CN104573731B - Fast target detection method based on convolutional neural networks - Google Patents
Fast target detection method based on convolutional neural networks Download PDFInfo
- Publication number
- CN104573731B CN104573731B CN201510061852.6A CN201510061852A CN104573731B CN 104573731 B CN104573731 B CN 104573731B CN 201510061852 A CN201510061852 A CN 201510061852A CN 104573731 B CN104573731 B CN 104573731B
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- layer
- sample
- image
- 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.)
- Expired - Fee Related
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 64
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 55
- 238000000034 method Methods 0.000 claims abstract description 50
- 238000012549 training Methods 0.000 claims abstract description 39
- 238000009499 grossing Methods 0.000 claims abstract description 10
- 238000010586 diagram Methods 0.000 claims description 35
- 230000004044 response Effects 0.000 claims description 31
- 238000011176 pooling Methods 0.000 claims description 22
- 239000000203 mixture Substances 0.000 claims description 18
- 238000005070 sampling Methods 0.000 claims description 18
- 230000004069 differentiation Effects 0.000 claims description 16
- 238000010606 normalization Methods 0.000 claims description 14
- 238000012360 testing method Methods 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 10
- 238000012937 correction Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 7
- 230000014509 gene expression Effects 0.000 claims description 7
- 239000013598 vector Substances 0.000 claims description 7
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 5
- 230000001629 suppression Effects 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 4
- 238000013461 design Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000002360 preparation method Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 241001465754 Metazoa Species 0.000 description 1
- 102000008297 Nuclear Matrix-Associated Proteins Human genes 0.000 description 1
- 108010035916 Nuclear Matrix-Associated Proteins Proteins 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 239000012634 fragment Substances 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 210000000299 nuclear matrix Anatomy 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Abstract
Fast target detection method based on convolutional neural networks, is related to computer vision technique.Convolutional neural networks parameter is trained first with training set, then solving the problems, such as that maximum pondization is lost feature and generated using the mode of expander graphs differentiates complete characteristics figure;The full connection weight of convolutional neural networks is regarded as linear classifier, general magnificent error of the linear classifier on complete characteristics are differentiated is estimated using possible approximate learning framework;Linear classifier number needed for being estimated according to general magnificent error and desired extensive error threshold, finally complete target detection with mode of the linear classifier based on smoothing windows on complete characteristics figure is differentiated.Significantly improve detection efficiency and target detection precision.
Description
Technical field
The present invention relates to computer vision technique, is specifically related to a kind of fast target detection based on convolutional neural networks
Method.
Background technology
One important sources in the human perception world are exactly to pass through image information, and research shows, the mankind obtain external information
In about 80%~90% information come from human eye acquisition image information.The mankind perceive energy to extraneous image information
Power is very high, can be with quickly positioning target and analysis target.Computer will have powerful visually-perceptible and understandability, just should
Possesses the powerful Target detection and identification ability of the similar mankind.Target detection is one of visually-perceptible and object understanding critically important
Premise work, the efficiency and precision of Target Acquisition decide the speed and effect of visually-perceptible.Once computer possesses similar
The powerful target detection box perception of the mankind, it is possible to preferably substitute manual work in every profession and trade, greatly save and be produced into
This.It also provides powerful support to daily intelligent Service industry in addition.Therefore, to the target detection skill in computer vision
Art is furtherd investigate, and is improved constantly the accuracy rate of detection, is had important practical significance.
The development trend for solving the two problems in academic circles at present is from heuristic is used to using machine learning
Method.Feature used also turns to the feature of task based access control extracted in self-adaptive from manual feature.The model of Target detection and identification
Also begin to occur the functional shift for detecting and recognizing multi-target detection and identification from single specific objective while carry out.Most typically
Example be exactly deep learning mould appearance, solve conventional target detection and identification model just for limited task target
The problem of detection and identification mission are effective.For example in 2001, the obverse face detection framework that Viola-Jone is proposed was based on
Harr features are quite effective to Face datection, but are not very good for side face face and pedestrian detection effect.Until 2005
Year, Dalal et al. proposes HOG features and using SVM to HOG (Histogram of corresponding to each smoothing windows
Gradient) after the strategy that feature is classified, vertical pedestrian's Detection results just have the breakthrough of a matter.However, HOG this
Manual feature, do not made for the Detection results of the targets such as the pedestrian of image classification and identification and any attitude, animal, plant
People is satisfied with.Then deformation model (Deformable Part Models:DPM the target detection for solving have deformation of) arising at the historic moment is asked
Topic.Although deformation model tries to solve the problems, such as because deformation causes target detection less than but deformations needed in its model
Be difficult preferably to capture in part reality, the feature that reason is become reconciled with regard to the good model of neither one carrys out identification component, therefore it
Effect is not very good on multi-class targets detection data set (PASCAL VOC, ImageNet).A nearest breakthrough sex work
It is the appearance of deep learning model.In the image classification and target detection data set ImageNet of maximum, based on deep learning
The raising for the Target detection and identification precision that one of model convolutional neural networks (CNN) are done is even more to exceed conventional one times of full accuracy
As many as.Nearest 2 years ImageNet data sets classification and the outstanding algorithm almost all of detection performance use convolutional neural networks,
They different network structures is different.ImageNet data sets epigraph classification at present and target detection highest precision difference
For 95% and 55%.
Although very high precision is improved on Target detection and identification based on the method for convolutional neural networks, due to
Convolutional neural networks network is complicated and computationally intensive, and it is not very high to apply the efficiency in target detection, and current many methods are all
It is object detection program to be accelerated based on GPU.A target image is given, target detection is done using smoothing windows strategy,
Even if being accelerated using GPU, its algorithm complex is still very big, extremely inefficient.In order to solve convolutional neural networks in target detection
On efficiency, the solution of academic circles at present main flow can be divided into three classes.The first kind is the method cut based on figure, first right
Given image carries out image segmentation, and some potential target areas are obtained by splitting block.Then with convolutional neural networks to this
A little target areas carry out feature extraction and classification, finally give the position of target.The shortcomings that this method is exactly to become dependent upon figure
As the quality of segmentation.Second class is to extract feature to original image by convolutional neural networks, then with smooth on characteristic pattern
Window strategy does the recurrence of target location and the classification of target.This method is extracting feature using convolutional neural networks to big figure
When, some can be lost to classifying and returning useful feature information, therefore the performance of last model be unable to reach it is optimal.3rd
Class method is then to find part with the advantage of convolutional neural networks classification, and then builds deformation model, using deformation model
Thought detects to target.But what the target detection in this classification and deformation model convolutional neural networks was performed separately
Way so that the Detection results of general frame are not very outstanding, and the efficiency of this model is nor very high in addition.
