (0.8,1), λωIt is redefined Such as formula (21),
If λω>=1, sωThere is similar view to the vehicle in target domain；In source domain, by conditional probability λω Sample and target domain sample when >=1 constitute new sample data set, in each iterative process, calculate every in source domain The availability of vehicle simultaneously selects new Sample Refreshment training set, with the training set retraining of renewal until fusion tag Softmax loss functions are restrained, and obtain optimization feature detection model each layer of weighting parameter W' and B'.
The further design of the model recognizing method based on CNN and domain adaptive learning is, in the step 9), Last full linking layer of the vehicle cab recognition network is FC8 layers, and the output of FC8 layers is sent to softmax classification layers, described The output of softmax classification layers is a probability distribution for covering four class vehicles, corresponds to the general of a certain vehicle in output result Rate is maximum, then the output result for judging vehicle to be identified be the vehicle, and the four classes vehicle includes motor bus, truck, box Buggy and car.
Advantages of the present invention is as follows：
(1) present invention uses the method based on computer vision technique, convenient for installation and maintenance, does not influence pavement life, no Traffic is influenceed, and device therefor is few, has the advantages that cost is low, and robustness is good, safe.
(2) present invention is by optimizing and revising Alexnet network structures and setting new object function to improve aspect of model table Show ability.
(3) relevance of the invention according to different field sample characteristics figure, the domain of feature based figure similarity measurement is established Adaptive-migration learning method simultaneously optimizes renewal to vehicle characteristics extraction initial model.
Brief description of the drawings
Fig. 1 is the schematic flow sheet of the model recognizing method based on CNN and domain adaptive learning.
Fig. 2 is the schematic diagram for increasing invariable rotary layer.
Fig. 3 is the schematic diagram for adding invariable rotary layer and Fisher differentiation diagnostic horizons.
Fig. 4 is domain adaptive learning flow chart.
Embodiment
The application is further illustrated below in conjunction with the accompanying drawings.
The vehicle cab recognition side based on CNN (convolutional neural networks) and domain adaptive learning provided such as Fig. 2, the present embodiment Method, obtains good effect on vehicle vehicle cab recognition, and whole algorithm realizes that step is as follows：
Step 1：Vehicle image gathers and pretreatment
Four kinds of motor bus, truck, box buggy and car vehicle images under natural scene are chosen respectively, are adopted altogether Collect 4000 vehicle images, every kind of each 1000 of vehicle, wherein 2500 are source domain sample, including CNN training samples Set X={ x1,x2,…,xhAnd CNN network model test sample set R={ γ1,γ2,…,γσ, xh、γσRepresent respectively Sample in source domain set, h, σ represent the number of sample in source domain set respectively；1500 are target domain sample, bag Include domain adaptive learning training sample setWith domain adaptive learning test sample set M={ δ1,δ2,…,δτ, tj、δτTarget domain sample is represented respectively, and T, τ represent the number of sample in target domain set respectively；Source domain is tested first Vehicle pictures are pre-processed in sample, and picture size is uniformly adjusted to 224 × 224 pixels, and label, mark are added to image The affiliated vehicle classification of the image, there are four kinds of motor bus, truck, box buggy and car classifications；Step 2：By vehicle image Database imports Alexnet networks and carries out pre-training, updates the parameter W of Alexnet network structuresiAnd Bi, i=1,2 ... m；
Further, Alexnet network structures share 8 layers, are made up of 5 convolutional layers, 3 full articulamentums, at each Excitation function ReLu and local acknowledgement's normalization (LRN) processing are contained in convolutional layer, then by down-sampled (pooling Processing)；Full articulamentum FC6 and FC7, are 6*6 characteristic patterns to size respectively using 4096 neurons, carry out a full link, Namely by the characteristic pattern of 6*6 sizes, carry out convolution and be changed into a characteristic point, then carry out a dropout at random from 4096 Lost in node some nodal informations (be namely worth clear 0), then just obtain 4096 new neurons, full articulamentum FC8's Output is sent to 4-way softmax layers, and it produces the distribution of a 4 class labels of covering.Alexnet network models Input is piece image, and specification is 224*224*3 (RGB image), is changed into 227*227*3 by pretreatment, and 96 used are big Small dimension is 11*11 filter filter, or referred to as convolution kernel, carries out feature extraction, output is then fusion tag Softmax loss functions.Fusion tag is used to represent the now error between the output result of network and label actual value.First Training sample is obtained using from image data set, boarding steps are carried out to whole CNN parameters according to the softmax loss functions of output Degree declines fine setting.In this step, 32 positive samples and 96 negative samples (totally 128 is sampled respectively for each stochastic gradient iteration Individual sample), learning rate is set to 0.0005, and momentum is set as 0.9, and the weight of all levels is all initially set to 0.0005, so as to Change the weight parameter W of Alexnet networksi(i=1,2 ... m) and biasing Bi(i=1,2 ... m) (m=8)；
Step 3：Full connection invariable rotary layer FCa is added after the 7th layer of CNN networks, designs new object function；Rotation Source domain training sample X is simultaneously added in former training sample X, obtains enhancing training sample χRI, according to enhancing training sample χRIThe network weight parameter W after addition invariable rotary layer is updated with object functionRI={ W1,W2,…,Wm,Wa,WbAnd biasing BRI ={ B1,B2,…,Bm,Ba,Bb}；
(1) assume that source domain vehicle image training sample is X, rotation transformation carried out to each vehicle image with angle φ, φ=10 ° in this patent, and be added in former training sample X, obtain strengthening training sample χRI。χRI={ X, TφX } it is one group of figure As the data set of rotation transformation, wherein Represent that it is φ to carry out angle to training samplekRotation Operation；
(2) full connection invariable rotary layer FCa, O are added after the 7th layer of CNN networksm(xu) it is expressed as the defeated of preceding layer FCm Go out, Oa(xu) exported for FCa layers, Ob(xu) it is expressed as softmax classification layers FCb output, (Wa,Ba) and (Wb,Bb) represent respectively The new parameter of FCa layers and FCb layers.
Oa(xu) and Ob(xu) calculation formula be respectively：Oa(xu)=κ (WaOm(xu)+Ba) and Wherein, κ (WaOm(xu)+Ba)=max (0, WaOm(xu)+Ba), It is ReLU and softmax nonlinear activation functions respectively.
The enhancing training sample of whole CNN networks is χRI={ xu|xu∈X∪TφX }, corresponding label is Represent sample xuCorresponding label.New network parameter WRIAnd BRICan be by a kind of new target letter Mathematics acquistion is arrived, and new object function is as follows：
Wherein, λ1And λ2It is weighting parameter, λ1And λ2Belong to [0,1], Section 1 M (χRI,γRI) be softmax layers damage Lose function.It is defined as
It isAnd logOb(xu) between Inner product, N is the sum of initial training sample in X, and K is pair In xuThe total degree of ∈ X rotation transformations, K=35 in this patent；
Section 2 R (X, TφX it is) a rotational invariance normalized constraint item, divides before rotation with postrotational training sample Wei not X and TφX, allow them that there is similar feature.Defining regularization constraint item is
Wherein Oa(xu) be FCa layers activation primitive output,It is the FCa layers activation based on sample after rotation The average output of function.It is defined as
Section 3It is weights attenuation term, for preventing over-fitting.Object function is：
Step 4：Full connection Fisher is added after invariable rotary layer above and distinguishes diagnostic horizon FCc, passes through input-result The CNN network weight parameters W after addition Fisher distinguishes diagnostic horizon is updated with object functionFD={ W1,W2,…,Wm,Wa,Wc,Wd} And BFD={ B1,B2,…,Bm,Ba,Bc,Bd}；
(1) Fisher is added after superincumbent invariable rotary layer FCa and distinguishes layer FCc, by invariable rotary floor and Fisherr areas Diagnostic horizon is divided to combine, CNN training samples are all real border frames of each object class,
It is defined as It is e-th of object class real border frame.Training sample is χFD={ xv, true tag isWith paired input-label (χFD,γFD) training with the addition of Fisher distinguishes the CNN network models of diagnostic horizon；
(2) random initializtion (Wc,Bc) and (Wd,Bd), counting loss function, distinguishing rule item and object function JFD (WFD,BFD), the parameter that upgrades in time WFD={ W1,W2,…,Wm,Wa,Wc,WdAnd BFD={ B1,B2,…,Bm,Ba,Bc,Bd, use Stochastic gradient descent method optimizes structure；To training sample xk∈χFD, Oa(xv) it is expressed as preceding layer FCa output, Oc(xv) be FCd layers export, Od(xv) it is expressed as softmax classification layers FCd output, (Wc,Bc) and (Wd,Bd) FCc layers and FCd are represented respectively The new parameter of layer.
Oc(xv) and Od(xv) calculation formula be respectively：
Oc(xv)=κ (WcOa(xv)+Bc)
Add the CNN network model object functions that Fisherr distinguishes diagnostic horizon
Wherein, λ3And λ4It is weighting parameter, λ3And λ4Belong to [0,1], Section 1 M (χFD,γFD) be softmax layers damage Lose function.It is defined as
|χFD| it is training sample χFDQuantity,
Section 2 F (χFD) it is the discrimination regularization constraint applied to CNN features, S is spaced in class by minimizingW(χFD) and S is spaced between maximization classB(χFD) obtain.
neRepresent the quantity of sample in e-th of object class, wherein meRepresented respectively with mAnd χFDAverage mark sheet Show；
Distinguish regularization term F (χFD) be
F(χFD)=tr (SW(χFD))-tr(SB(χFD))
The summation of the mark computing, i.e. matrix the elements in a main diagonal of tr () representing matrix；
The object function of the step is integrated into
Step 5：Extract the characteristic pattern of source domain vehicle and target domain vehicle respectively using above-mentioned improvement CNN networks；Will Source domain strengthens sample and the vehicle image of target domain training sample is separately input to by improving the CNN network models trained In, the feature of two field vehicle images is extracted, obtains corresponding characteristic pattern.
Assuming that the collection of input feature vector figure is combined intoOutput characteristic atlas is combined into Wherein Rk-1And RkIt is set x respectivelykAnd zkThe number of element, k represent the number of plies of convolution.Due to Alexnet network convolutional layer bags Convolution unit and sub-sample unit are included, the intermediate features atlas between this Unit two is expressed asOften Individual characteristic pattern ykIt is calculated as in convolution unit Represent characteristic pattern xkAnd convolution kernelBetween Convolution,Represent biasing.F (x)=max (0, x) is activation primitive.After convolution unit, sub-sample unit uses fixed big Every characteristic pattern of Nuclear receptor co repressor in small averagePond characteristic pattern corresponding to formation Convolution kernel CollectionCharacteristic pattern is accordingly
Step 6：The cosine similarity between target domain training sample and source domain enhancing training sample characteristic pattern is calculated, The shared convolution kernel of each convolutional layer of CNN networks or unshared convolution kernel are determined, retains each convolutional layer and shares convolution kernel Weight and biasing, based on target domain sample, using stochastic gradient descent method update the unshared convolution kernel of convolutional layer weight and Biasing.