The content of the invention
It is an object of the invention to propose a kind of fast target detection method based on convolutional neural networks.
The present invention comprises the following steps:
A training sample set (x) is preparedi, yi), i=1 ... ..., N, N is number of training, and N is natural number, xiRepresent training
The image of fixed size corresponding to sample, comprising target and target, the image full of picture is positive sample, and other images are negative samples
This;yiRepresent sample class vector:
B) m batches are divided by all training samples, the convolutional neural networks that m-2 lot samples are originally put into design are calculated with backpropagation
Method is trained, and 2 lot sample sheets are used to test, and the convolutional neural networks include convolutional layer, maximum pond (max-pooling) layer drawn game
Portion's contrast normalization (local contrast normalization) three composition of layer, these three compositions play the part of nonlinear function
Function, it is therefore an objective to the original image in manifold is mapped to theorem in Euclid space, the activation primitive of convolution uses non-linear school
Positive unit (Rectifier linear unit), can make it that the aspect ratio after convolution is sparse, design comprising these three into
After the network structure model divided, model is trained with back-propagation algorithm training, obtains parameter W;
C) extraction step B) in the parameter W trained, solved by the way of expander graphs in convolutional neural networks maximum
The problem of information, is lost in pond (max-pooling), gives test image, conventional maximum pond (max-pooling) is (namely
Using skew A be used as starting point mode) operation after can only obtain expander graphs (A), other be advantageous to classify information can lose,
Therefore for each being offset in 2 × 2 convolution kernel, a corresponding characteristic pattern, referred to as expander graphs are all preserved;For each layer
Max-pooling cores size is K, and whole convolutional neural networks have p layers, each maximum pond (max-pooling) after extension
Down-sampling layer feature map number extend to 2KTimes, whole network expands to o=(2K)pTimes, then parameter W is used for arbitrarily large
Small image to be detected, the extension feature figure before full connection is obtained, it is referred to as to differentiate complete characteristics (Discriminative
Complete Features);Given input picture x, the wave filter group K trained and skew b, the output of convolutional layer can be with
Write as the form of (formula one):
(formula one)
Wherein MjThe index of input feature vector figure selected by expression;L represents the index of current layer;I and j represent respectively input and
Output characteristic index of the picture;O represents segment index;F represent activation primitive, used here as linearity correction unit function f (x)=max (x,
0);* convolution operation is represented;
Maximum pond (max-pooling) layer can be write as the form of (formula two):
(formula two)
Wherein, m and n represents the pixel index of current layer respectively;S represents the size of down-sampling core;Before p and q is represented respectively
One layer of pixel index, and meet p=s* (m-1)+k+1, q=s* (n-1)+k+1, wherein 0 < k < s;Branch is used to select
Corresponding columns and rows;
In order that the feature that must be obtained has the property of Shandong nation, local contrast can be normalized layer and introduce segment processing
In link, the layer can be write as the form of (formula three):
(formula three)
Wherein, the number for the neighbouring competition characteristic pattern that r expressions are formulated, the feature map number of N expressions current layer altogether, k, α,
β is some hyper parameters, can specify suitable floating-point values in the training process, after CNN model trainings are good, three more than
Individual formula, it can obtain and differentiate that complete characteristics provide support to be follow-up based on smoothing windows scanning;
D the full connection of convolutional neural networks) is regarded as a linear classifier, is directly based upon and differentiates that complete characteristics figure is examined
Survey, estimate that linear classifier is differentiating using possible approximate study (probably approximately correct) framework
Extensive error R [ω] on complete characteristics, then according to built-up pattern (ensemble model) and desired extensive error H
Training pattern number q required for [ω] is calculated;
(formula four)
E) each training pattern will be obtained in q model and performs feedforward operation in image to be detected respectively, each mould
Type obtains the differentiation complete characteristics figure before o groups connect entirely, and then every group of differentiation complete characteristics figure is contracted with arest neighbors interpolation algorithm
Put, obtain n*o groups and differentiate complete characteristics figure, q model amounts to n*o*q groups and differentiate complete characteristics figure, then complete in every group of differentiation
Directly intensive smoothing windows classification work is performed on standby characteristic pattern with linear classifier and obtains n*o*q group response diagrams, wherein linear point
Class device and differentiate complete graph dot product operations can change into convolution operation, due to use step B) in non-linear correction unit obtain
To sparse features figure, convolution can accelerate sort operation using sparse Fourier transform;
F n*q group response diagrams greatly) are worth to using non-to every o groups response diagram in n*o*q, then in n*q response diagrams
Every n response diagram perform non-maxima suppression obtain q groups have true yardstick response diagram, q groups response diagram perform and fortune
Calculate and obtain a final response diagram with true yardstick, calculate the barycenter in each UNICOM region in final response diagram:
G) barycenter and true yardstick are mapped in former mapping to be checked, drawn pair according to each centroid position and scale-value
The rectangle frame answered, complete target detection.