By vehicle image enhancing training sample χRIIt is expressed asThat is source domain enhancing training sample,For target domain training sample.
The vehicle image for having same alike result in A is divided into η classes by K mean cluster algorithm, i.e.,Each Class set AμIn vehicle image obtain similar characteristic pattern by convolutional layer, be calculated as per a kind of average characteristics figure| | represent the number of data.Then found and target figure in the η class images divided As tjThe characteristic pattern being consistent, μ classes and target image tjWith maximum cosine average similarity, computational methods areCos () represents the cosine similarity between two images.It Afterwards, convolution kernel f is estimated by following rulel kWhether it is each tjShared convolution kernel.
Wherein ε is decision-making value.IfThen based on most of in target domain B Sample all thinks this layerIt is shared convolution kernel, retainsWeight and biasing；If Then need to updateWeight and biasing.Loss function is expressed asAllow characteristic patternApproach its corresponding Feature MappingSo as to update unshared convolution kernel collectionMake the parameter of grader can be Used in aiming field.Unshared convolution kernel is updated using stochastic gradient descent algorithm
Step 7：Image in target domain training sample is subjected to convolution, calculates the cosine similarity between characteristic pattern, from And the average similarity matrix of target domain is obtained, clustered using analytic hierarchy process (AHP) (AHP), so as to every a kind of similar features figure only Retain one, simplify structure, accelerate detection speed.The order for optimizing structure will from back to front.
By the vehicle image of target domain, substitution trains CNN networks, obtains convolution characteristic pattern, calculates between characteristic pattern Cosine similarity, and be combined into similarity matrix P and q is corresponding spy Levy the index of figure.Then the average similarity matrix computations of target domain image areT represents target domain sample Quantity, according toCharacteristic pattern is clustered using analytic hierarchy process (AHP) (AHP).Merge the Feature Mapping similar with layer, will Similar convolution kernel only retains one in each convolutional layer of CNN networks based on target domain sample, and it is superfluous to remove each convolutional layer Remaining convolution kernel, accelerate detection speed；
Assuming that characteristic patternAnd characteristic patternBelong to identical class,QuiltInstead of.According to formula, convolution kernel collectionOnly use In formationWhenQuiltInstead of when,It can delete.On the other hand, next stage withThe filter of connection can not be straight Connect deletion.
AllowIt can obtain
IfIt can obtain
Next stage withThe convolution kernel of connection can be deleted after new convolution kernel is added to.
Pay attention to, if with reference to the characteristic pattern of the last stage, convolutionIt should be replaced by matrix inner products,WithIt is a part for grader weight.
Step 8：Whole CNN networks are finely tuned with the training sample of target domain renewal.
Some vehicles in source domain enhancing training sample A and the vehicle in target domain training sample E have phase As distribution character, can expand as the new training sample of target domain.Use the ratio λ of target density and source densityωAs To source domain sample sωAvailability discrimination standard.λωUse condition definition of probability is
p(sω| E) and p (sω| A) s is represented respectivelyωProbability in source domain and target domain, conditional probability modeling are as follows
ot(sω) and os(sω) it is illustrated respectively in the target domain detector of previous migration and the s of source domain detectorωIt is defeated Go out value.If sωCan correctly it be classified by cross-cutting, then p (sω| E) ∈ (0.8,1) and p (sω| A) ∈ (0.8,1), λωBy again It is defined as
If λω>=1, sωThere is similar view to the vehicle in target domain, and micro-adjustment feature can be helped to extract Model.In source domain A, by conditional probability λωSample and target domain E samples when >=1 constitute new sample data set D, In each iterative process, calculate the availability of each vehicle in A and select new Sample Refreshment training set D.Again with renewal Training set D retraining obtains optimization feature detection model each layer of weighting parameter W' and B' after convergence.
Step 9：Vehicle cab recognition, each test sample in target domain test sample set M is substituted into the vehicle and known Other network, FC8 layers are last full linking layers, and the output of full articulamentum is sent to softmax classification layers, produces one and cover The distribution of 4 class labels；When inputting vehicle image, four probability numbers can be obtained, represent the vehicle for motor bus, bulk production respectively Car, box buggy, the probability of car, which maximum probability, any class vehicle be just judged to；Input a vehicle image, output four Individual numerical value, if the numerical value of motor bus is maximum, judge the vehicle for motor bus；If the numerical value of truck is maximum, judge that the vehicle is Truck；If the numerical value of box buggy is maximum, it is box buggy to judge the vehicle；If the numerical value of car is maximum, judging should Vehicle is car.
The foregoing is only a preferred embodiment of the present invention, but protection scope of the present invention be not limited thereto, Any one skilled in the art the invention discloses technical scope in, the change or replacement that can readily occur in, It should all be included within the scope of the present invention.Therefore, protection scope of the present invention should be with scope of the claims It is defined.

Claims (10)

1. a kind of model recognizing method based on CNN and domain adaptive learning, it is characterised in that comprise the following steps：
Step 1) gathers the vehicle image for including various under natural scene respectively, forms vehicle image data base, described A part is source domain sample in image, and source domain sample includes source domain training and source domain test sample, remainder are Target domain sample object field sample is trained including target domain and source domain test sample, to described after collection vehicle image Vehicle pictures in vehicle image data base are pre-processed；
Step 2) builds CNN network models, and the vehicle image data base is imported into Alexnet networks carries out pre-training, renewal The weight parameter W of Alexnet network structuresiWith offset parameter Bi, WiRepresent the weight of the i-th layer network, BiRepresent the i-th layer network Biasing, i=1,2 ... m；
Step 3) adds invariable rotary layer FCa in CNN networks, rotates source domain training sample, obtains source domain enhancing training Sample, training sample and first object function are strengthened according to the source domain, the network weight after invariable rotary layer is added in training Parameter WRI={ W1,W2,…,Wm,Wa,WbAnd biasing BRI={ B1,B2,…,Bm,Ba,Bb}；
Step 4), which adds Fisher after invariable rotary layer and distinguishes layer FCc and formed, improves CNN network models, by input label and Second object function, training renewal with the addition of Fisher and distinguish the CNN network weight parameters W after diagnostic horizonFD={ W1,W2,…, Wm,Wa,Wc,WdAnd BFD={ B1,B2,…,Bm,Ba,Bc,Bd}；
Step 5) extracts target domain training sample and source domain respectively by the improvement CNN network models strengthens training sample Characteristic pattern, by source domain strengthen training sample and target domain training sample vehicle image be separately input to it is trained Improve in CNN network models, extract the feature of the vehicle image of source domain and target domain respectively, obtain corresponding characteristic pattern；
The cosine that step 6) is calculated between target domain training sample characteristic pattern and source domain enhancing training sample characteristic pattern is similar Degree, determines the shared convolution kernel of improved CNN networks convolutional layer or unshared convolution kernel, retains each convolutional layer and shares convolution kernel Weight and biasing, based on target domain training sample, use the stochastic gradient descent method renewal unshared convolution kernel of convolutional layer Weight and biasing；
Step 7) convolution target domain training sample image, characteristic pattern is obtained, calculate the cosine similarity between characteristic pattern, obtained The average similarity matrix of target domain training sample, is clustered using analytic hierarchy process (AHP), and a similar features figure is retained per class；
Step 8) finely tunes whole CNN networks by the target domain training sample, ultimately forms vehicle cab recognition network；
Each test sample in the target domain test sample is substituted into the vehicle cab recognition network, vehicle net by step 9) Network judges the vehicle of vehicle according to the output result of vehicle cab recognition network.
2. the model recognizing method according to claim 1 based on CNN and domain adaptive learning, it is characterised in that the step It is rapid 1) in, a part is source domain sample in described image, including CNN training sample set X={ x1,x2,…,xhAnd CNN network model test sample set R={ γ1,γ2,…,γσ, xh、γσThe sample in source domain set, h, σ are represented respectively The number of sample in source domain set is represented respectively；Remainder is target domain sample, including domain adaptive learning training sample This setWith domain adaptive learning test sample set M={ δ1,δ2,…,δτ, tj、δτTarget domain is represented respectively Sample, T, τ represent the number of sample in target domain set respectively.
3. the model recognizing method according to claim 1 based on CNN and domain adaptive learning, it is characterised in that step 1) In, the pretreatment operation unifies picture size, and adds label to image, marks the affiliated vehicle classification of the image.
4. the model recognizing method according to claim 3 based on CNN and domain adaptive learning, it is characterised in that the step It is rapid 2) in, Alexnet network structures are made up of five convolutional layers sequentially set gradually with three full articulamentums, Alexnet networks export the softmax loss functions of fusion tag, according to softmax loss functions to weight parameter WiWith it is inclined Put parameter BiStochastic gradient descent fine setting is carried out, the fusion tag is real for the output result of Alexnet networks and the label Error between actual value.
5. the model recognizing method according to claim 4 based on CNN and domain adaptive learning, it is characterised in that the step It is rapid 3) to comprise the following steps：
The source domain training sample of vehicle image 3-1) is set as X, rotation transformation is carried out to each vehicle image with angle φ, and It is added in former source domain training sample X, obtains source domain enhancing training sample χRI, χRI={ X, TφX } it is one group of image rotation The data set of conversion, wherein Represent that it is φ to carry out angle to training samplekRotation process；
Full connection rotation 3-2) is added after the full articulamentum of Alexnet network structure layer 7s in CNN network models not Change layer FCa, by Om(xu) it is expressed as preceding layer FCm output, Oa(xu) exported for FCa layers, Ob(xu) it is expressed as softmax classification Layer FCb output, (Wa,Ba) and (Wb,Bb) new parameter of FCa layers and FCb layers is represented respectively；
Oa(xu) and Ob(xu) calculation formula respectively such as formula (1), formula (2)：
Oa(xu)=κ (WaOm(xu)+Ba) (1)
Wherein, κ (WaOm(xu)+Ba)=max (0, WaOm(xu)+Ba),
It is that ReLU and softmax are non-respectively Linear activation primitive；
The enhancing training sample of whole CNN network models is set as χRI={ X, TφX }, corresponding label is Represent sample xuCorresponding label；New network parameter WRIAnd BRILearnt by new object function Arrive, new object function such as formula (3)：
Wherein, λ1And λ2It is weighting parameter, λ1And λ2Belong to [0,1], M (χRI,γRI) be softmax layers loss function, table Up to formula such as formula (4)；R(X,TφX it is) a rotational invariance normalized constraint item, expression formula such as formula (5), before rotation and after rotation Training sample be respectively X and TφX,It is weights attenuation term, for preventing over-fitting；
In formula (4),It isAnd logOb(xu) between Inner products, N is the sum of initial training sample in X, K It is for xuThe total degree of ∈ X rotation transformations,
Wherein, Oa(xu) be FCa layers activation primitive output,It is the FCa layer activation primitives based on sample after rotation Average output, such as formula (6)；
The first object function such as formula (7)：