In step A) in, the preparation training sample set may include following sub-step:
A1 the image block to the frame that sets the goal) is extracted from training image, then zooms to fixed size size as positive sample,
Each gives the image block for the frame that sets the goal to obtain N as a sample0Individual image block, i.e. N0Individual sample Xi, i=1 ... ... N;
A2) for negative sample image acquisition, typically around positive sample image block extract with its without very greatly it is overlapping
, the image block of arbitrary size zoom to fixed dimension as negative sample Xi, the extraction of negative sample will have as far as possible to be represented
Property, most of scene images are covered, extract N altogether1Negative sample, i=1 ... ... N;
A3) N=N0+N1。
In step B) in, it is described to may include following sub-step using the training of m this progress of lot sample convolutional neural networks:
B1) N number of sample needed when in batches it is random break up order, be then divided into m lot sample sheets, purpose in batches is
The Grad needed for neural metwork training is calculated with small quantities of sample, order is broken up and is advantageous to obtain more reasonably gradient direction;
B2) well-designed convolutional neural networks include convolutional layer, maximum pond (max-pooling) layer and local contrast
Normalize (local contrast normalization) three composition of layer, the first two composition is required composition, the 3rd into
Point be that non-linear correction unit is selected according to any optional composition, the activation primitive of convolution, by be combined these three into
Divide the network structure that can be obtained for different target detection;
B3) filter number and feature map number, wave filter size, down-sampling core required in network structure is set
Hyper parameter required for size, every layer of learning rate and local contrast normalization;
B4) during convolutional neural networks are trained, using momentum and at random training skills is thrown away;
B5 deconditioning) is determined when according to checking curve;
B6) the extracting parameter W from the model trained.
In step C) in, it is described to solve maximum pond (max- in convolutional neural networks by the way of expander graphs
Pooling the problem of) losing information may include following sub-step:
C1) W extracted in step B is used for during the feedforward operation of image to be detected, met for each characteristic pattern
It is new as one for each skew in the size K of down-sampling core when losing the situation of discriminant information to max-pooling
Beginning offset, for the full down-sampling core that then tiles to whole image to be detected, each beginning offset obtains an expander graphs,
Current down-sampling layer can generate 2KIndividual expander graphs, the expander graphs that all characteristic patterns are obtained using same skew are referred to as one section
(fragment) 2, can be generated after down-samplingKThe characteristic pattern of section, if p layer down-sampling layers in whole network, o=can be obtained
(2K)pThe characteristic pattern of section;
C2) the follow-up convolution operation of down-sampling to every section, it is necessary to use identical convolution operation;
C3) local contrast normalization layer needs to use identical local contrast normalization operation to every section;
C4) during treating test image and going to feedforward operation, full articulamentum preceding layer obtain (2K)pSection
Characteristic pattern is referred to as differentiating complete characteristics figure.
In step D) in, described be directly based upon differentiates that complete characteristics figure detects (classification) and may include following sub-step:
D1) by (2 in step CK)pThe characteristic pattern of section all using arest neighbors interpolation algorithm n times, obtains n* (2K)pSection is sentenced
Other complete characteristics figure;
D2) differentiate that doing convolution with it with linear classifier on complete characteristics figure obtains n* (2 at every sectionK)pResponse diagram;
D3) n* (2K)pResponse diagram all zoom to same size and record pantograph ratio, with non-maxima suppression algorithm
Obtain the pantograph ratio of final response diagram and corresponding peak response figure.
The present invention trains convolutional neural networks parameter first with training set, is then solved most using the mode of expander graphs
Great Chiization (max-pooling), which is lost the problem of feature and generated, differentiates complete characteristics figure;The full connection of convolutional neural networks
Weight regards linear classifier as, estimates general China of the linear classifier on complete characteristics are differentiated using possible approximate learning framework
Error;Linear classifier number needed for being estimated according to general magnificent error and desired extensive error threshold is finally complete in differentiation
On characteristic pattern target detection is completed with mode of the linear classifier based on smoothing windows.
The present invention solves the problems, such as the Character losing in Max-pooling layers by the way of expander graphs, and expander graphs
Move to other layers;All expander graphs before full connection are referred to as differentiating complete characteristics figure, and the weight connected entirely is regarded as
One linear classifier, detection efficiency can be significantly improved by directly doing detection on complete characteristics figure is differentiated;Using possible approximate
Learn (probably approximately correct) framework to estimate that linear classifier is general on complete characteristics are differentiated
Change error R [ω], then the instruction according to required for calculating built-up pattern (ensemble model) and desired extensive error H [ω]
Practice model number q, final prediction error is reduced according to the prediction of q model, and then improve target detection precision.
Brief description of the drawings
Fig. 1 is the detection framework schematic diagram of the embodiment of the present invention.
Fig. 2 is the schematic diagram of the expander graphs of the embodiment of the present invention.
Fig. 3 is the detection example figure of the embodiment of the present invention.
Fig. 4 is the testing result figure of the embodiment of the present invention.Wherein left frame is the inventive method testing result, and left frame is
The spacious testing result for regarding the method that scientific and technological (Megvii) Co., Ltd proposes in Beijing.
Fig. 5 is the of the invention and ROC curve figure of contrast of other several object detection methods on FDDB data sets.
Wherein imaginary curve is (entitled:The method of the present invention) it is method of the invention;
Method 1 corresponds to method (B.Yang, J.Yan, the Z.Lei and of B.Yang et al. propositions
S.Z.Li.Aggregate channel features for multi-view face
detection..International Joint Conference on Biometrics,2014);
Method 2 corresponds to method (H.Li, Z.Lin, J.Brandt, the X.Shen and of H.Li et al. propositions
G.Hua.Efficient Boosted Exemplar-based Face Detection.CVPR 2014);
Method 3 corresponds to method (J.Yan, Z.Lei, L.Wen the and S.Z.Li.The of J.Yan et al. propositions
Fastest Deformable Part Model for Object Detection.CVPR 2014);
Method 4 corresponds to the spacious method for regarding scientific and technological (Megvii) Co., Ltd and proposing in Beijing;
Method 5 corresponds to method (M.Mathias, R.Benenson, the M.Pedersoli of M.Mathias et al. propositions
and L.Van Gool.Face detection without bells and whistles.ECCV 2014);
Method 6 corresponds to method (X.Shen, Z.Lin, the J.Brandt and of X.Shen et al. propositions
Y.Wu.Detecting and Aligning Faces by Image Retrieval.CVPR 2013);
Method 7 corresponds to method (the J.Li and Y.Zhang.Learning SURF cascade of J.Li et al. propositions
for fast and accurate object detection.CVPR 2013.);
Method 8 corresponds to method (J.Li, T.Wang and the Y.Zhang.Face Detection of J.Li et al. propositions
using SURF Cascade.ICCV 2011BeFIT workshop.);
Method 9 corresponds to method (the P.Viola and M.Jones.Robust real-time of Viola et al. propositions
object detection.In IJCV,2001);
Method 10 correspond to A.Giusti et al. proposition method (A.Giusti, D.C.Ciresan, J.Masci,
L.M.Gambardella,and J.Schmidhuber.Fast image scanning with deep max-pooling
convolutional neural networks.In ICIP,2013)。
Embodiment
The method of the present invention is elaborated with reference to the accompanying drawings and examples, the present embodiment is with the technology of the present invention side
Implemented under premised on case, give embodiment and specific operation process, but protection scope of the present invention be not limited to it is following
Embodiment.