6. the model recognizing method according to claim 1 based on CNN and domain adaptive learning, it is characterised in that the step It is rapid 4) to comprise the following steps：
Fisher 4-1) is added after invariable rotary layer FCa and distinguishes diagnostic horizon FCc, invariable rotary layer and Fisher are distinguished Diagnostic horizon is combined, and CNN training samples are all real border frames of each object class, are defined as It is e-th of object class real border frame；Training sample is χFD={ xv, it is corresponding Output result isBy input results to (χFD,γFD) training；
4-2) random initializtion (Wc,Bc) and (Wd,Bd), counting loss function, distinguishing rule item and the second object function JFD (WFD, BFD), the parameter that upgrades in time WFD={ W1,W2,…,Wm,Wa,Wc,WdAnd BFD={ B1,B2,…,Bm,Ba,Bc,Bd, use Stochastic gradient descent method optimizes structure；To training sample xk∈χFD, Oa(xv) it is expressed as preceding layer FCa output, Oc(xv) be FCd layers export, Od(xv) it is expressed as softmax classification layers FCd output, (Wc,Bc) and (Wd,Bd) FCc layers and FCd are represented respectively The new parameter of layer；
Oc(xv) and Od(xv) calculation formula respectively such as formula (8), formula (9)：
Oc(xv)=κ (WcOa(xv)+Bc) (8)
The second target object function such as formula (10)
Wherein, λ3And λ4It is weighting parameter, λ3And λ4Belong to [0,1], M (χFD,γFD) be softmax layers loss function, M (χFD,γFD) expression formula such as formula (11)；F(χFD) it is the discrimination regularization constraint applied to CNN features, between minimizing in class Every SW(χFD) maximize class between be spaced SB(χFD) obtain, SW(χFD)、SB(χFD) expression formula respectively such as formula (12), formula (13)；
In formula (11), | χFD| it is training sample χFDQuantity,
In formula (12), formula (13), neRepresent the quantity of sample in e-th of object class, wherein meRepresented respectively with mAnd χFDIt is flat Equal feature such as formula (14), formula (15)；
Distinguish regularization term F (χFD) such as formula (16),
F(χFD)=tr (SW(χFD))-tr(SB(χFD)) (16)
In formula (16), tr (SW(χFD)) and tr (SB(χFD)) summation of the mark computing, i.e. matrix the elements in a main diagonal of representing matrix；
Second object function is integrated into：
7. the model recognizing method according to claim 4 based on CNN and domain adaptive learning, it is characterised in that the step 5) in, it is assumed that the collection of input feature vector figure is combined intoOutput characteristic atlas is combined into Wherein Rk-1And RkIt is set x respectivelykAnd zkThe number of element, k represent the number of plies of convolution；Alexnet networks convolutional layer includes convolution Unit and sub-sample unit, the intermediate features atlas between convolution unit and sub-sample unit are expressed as Each characteristic pattern ykIt is calculated as in convolution unit Represent characteristic pattern xkAnd convolution kernelIt Between convolution,Biasing is represented, F (x)=max (0, x) is activation primitive, and sub-sampling layer uses fixed big after convolution unit Every characteristic pattern of Nuclear receptor co repressor in small averagePond characteristic pattern corresponding to formation Convolution kernel CollectionCharacteristic pattern is accordingly
8. the model recognizing method according to claim 1 based on CNN and domain adaptive learning, it is characterised in that the step Suddenly in (6), vehicle image is strengthened into training sample χRIIt is expressed asThat is source domain enhancing training sample,For target domain training sample；
The vehicle image for having same alike result in A is divided into η classes by K mean cluster algorithm, i.e.,Each class set Aμ In vehicle image obtain similar characteristic pattern by convolutional layer, be calculated as per a kind of average characteristics figure |Aμ| the number of data is represented, is found and target image t in the η class images dividedjThe characteristic pattern being consistent, μ classes and target image tjWith maximum cosine average similarity, computational methods areCos () table Show the cosine similarity between two images；
Setting estimation convolution kernel fl kWhether it is each tjShared convolution kernel decision condition formula such as formula (17)；
In formula (17), ε is decision-making value, and the decision rule of convolution (17) is：IfThen The f of this layer is all thought based on most numerical example in target domain training samplel kIt is shared convolution kernel, retains fl kWeight and Biasing；IfThen need to update fl kWeight and biasing；Softmax loss functions are expressed asSo that characteristic patternApproach its corresponding Feature MappingIt is non-common so as to update Enjoy convolution kernel collection fl k。
9. the model recognizing method according to claim 1 based on CNN and domain adaptive learning, it is characterised in that the step Suddenly in (8), the ratio λ of target density and source density is usedωAs to source domain sample sωAvailability discrimination standard, λωUse Conditional probability is defined such as formula (18)
p(sω| E) and p (sω| A) s is represented respectivelyωProbability in source domain and target domain, conditional probability are modeled such as formula (19), formula (20),
ot(sω) and os(sω) it is illustrated respectively in the target domain detector of previous migration and the s of source domain detectorωOutput Value, if sωCorrectly classified by cross-cutting, then p (sω| E) ∈ (0.8,1) and p (sω| A) ∈ (0.8,1), λωBe redefined as Formula (21),
If λω>=1, sωThere is similar view to the vehicle in target domain；In source domain, by conditional probability λωWhen >=1 Sample and target domain sample constitute new sample data set, in each iterative process, calculate source domain in each car Availability and select new Sample Refreshment training set, with the training set retraining of renewal until the softmax of fusion tag Loss function is restrained, and obtains optimization feature detection model each layer of weighting parameter W' and B'.
10. the model recognizing method according to claim 4 based on CNN and domain adaptive learning, it is characterised in that described In step 9), last full linking layer of the vehicle cab recognition network is FC8 layers, and the output of FC8 layers is sent to softmax points Class layer, the output of the softmax classification layer is a probability distribution for covering four class vehicles, is corresponded in output result a certain The maximum probability of vehicle, the then output result for judging vehicle to be identified are the vehicle, and the four classes vehicle includes motor bus, bulk production Car, box buggy and car.
CN201710670157.9A 2017-08-07 2017-08-07 Vehicle type identification method based on CNN and domain adaptive learning Active CN107527068B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710670157.9A CN107527068B (en) 2017-08-07 2017-08-07 Vehicle type identification method based on CNN and domain adaptive learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710670157.9A CN107527068B (en) 2017-08-07 2017-08-07 Vehicle type identification method based on CNN and domain adaptive learning