Referring to Fig. 1, the embodiment of the present invention comprises the following steps:
A. training sample set (x is preparedi, yi), i=1 ... ..., N, N are number of training, and N is natural number.xiRepresent training
The image of fixed size corresponding to sample, comprising target and target, the image full of picture is positive sample, and other images are negative samples
This.yiRepresent sample class vector:
B. divide m batches by all training samples, originally put m-2 lot samples into well-designed convolutional neural networks reversely biography
Algorithm for Training is broadcast, 2 lot sample sheets are used to test.Well-designed convolutional neural networks include convolutional layer, maximum pond (max-
Pooling) layer and local contrast normalization (local contrast normalization) three composition of layer.These three compositions
The function of performer's nonlinear function, it is therefore an objective to which the original image in manifold is mapped to theorem in Euclid space.The activation of convolution
Function uses non-linear correction unit (Rectifier linear unit), can make it that the aspect ratio after convolution is sparse.If
After counting the network structure model comprising these three compositions, model is trained with back-propagation algorithm training, obtains parameter
W。
C. the parameter W trained in extraction step B, solves max- in convolutional neural networks by the way of expander graphs
Pooling loses the problem of information.Given test image, conventional maximum pond (max-pooling) is (namely using skew A
Mode as starting point) operation after can only obtain expander graphs (A), other be advantageous to classification information can lose.Therefore it is directed to
Each offset in 2*2 convolution kernel, all preserve a corresponding characteristic pattern, referred to as expander graphs, as shown in Fig. 2 (A), (B),
(C), (D) is different expander graphs corresponding to different skews.
It is K for each layer of maximum pond (max-pooling) core size, whole convolutional neural networks have p layers, extend it
Each max-pooling down-sampling layer feature map number extends to 2 afterwardsKTimes, whole network expands to o=(2K)pTimes, then
Parameter W is used for image to be detected of arbitrary size, obtains the extension feature figure before full connection, it is referred to as to differentiate complete characteristics
(Discriminative Complete Features), as shown in figure 3, (a) and (c) is referred to as differentiating complete characteristics in Fig. 3, its
In (c) (a) obtained by arest neighbors difference arithmetic.
Given input picture x, the wave filter group K trained and skew b, the output of convolutional layer can be write as (formula one)
Form:
(formula one)
Wherein MjThe index of input feature vector figure selected by expression;L represents the index of current layer;I and j represent respectively input and
Output characteristic index of the picture;O represents segment index;F represent activation primitive, used here as linearity correction unit function f (x)=max (x,
0);* convolution operation is represented.
Maximum pond (max-pooling) layer can be write as the form of (formula two):
(formula two)
Wherein, m and n represents the pixel index of current layer respectively;S represents the size of down-sampling core;Before p and q is represented respectively
One layer of pixel index, and meet p=s* (m-1)+k+1, q=s* (n-1)+k+1, wherein 0 < k < s;Branch is used to select
Corresponding columns and rows.
In order that the feature that must be obtained has the property of Shandong nation, local contrast can be normalized layer and introduce segment processing
In link, the layer can be write as the form of (formula three):
(formula three)
Wherein, the number for the neighbouring competition characteristic pattern that r expressions are formulated, the feature map number of N expressions current layer altogether, k, α,
β is some hyper parameters, can specify suitable floating-point values in the training process.After CNN model trainings are good, 3 more than
Individual formula, it can obtain and differentiate that complete characteristics provide support to be follow-up based on smoothing windows scanning.
D. the full connection of convolutional neural networks is regarded as a linear classifier, be directly based upon differentiate complete characteristics figure on and
It is not that former detection figure does detection and can greatly improve detection speed.Using possible approximate study (probably
Approximately correct) framework come estimate linear classifier differentiate complete characteristics on extensive error R [ω], so
The training pattern number q according to required for calculating built-up pattern (ensemble model) and desired extensive error H [ω] afterwards.
(formula four)
E. each training pattern will be obtained in q model and performs feedforward operation in image to be detected respectively, each mould
Type obtains the differentiation complete characteristics figure before o groups connect entirely, and then every group of differentiation complete characteristics figure is contracted with arest neighbors interpolation algorithm
Put, obtain n*o groups and differentiate complete characteristics figure, q model amounts to n*o*q groups and differentiate complete characteristics figure.Then it is complete in every group of differentiation
Directly intensive smoothing windows classification work, which is performed, on standby characteristic pattern with linear classifier obtains n*o*q group response diagrams.Wherein linear point
The dot product operations of class device and differentiation complete graph can change into convolution operation, due to being obtained using the non-linear correction unit in step B
To sparse features figure, convolution can accelerate sort operation using sparse Fourier transform.As shown in figure 3, (b) in Fig. 3 and
(d) it is respectively that linear classifier obtains in differentiation complete characteristics (a) and (c) convolution, is obtained wherein (e) is (d) scaling.