Publications (2)

Publication Number Publication Date
CN107527068A true CN107527068A (en) 2017-12-29
CN107527068B CN107527068B (en) 2020-12-25

Family

ID=60680760

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710670157.9A Active CN107527068B (en) 2017-08-07 2017-08-07 Vehicle type identification method based on CNN and domain adaptive learning

Country Status (1)

Country Link
CN (1) CN107527068B (en)

Cited By (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108229589A (en) * 2018-02-09 2018-06-29 天津师范大学 A kind of ground cloud atlas sorting technique based on transfer learning
CN108491880A (en) * 2018-03-23 2018-09-04 西安电子科技大学 Object classification based on neural network and position and orientation estimation method
CN108492575A (en) * 2018-04-11 2018-09-04 济南浪潮高新科技投资发展有限公司 A kind of intelligent vehicle type identifier method
CN108921201A (en) * 2018-06-12 2018-11-30 河海大学 Dam defect identification and classification method based on feature combination with CNN
CN109033993A (en) * 2018-06-29 2018-12-18 南京行者易智能交通科技有限公司 A kind of method and device of image recognition detection switch door
CN109063719A (en) * 2018-04-23 2018-12-21 湖北工业大学 A kind of image classification method of co-ordinative construction similitude and category information
CN109284590A (en) * 2018-09-29 2019-01-29 武汉极意网络科技有限公司 Access method, equipment, storage medium and the device of behavior safety protection
CN109359684A (en) * 2018-10-17 2019-02-19 苏州大学 Fine granularity model recognizing method based on Weakly supervised positioning and subclass similarity measurement
CN109359666A (en) * 2018-09-07 2019-02-19 佳都新太科技股份有限公司 A kind of model recognizing method and processing terminal based on multiple features fusion neural network
CN109657552A (en) * 2018-11-16 2019-04-19 北京邮电大学 The vehicle type recognition device being cold-started across scene and method are realized based on transfer learning
CN109711529A (en) * 2018-11-13 2019-05-03 中山大学 A kind of cross-cutting federal learning model and method based on value iterative network
CN109840518A (en) * 2018-09-07 2019-06-04 电子科技大学 A kind of visual pursuit method of combining classification and domain adaptation
CN110147837A (en) * 2019-05-14 2019-08-20 中国电子科技集团公司第二十八研究所 The intensive object detection method of any direction, system and the equipment focused based on feature
CN110287847A (en) * 2019-06-19 2019-09-27 长安大学 Vehicle grading search method based on Alexnet-CLbpSurf multiple features fusion
WO2019200748A1 (en) * 2018-04-17 2019-10-24 平安科技（深圳）有限公司 Transfer learning method, device, computer device, and storage medium
CN110569764A (en) * 2019-08-28 2019-12-13 北京工业大学 mobile phone model identification method based on convolutional neural network
CN110674689A (en) * 2019-08-19 2020-01-10 浙江省北大信息技术高等研究院 Vehicle re-identification method and system based on feature embedding space geometric constraint
CN110751869A (en) * 2019-10-12 2020-02-04 南京摄星智能科技有限公司 Simulated environment and battlefield situation strategy transfer technology based on countermeasure discrimination migration method
CN110825381A (en) * 2019-09-29 2020-02-21 南京大学 CNN-based bug positioning method combining source code semantics and grammatical features
CN110942323A (en) * 2018-09-25 2020-03-31 优估（上海）信息科技有限公司 Evaluation model construction method, device and system
CN111079833A (en) * 2019-12-16 2020-04-28 腾讯科技（深圳）有限公司 Image recognition method, image recognition device and computer-readable storage medium
CN111144464A (en) * 2019-12-17 2020-05-12 金陵科技学院 Fruit automatic identification method based on CNN-Kmeans algorithm
CN111488972A (en) * 2020-04-09 2020-08-04 北京百度网讯科技有限公司 Data migration method and device, electronic equipment and storage medium
CN111860582A (en) * 2020-06-11 2020-10-30 北京市威富安防科技有限公司 Image classification model construction method and device, computer equipment and storage medium
CN111914911A (en) * 2020-07-16 2020-11-10 桂林电子科技大学 Vehicle re-identification method based on improved depth relative distance learning model
CN112990315A (en) * 2021-03-17 2021-06-18 北京大学 3D shape image classification method of equal-variation 3D convolution network based on partial differential operator
CN114677545A (en) * 2022-03-29 2022-06-28 电子科技大学 Lightweight image classification method based on similarity pruning and efficient module
US11580405B2 (en) 2019-12-26 2023-02-14 Sap Se Domain adaptation of deep neural networks
WO2024045319A1 (en) * 2022-08-30 2024-03-07 北京龙智数科科技服务有限公司 Face image clustering method and apparatus, electronic device, and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030063796A1 (en) * 2001-09-28 2003-04-03 Koninklijke Philips Electronics N.V. System and method of face recognition through 1/2 faces
CN103425996A (en) * 2013-08-01 2013-12-04 华南理工大学 Parallel distribution type large-scale image recognition method
US20160140425A1 (en) * 2014-11-14 2016-05-19 Thomson Licensing Method and apparatus for image classification with joint feature adaptation and classifier learning
CN106250812A (en) * 2016-07-15 2016-12-21 汤平 A kind of model recognizing method based on quick R CNN deep neural network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030063796A1 (en) * 2001-09-28 2003-04-03 Koninklijke Philips Electronics N.V. System and method of face recognition through 1/2 faces
CN103425996A (en) * 2013-08-01 2013-12-04 华南理工大学 Parallel distribution type large-scale image recognition method
US20160140425A1 (en) * 2014-11-14 2016-05-19 Thomson Licensing Method and apparatus for image classification with joint feature adaptation and classifier learning
CN106250812A (en) * 2016-07-15 2016-12-21 汤平 A kind of model recognizing method based on quick R CNN deep neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
戴礼荣等: "基于深度学习的语音识别技术现状与展望", 《数据采集与处理》 *
茅正冲等: "计算机视觉技术的车型识别应用研究", 《单片机与嵌入式系统应用》 *