F. n*q group response diagrams greatly are worth to using non-to every o groups response diagram in n*o*q, then in n*q response diagrams
Every n response diagram perform non-maxima suppression obtain q groups have true yardstick response diagram, q groups response diagram perform and fortune
Calculate and obtain a final response diagram with true yardstick, calculate the barycenter in each UNICOM region in final response diagram.
G. barycenter and true yardstick are mapped in former mapping to be checked, drawn pair according to each centroid position and scale-value
The rectangle frame answered, target detection is completed, Fig. 4 is testing result.
The present invention and other based on convolutional neural networks the time required to target detection on comparative result be shown in Table 1.
Table 1
Method | CPU time (s) | GPU time (s) | Total time (s) |
Method 11 | 2.3 | 25.08 | 28.1 |
Method 12 | 43.2 | 0 | 43.2 |
Method 13 | 2.3 | 0.25 | 2.55 |
The method of the present invention | 1.3 | 0 | 1.3 |
In table 1, method 11 be Fabian et al. propose method (Fabian Nasse, Christian Thurau,
and Gernot A.Fink,“Face detection using gpu-based convolutional neural
networks,”in CAIP,2009,pp.83–90);
Method 12 be A.Giusti et al. propose method (A.Giusti, D.C.Ciresan, J.Masci,
L.M.Gambardella,and J.Schmidhuber.Fast image scanning with deep max-pooling
convolutional neural networks.In ICIP,2013);
Method 13 is method (K.He, X.Zhang, S.Ren, the and J.Sun.Spatial that K.He et al. is proposed
pyramid pooling in deep convolutional networks for visual recognition.In
ECCV,2014)。
The present invention directly does and classified that (multiple dimensioned differentiation complete characteristics figure is by most on multiple dimensioned differentiation complete characteristics figure
Neighbour's interpolation algorithm scales to obtain), the weight vectors that linear classifier is changed into change into nuclear matrix, and linear classification can pass through core
Matrix and multiple dimensioned differentiation complete characteristics picture scroll product are completed.Due to differentiating that complete characteristics are sparse, convolution speed can pass through
Sparse Fourier transform obtains.Further, since directly detected on complete characteristics figure is differentiated, rather than in artwork, target inspection
Degree of testing the speed greatly improves.
Claims (5)
1. the fast target detection method based on convolutional neural networks, it is characterised in that comprise the following steps:
A training sample set (x) is preparedi, yi), i=1 ... ..., N, N is number of training, and N is natural number, xiRepresent training sample
The image of corresponding fixed size, comprising target and target, the image full of picture is positive sample, and other images are negative samples;yi
Represent sample class vector:
B) m batches are divided by all training samples, the convolutional neural networks that m-2 lot samples are originally put into design are instructed with back-propagation algorithm
Practice, 2 lot sample sheets are used to test, and the convolutional neural networks include convolutional layer, maximum pond layer and local contrast normalization layer three
Individual composition, these three compositions play the part of the function of nonlinear function, it is therefore an objective to the original image in manifold are mapped to European
Space, the activation primitive of convolution uses non-linear correction unit, after designing the network structure model comprising these three compositions, uses
Back-propagation algorithm training is trained to model, obtains parameter W;
C) extraction step B) in the parameter W trained, solve maximum pond in convolutional neural networks by the way of expander graphs
The problem of losing information, test image is given, for each being offset in 2 × 2 convolution kernel, all preserve a corresponding feature
Figure, referred to as expander graphs;It is K for each layer of max-pooling cores size, whole convolutional neural networks have p layers, every after extension
The down-sampling layer feature map number in individual maximum pond extends to 2KTimes, whole network expands to o=(2K)pTimes, then parameter W
For image to be detected of arbitrary size, the extension feature figure before full connection is obtained, it is referred to as to differentiate complete characteristics;Given input
Image x, the wave filter group K trained and skew b, the output of convolutional layer are write as the form of formula one:
Wherein MjThe index of input feature vector figure selected by expression;L represents the index of current layer;I and j represents to input and export respectively
Feature index of the picture;O represents segment index;F represents activation primitive, used here as linearity correction unit function f (x)=max (x, 0);*
Represent convolution operation;
Maximum pond layer is write as the form of formula two:
Wherein, m and n represents the pixel index of current layer respectively;S represents the size of down-sampling core;P and q represent preceding layer respectively
Pixel index, and meet p=s* (m-1)+k+1, q=s* (n-1)+k+1, wherein 0 < k < s;Branch is used to select correspondingly
Columns and rows;
In order that the feature that must be obtained has the property of Shandong nation, local contrast normalization layer is introduced into the link of segment processing,
This layer is write as the form of formula three:
Wherein, r represents the number for the neighbouring competition characteristic pattern formulated, and N represents the feature map number of current layer altogether, k, α, and β is
Some hyper parameters, suitable floating-point values are specified in the training process, after CNN model trainings are good, according to three above formula,
Obtain and differentiate that complete characteristics provide support to be follow-up based on smoothing windows scanning;
D the full connection of convolutional neural networks) is regarded as a linear classifier, is directly based upon and differentiates that complete characteristics figure detects,
Extensive error R [ω] of the linear classifier on complete characteristics are differentiated is estimated using possible approximate learning framework, then basis
Training pattern number q required for built-up pattern and desired extensive error H [ω] calculate;
E) each training pattern will be obtained in q model and performs feedforward operation in image to be detected respectively, each model obtains
The differentiation complete characteristics figure before the full connection of o groups is taken, then every group of differentiation complete characteristics figure is scaled with arest neighbors interpolation algorithm, obtained
Differentiate complete characteristics figure to n*o groups, q model amounts to n*o*q groups and differentiate complete characteristics figure, then in every group of differentiation complete characteristics
Directly perform the classification work of intensive smoothing windows with linear classifier on figure and obtain n*o*q group response diagrams, wherein linear classifier with
Differentiate that the dot product operations of complete graph change into convolution operation;
F n*q group response diagrams greatly) are worth to using non-to every o groups response diagram in n*o*q, then to every in n*q response diagrams
N response diagram performs non-maxima suppression and obtains response diagram of the q groups with true yardstick, and q groups response diagram is performed and obtained with computing
The final response diagram that there is true yardstick to one, calculate the barycenter in each UNICOM region in final response diagram:
G) barycenter and true yardstick are mapped in former mapping to be checked, according to corresponding to being drawn each centroid position and scale-value
Rectangle frame, complete target detection.