Cited By (43)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108229589A (en) * 2018-02-09 2018-06-29 天津师范大学 A kind of ground cloud atlas sorting technique based on transfer learning
CN108229589B (en) * 2018-02-09 2021-10-15 天津师范大学 Foundation cloud picture classification method based on transfer learning
CN108491880A (en) * 2018-03-23 2018-09-04 西安电子科技大学 Object classification based on neural network and position and orientation estimation method
CN108492575A (en) * 2018-04-11 2018-09-04 济南浪潮高新科技投资发展有限公司 A kind of intelligent vehicle type identifier method
WO2019200748A1 (en) * 2018-04-17 2019-10-24 平安科技（深圳）有限公司 Transfer learning method, device, computer device, and storage medium
CN109063719A (en) * 2018-04-23 2018-12-21 湖北工业大学 A kind of image classification method of co-ordinative construction similitude and category information
CN109063719B (en) * 2018-04-23 2022-02-01 湖北工业大学 Image classification method combining structure similarity and class information
CN108921201A (en) * 2018-06-12 2018-11-30 河海大学 Dam defect identification and classification method based on feature combination with CNN
CN109033993A (en) * 2018-06-29 2018-12-18 南京行者易智能交通科技有限公司 A kind of method and device of image recognition detection switch door
CN109840518B (en) * 2018-09-07 2022-03-25 电子科技大学 Visual tracking method combining classification and domain adaptation
CN109359666A (en) * 2018-09-07 2019-02-19 佳都新太科技股份有限公司 A kind of model recognizing method and processing terminal based on multiple features fusion neural network
CN109840518A (en) * 2018-09-07 2019-06-04 电子科技大学 A kind of visual pursuit method of combining classification and domain adaptation
CN110942323A (en) * 2018-09-25 2020-03-31 优估（上海）信息科技有限公司 Evaluation model construction method, device and system
CN109284590A (en) * 2018-09-29 2019-01-29 武汉极意网络科技有限公司 Access method, equipment, storage medium and the device of behavior safety protection
CN109359684A (en) * 2018-10-17 2019-02-19 苏州大学 Fine granularity model recognizing method based on Weakly supervised positioning and subclass similarity measurement
CN109359684B (en) * 2018-10-17 2021-10-29 苏州大学 Fine-grained vehicle type identification method based on weak supervision positioning and subcategory similarity measurement
CN109711529A (en) * 2018-11-13 2019-05-03 中山大学 A kind of cross-cutting federal learning model and method based on value iterative network
CN109711529B (en) * 2018-11-13 2022-11-08 中山大学 Cross-domain federated learning model and method based on value iterative network
CN109657552A (en) * 2018-11-16 2019-04-19 北京邮电大学 The vehicle type recognition device being cold-started across scene and method are realized based on transfer learning
CN109657552B (en) * 2018-11-16 2022-11-29 北京邮电大学 Vehicle type recognition device and method for realizing cross-scene cold start based on transfer learning
CN110147837A (en) * 2019-05-14 2019-08-20 中国电子科技集团公司第二十八研究所 The intensive object detection method of any direction, system and the equipment focused based on feature
CN110287847A (en) * 2019-06-19 2019-09-27 长安大学 Vehicle grading search method based on Alexnet-CLbpSurf multiple features fusion
CN110674689A (en) * 2019-08-19 2020-01-10 浙江省北大信息技术高等研究院 Vehicle re-identification method and system based on feature embedding space geometric constraint
CN110674689B (en) * 2019-08-19 2022-05-06 杭州未名信科科技有限公司 Vehicle re-identification method and system based on feature embedding space geometric constraint
CN110569764A (en) * 2019-08-28 2019-12-13 北京工业大学 mobile phone model identification method based on convolutional neural network
CN110569764B (en) * 2019-08-28 2023-12-22 北京工业大学 Mobile phone model identification method based on convolutional neural network
CN110825381A (en) * 2019-09-29 2020-02-21 南京大学 CNN-based bug positioning method combining source code semantics and grammatical features
CN110751869A (en) * 2019-10-12 2020-02-04 南京摄星智能科技有限公司 Simulated environment and battlefield situation strategy transfer technology based on countermeasure discrimination migration method
CN110751869B (en) * 2019-10-12 2021-11-05 南京星耀智能科技有限公司 Simulated environment and battlefield situation strategy transfer technology based on countermeasure discrimination migration method
CN111079833B (en) * 2019-12-16 2022-05-06 腾讯医疗健康(深圳)有限公司 Image recognition method, image recognition device and computer-readable storage medium
CN111079833A (en) * 2019-12-16 2020-04-28 腾讯科技（深圳）有限公司 Image recognition method, image recognition device and computer-readable storage medium
CN111144464A (en) * 2019-12-17 2020-05-12 金陵科技学院 Fruit automatic identification method based on CNN-Kmeans algorithm
US11580405B2 (en) 2019-12-26 2023-02-14 Sap Se Domain adaptation of deep neural networks
CN111488972B (en) * 2020-04-09 2023-08-08 北京百度网讯科技有限公司 Data migration method, device, electronic equipment and storage medium
CN111488972A (en) * 2020-04-09 2020-08-04 北京百度网讯科技有限公司 Data migration method and device, electronic equipment and storage medium
CN111860582A (en) * 2020-06-11 2020-10-30 北京市威富安防科技有限公司 Image classification model construction method and device, computer equipment and storage medium
CN111860582B (en) * 2020-06-11 2021-05-11 北京市威富安防科技有限公司 Image classification model construction method and device, computer equipment and storage medium
CN111914911B (en) * 2020-07-16 2022-04-08 桂林电子科技大学 Vehicle re-identification method based on improved depth relative distance learning model
CN111914911A (en) * 2020-07-16 2020-11-10 桂林电子科技大学 Vehicle re-identification method based on improved depth relative distance learning model
CN112990315A (en) * 2021-03-17 2021-06-18 北京大学 3D shape image classification method of equal-variation 3D convolution network based on partial differential operator
CN112990315B (en) * 2021-03-17 2023-10-20 北京大学 3D shape image classification method of constant-variation 3D convolution network based on partial differential operator
CN114677545A (en) * 2022-03-29 2022-06-28 电子科技大学 Lightweight image classification method based on similarity pruning and efficient module
WO2024045319A1 (en) * 2022-08-30 2024-03-07 北京龙智数科科技服务有限公司 Face image clustering method and apparatus, electronic device, and storage medium

Also Published As

Publication number Publication date
CN107527068B (en) 2020-12-25

Similar Documents

Publication Publication Date Title
CN107527068A (en) Model recognizing method based on CNN and domain adaptive learning
CN111368896B (en) Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network
CN106250812B (en) A kind of model recognizing method based on quick R-CNN deep neural network
WO2020232905A1 (en) Superobject information-based remote sensing image target extraction method, device, electronic apparatus, and medium
CN107392925B (en) Remote sensing image ground object classification method based on super-pixel coding and convolutional neural network
CN111079674B (en) Target detection method based on global and local information fusion
CN107133974B (en) Gaussian Background models the vehicle type classification method combined with Recognition with Recurrent Neural Network
DE112020005609T5 (en) Domain adaptation for semantic segmentation by exploiting weak labels
CN104134068B (en) Monitoring vehicle characteristics based on sparse coding represent and sorting technique
CN108549926A (en) A kind of deep neural network and training method for refining identification vehicle attribute
EP3029606A2 (en) Method and apparatus for image classification with joint feature adaptation and classifier learning
CN106920243A (en) The ceramic material part method for sequence image segmentation of improved full convolutional neural networks
CN113239924B (en) Weak supervision target detection method and system based on transfer learning
CN110717553A (en) Traffic contraband identification method based on self-attenuation weight and multiple local constraints
CN107341447A (en) A kind of face verification mechanism based on depth convolutional neural networks and evidence k nearest neighbor
CN111160553B (en) Novel field self-adaptive learning method
US20220301173A1 (en) Method and system for graph-based panoptic segmentation
CN106909902A (en) A kind of remote sensing target detection method based on the notable model of improved stratification
CN110281949B (en) Unified hierarchical decision-making method for automatic driving
CN109977994A (en) A kind of presentation graphics choosing method based on more example Active Learnings
EP3752955A1 (en) Image segmentation
CN107103308A (en) A kind of pedestrian's recognition methods again learnt based on depth dimension from coarse to fine
CN104954741A (en) Tramcar on-load and no-load state detecting method and system based on deep-level self-learning network
CN110852358A (en) Vehicle type distinguishing method based on deep learning
CN108345900A (en) Pedestrian based on color and vein distribution characteristics recognition methods and its system again

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
CB02 Change of applicant information
CB02 Change of applicant information

Address after: 210044 No. 219 Ning six road, Jiangbei new district, Nanjing, Jiangsu

Applicant after: Nanjing University of Information Science and Technology

Address before: No. 69, Jianye District, Jianye District, Nanjing, Jiangsu

Applicant before: Nanjing University of Information Science and Technology

GR01 Patent grant
GR01 Patent grant