2. the fast target detection method based on convolutional neural networks as claimed in claim 1, it is characterised in that in step A) in,
The preparation training sample set includes following sub-step:
A1 the image block to the frame that sets the goal) is extracted from training image, then zooms to fixed size size as positive sample, it is each
The individual image block to the frame that sets the goal obtains N as a sample0Individual image block, i.e. N0Individual sample Xi, i=1 ... ... N0;
A2) for negative sample image acquisition, around positive sample image block extract with its without very greatly it is overlapping, arbitrarily large
Small image block zooms to fixed dimension as negative sample Xi’;The extraction of negative sample is near possible representative, to cover
Most of scene images;N is extracted altogether1Negative sample, i=1 ... ... N1;
A3) N=N0+N1。
3. the fast target detection method based on convolutional neural networks as claimed in claim 1, it is characterised in that in step B) in,
The utilizationThe training of this progress of lot sample convolutional neural networks includes following sub-step:
B1) N number of sample needed when in batches it is random break up order, be then divided intoLot sample sheet;
B2) convolutional neural networks include convolutional layer, maximum pond layer and local contrast normalization three composition of layer, the first two composition
It is required composition, the 3rd composition is to select non-linear correction unit according to any optional composition, the activation primitive of convolution, is led to
Cross and be combined the network structure that these three compositions obtain being directed to different target detection;
B3) set network structure in required for filter number and feature map number, wave filter size, down-sampling core it is big
Hyper parameter required for small, every layer of learning rate and local contrast normalization;
B4) during convolutional neural networks are trained, using momentum and at random training skills is thrown away;
B5 deconditioning) is determined when according to checking curve;
B6) the extracting parameter W from the model trained.
4. the fast target detection method based on convolutional neural networks as claimed in claim 1, it is characterised in that in step C) in,
It is described to solve the problems, such as that maximum pondization loss information includes following sub-step in convolutional neural networks by the way of expander graphs:
C1) step B) in extraction W be used for image to be detected feedforward operation during, run into for each characteristic pattern
When max-pooling loses the situation of discriminant information, new as one for each skew in the size K of down-sampling core
Beginning offset, for the full down-sampling core that then tiles to whole image to be detected, each beginning offset obtains an expander graphs, when
Front lower sample level can generate 2KIndividual expander graphs, all characteristic patterns are referred to as one section using the same obtained expander graphs of offseting, under adopt
2 can be generated after sampleKThe characteristic pattern of section, if there is z layer down-sampling layers in whole network, obtains o=(2K)zThe characteristic pattern of section;
C2) the follow-up convolution operation of down-sampling to every section, it is necessary to use identical convolution operation;
C3) local contrast normalization layer needs to use identical local contrast normalization operation to every section;
C4) during treating test image and going to feedforward operation, full articulamentum preceding layer obtain (2K)zThe feature of section
Figure is referred to as differentiating complete characteristics figure.
5. the fast target detection method based on convolutional neural networks as claimed in claim 1, it is characterised in that in step D) in,
Described be directly based upon differentiates that complete characteristics figure does detection and includes following sub-step:
D1) by step C) in (2K)zThe characteristic pattern of section all using arest neighbors interpolation algorithm n times, obtains n* (2K)zThe differentiation of section is complete
Standby characteristic pattern;
D2) differentiate that doing convolution with it with linear classifier on complete characteristics figure obtains n* (2 at every sectionK)zResponse diagram;
D3) n* (2K)zResponse diagram all zoom to same size and record pantograph ratio, obtained with non-maxima suppression algorithm
The pantograph ratio of final response diagram and corresponding peak response figure.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510061852.6A CN104573731B (en) | 2015-02-06 | 2015-02-06 | Fast target detection method based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510061852.6A CN104573731B (en) | 2015-02-06 | 2015-02-06 | Fast target detection method based on convolutional neural networks |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104573731A CN104573731A (en) | 2015-04-29 |
CN104573731B true CN104573731B (en) | 2018-03-23 |
Family
ID=53089751
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510061852.6A Expired - Fee Related CN104573731B (en) | 2015-02-06 | 2015-02-06 | Fast target detection method based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104573731B (en) |
Families Citing this family (62)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9998130B2 (en) * | 2016-07-06 | 2018-06-12 | Hrl Laboratories, Llc | Method to perform convolutions between arbitrary vectors using clusters of weakly coupled oscillators |
CN104992223B (en) * | 2015-06-12 | 2018-02-16 | 安徽大学 | Intensive Population size estimation method based on deep learning |
US10614339B2 (en) | 2015-07-29 | 2020-04-07 | Nokia Technologies Oy | Object detection with neural network |
US10970617B2 (en) | 2015-08-21 | 2021-04-06 | Institute Of Automation Chinese Academy Of Sciences | Deep convolutional neural network acceleration and compression method based on parameter quantification |
CN105160310A (en) * | 2015-08-25 | 2015-12-16 | 西安电子科技大学 | 3D (three-dimensional) convolutional neural network based human body behavior recognition method |
US10332028B2 (en) * | 2015-08-25 | 2019-06-25 | Qualcomm Incorporated | Method for improving performance of a trained machine learning model |
CN105205453B (en) * | 2015-08-28 | 2019-01-08 | 中国科学院自动化研究所 | Human eye detection and localization method based on depth self-encoding encoder |
CN105120130B (en) * | 2015-09-17 | 2018-06-29 | 京东方科技集团股份有限公司 | A kind of image raising frequency system, its training method and image raising frequency method |
CN105184271A (en) * | 2015-09-18 | 2015-12-23 | 苏州派瑞雷尔智能科技有限公司 | Automatic vehicle detection method based on deep learning |
US10614354B2 (en) * | 2015-10-07 | 2020-04-07 | Altera Corporation | Method and apparatus for implementing layers on a convolutional neural network accelerator |
CN105335716B (en) * | 2015-10-29 | 2019-03-26 | 北京工业大学 | A kind of pedestrian detection method extracting union feature based on improvement UDN |
US9965705B2 (en) * | 2015-11-03 | 2018-05-08 | Baidu Usa Llc | Systems and methods for attention-based configurable convolutional neural networks (ABC-CNN) for visual question answering |
CN105279556B (en) * | 2015-11-05 | 2017-11-07 | 国家卫星海洋应用中心 | A kind of Enteromorpha detection method and device |
CN105426919B (en) * | 2015-11-23 | 2017-11-14 | 河海大学 | The image classification method of non-supervisory feature learning is instructed based on conspicuousness |
CN105468335B (en) * | 2015-11-24 | 2017-04-12 | 中国科学院计算技术研究所 | Pipeline-level operation device, data processing method and network-on-chip chip |
CN106778604B (en) * | 2015-12-15 | 2020-04-14 | 西安电子科技大学 | Pedestrian re-identification method based on matching convolutional neural network |
US10360477B2 (en) * | 2016-01-11 | 2019-07-23 | Kla-Tencor Corp. | Accelerating semiconductor-related computations using learning based models |
CN105718890A (en) * | 2016-01-22 | 2016-06-29 | 北京大学 | Method for detecting specific videos based on convolution neural network |
CN105740892A (en) * | 2016-01-27 | 2016-07-06 | 北京工业大学 | High-accuracy human body multi-position identification method based on convolutional neural network |
CN108475331B (en) * | 2016-02-17 | 2022-04-05 | 英特尔公司 | Method, apparatus, system and computer readable medium for object detection |
CN105821538B (en) * | 2016-04-20 | 2018-07-17 | 广州视源电子科技股份有限公司 | The detection method and system of spun yarn fracture |
CN111860772B (en) * | 2016-04-29 | 2024-01-16 | 中科寒武纪科技股份有限公司 | Device and method for executing artificial neural network mapping operation |
CN106019359A (en) * | 2016-05-17 | 2016-10-12 | 浪潮集团有限公司 | Earthquake prediction system based on neural network |
WO2017200597A1 (en) * | 2016-05-20 | 2017-11-23 | Google Llc | Progressive neural networks |
CN109154990B (en) * | 2016-06-03 | 2023-10-03 | 英特尔公司 | Finding convolutional layers in convolutional neural networks |
CN106021990B (en) * | 2016-06-07 | 2019-06-25 | 广州麦仑信息科技有限公司 | A method of biological gene is subjected to classification and Urine scent with specific character |
CN106096655B (en) * | 2016-06-14 | 2019-08-27 | 厦门大学 | A kind of remote sensing image airplane detection method based on convolutional neural networks |
CN106203496B (en) * | 2016-07-01 | 2019-07-12 | 河海大学 | Hydrographic curve extracting method based on machine learning |
CN106504233B (en) * | 2016-10-18 | 2019-04-09 | 国网山东省电力公司电力科学研究院 | Unmanned plane inspection image electric power widget recognition methods and system based on Faster R-CNN |
CN106778835B (en) * | 2016-11-29 | 2020-03-24 | 武汉大学 | Remote sensing image airport target identification method fusing scene information and depth features |
CN106780512B (en) * | 2016-11-30 | 2020-01-17 | 厦门美图之家科技有限公司 | Method, application and computing device for segmenting image |
CN106845528A (en) * | 2016-12-30 | 2017-06-13 | 湖北工业大学 | A kind of image classification algorithms based on K means Yu deep learning |
CN107038448B (en) * | 2017-03-01 | 2020-02-28 | 中科视语(北京)科技有限公司 | Target detection model construction method |
CN108229675B (en) * | 2017-03-17 | 2021-01-01 | 北京市商汤科技开发有限公司 | Neural network training method, object detection method, device and electronic equipment |
CN108629354B (en) * | 2017-03-17 | 2020-08-04 | 杭州海康威视数字技术股份有限公司 | Target detection method and device |
CN107124609A (en) * | 2017-04-27 | 2017-09-01 | 京东方科技集团股份有限公司 | A kind of processing system of video image, its processing method and display device |
CN107220652B (en) * | 2017-05-31 | 2020-05-01 | 北京京东尚科信息技术有限公司 | Method and device for processing pictures |
CN107527355B (en) * | 2017-07-20 | 2020-08-11 | 中国科学院自动化研究所 | Visual tracking method and device based on convolutional neural network regression model |
CN109325385A (en) * | 2017-07-31 | 2019-02-12 | 株式会社理光 | Target detection and region segmentation method, device and computer readable storage medium |
CN107563303B (en) * | 2017-08-09 | 2020-06-09 | 中国科学院大学 | Robust ship target detection method based on deep learning |
CN107292886B (en) * | 2017-08-11 | 2019-12-31 | 厦门市美亚柏科信息股份有限公司 | Target object intrusion detection method and device based on grid division and neural network |
CN107506774A (en) * | 2017-10-09 | 2017-12-22 | 深圳市唯特视科技有限公司 | A kind of segmentation layered perception neural networks method based on local attention mask |
CN107766643B (en) * | 2017-10-16 | 2021-08-03 | 华为技术有限公司 | Data processing method and related device |
CN107944354B (en) * | 2017-11-10 | 2021-09-17 | 南京航空航天大学 | Vehicle detection method based on deep learning |
WO2019099899A1 (en) * | 2017-11-17 | 2019-05-23 | Facebook, Inc. | Analyzing spatially-sparse data based on submanifold sparse convolutional neural networks |
CN108171796A (en) * | 2017-12-25 | 2018-06-15 | 燕山大学 | A kind of inspection machine human visual system and control method based on three-dimensional point cloud |
CN108280453B (en) * | 2018-01-08 | 2020-06-16 | 西安电子科技大学 | Low-power-consumption rapid image target detection method based on deep learning |
CN110390344B (en) * | 2018-04-19 | 2021-10-26 | 华为技术有限公司 | Alternative frame updating method and device |
CN108830280B (en) * | 2018-05-14 | 2021-10-26 | 华南理工大学 | Small target detection method based on regional nomination |
CN108830300A (en) * | 2018-05-28 | 2018-11-16 | 深圳市唯特视科技有限公司 | A kind of object transmission method based on mixing supervisory detection |
CN108875819B (en) * | 2018-06-08 | 2020-10-27 | 浙江大学 | Object and component joint detection method based on long-term and short-term memory network |
CN109189965A (en) * | 2018-07-19 | 2019-01-11 | 中国科学院信息工程研究所 | Pictograph search method and system |
CN109466725B (en) * | 2018-10-11 | 2021-05-18 | 重庆邮电大学 | Intelligent water surface floater fishing system based on neural network and image recognition |
CN109376787B (en) * | 2018-10-31 | 2021-02-26 | 聚时科技(上海)有限公司 | Manifold learning network and computer vision image set classification method based on manifold learning network |
CN109753903B (en) * | 2019-02-27 | 2020-09-15 | 北航(四川)西部国际创新港科技有限公司 | Unmanned aerial vehicle detection method based on deep learning |
CN110135312B (en) * | 2019-05-06 | 2022-05-03 | 电子科技大学 | Rapid small target detection method based on hierarchical LCM |
CN110390394B (en) * | 2019-07-19 | 2021-11-05 | 深圳市商汤科技有限公司 | Batch normalization data processing method and device, electronic equipment and storage medium |
CN110674829B (en) * | 2019-09-26 | 2023-06-02 | 哈尔滨工程大学 | Three-dimensional target detection method based on graph convolution attention network |
CN111612051B (en) * | 2020-04-30 | 2023-06-20 | 杭州电子科技大学 | Weak supervision target detection method based on graph convolution neural network |
CN112862195B (en) * | 2021-02-19 | 2023-06-20 | 金陵科技学院 | SFT-ALS-based time series vermicelli fluctuation prediction method |
WO2023220892A1 (en) * | 2022-05-16 | 2023-11-23 | Intel Corporation | Expanded neural network training layers for convolution |
CN114657513B (en) * | 2022-05-23 | 2022-09-20 | 河南银金达新材料股份有限公司 | Preparation method of antibacterial regenerated polyester film |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103810503A (en) * | 2013-12-26 | 2014-05-21 | 西北工业大学 | Depth study based method for detecting salient regions in natural image |
CN104063719A (en) * | 2014-06-27 | 2014-09-24 | 深圳市赛为智能股份有限公司 | Method and device for pedestrian detection based on depth convolutional network |
CN104281853A (en) * | 2014-09-02 | 2015-01-14 | 电子科技大学 | Behavior identification method based on 3D convolution neural network |
CN104680508A (en) * | 2013-11-29 | 2015-06-03 | 华为技术有限公司 | Convolutional neural network and target object detection method based on convolutional neural network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7274832B2 (en) * | 2003-11-13 | 2007-09-25 | Eastman Kodak Company | In-plane rotation invariant object detection in digitized images |
-
2015
- 2015-02-06 CN CN201510061852.6A patent/CN104573731B/en not_active Expired - Fee Related
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104680508A (en) * | 2013-11-29 | 2015-06-03 | 华为技术有限公司 | Convolutional neural network and target object detection method based on convolutional neural network |
CN103810503A (en) * | 2013-12-26 | 2014-05-21 | 西北工业大学 | Depth study based method for detecting salient regions in natural image |
CN104063719A (en) * | 2014-06-27 | 2014-09-24 | 深圳市赛为智能股份有限公司 | Method and device for pedestrian detection based on depth convolutional network |
CN104281853A (en) * | 2014-09-02 | 2015-01-14 | 电子科技大学 | Behavior identification method based on 3D convolution neural network |
Also Published As
Publication number | Publication date |
---|---|
CN104573731A (en) | 2015-04-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104573731B (en) | Fast target detection method based on convolutional neural networks | |
CN109543606B (en) | Human face recognition method with attention mechanism | |
CN107844795B (en) | Convolutional neural networks feature extracting method based on principal component analysis | |
CN109034210A (en) | Object detection method based on super Fusion Features Yu multi-Scale Pyramid network | |
CN104392463B (en) | Image salient region detection method based on joint sparse multi-scale fusion | |
CN111860171B (en) | Method and system for detecting irregular-shaped target in large-scale remote sensing image | |
CN108171112A (en) | Vehicle identification and tracking based on convolutional neural networks | |
CN110807422A (en) | Natural scene text detection method based on deep learning | |
CN109446922B (en) | Real-time robust face detection method | |
CN109741318A (en) | The real-time detection method of single phase multiple dimensioned specific objective based on effective receptive field | |
CN110827312B (en) | Learning method based on cooperative visual attention neural network | |
CN107944459A (en) | A kind of RGB D object identification methods | |
CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
CN106778768A (en) | Image scene classification method based on multi-feature fusion | |
CN110781962B (en) | Target detection method based on lightweight convolutional neural network | |
CN108520203A (en) | Multiple target feature extracting method based on fusion adaptive more external surrounding frames and cross pond feature | |
CN109948457B (en) | Real-time target recognition method based on convolutional neural network and CUDA acceleration | |
CN105046278B (en) | The optimization method of Adaboost detection algorithm based on Haar feature | |
CN112163498A (en) | Foreground guiding and texture focusing pedestrian re-identification model establishing method and application thereof | |
CN112528845A (en) | Physical circuit diagram identification method based on deep learning and application thereof | |
CN106529441A (en) | Fuzzy boundary fragmentation-based depth motion map human body action recognition method | |
CN110969101A (en) | Face detection and tracking method based on HOG and feature descriptor | |
CN114492634A (en) | Fine-grained equipment image classification and identification method and system | |
CN112102379B (en) | Unmanned aerial vehicle multispectral image registration method | |
CN111160372B (en) | Large target identification method based on high-speed convolutional neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180323 |