CN107092960A - A kind of improved parallel channel convolutional neural networks training method - Google Patents
A kind of improved parallel channel convolutional neural networks training method Download PDFInfo
- Publication number
- CN107092960A CN107092960A CN201710247556.4A CN201710247556A CN107092960A CN 107092960 A CN107092960 A CN 107092960A CN 201710247556 A CN201710247556 A CN 201710247556A CN 107092960 A CN107092960 A CN 107092960A
- Authority
- CN
- China
- Prior art keywords
- eigenmatrix
- layer
- neural networks
- convolutional neural
- convolutional
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
A kind of improved parallel channel convolutional neural networks training method.It carries out feature extraction using direct-connected and convolutional channel to data in convolutional neural networks, obtains eigenmatrix;Eigenmatrix is merged, and carries out Data Dimensionality Reduction;Convolutional neural networks are trained, the penalty values of this network training are calculated;Calculate each layer error term and weights gradient;Judge whether network restrains according to penalty values, if do not restrained, be trained according to weights gradient adjustment convolutional neural networks initiation parameter and again, the steps such as network training result are exported if having restrained.The present invention can ensure the circulation of data in a network by the introducing of direct channel, gradient unstable problem when overcoming the training of deep layer convolutional neural networks, can train the network of deeper;Using maximum pondization and average pond, it can be consistent the eigenmatrix dimension between feature extraction twice and the advantage of two kinds of pond methods can be combined.
Description
Technical field
The invention belongs to deep learning and big data technical field, and in particular to a kind of improved parallel channel convolution god
Through network training method.
Background technology
With the development of society, the arriving in big data epoch, associated technology is continued to develop and innovated.Deep learning
Because it can improve classification accuracy rate using mass data and by the training of deeper layer network, some row are obtained in recent years and are broken through
Property progress.Scholars attempt to lift its performance by increasing the scale of convolutional neural networks, and it is most simple to increase network size
Single mode is exactly to increase " depth ".
But the depth network that the structure based on traditional convolutional neural networks is built, with the increase of the network number of plies, precision
Saturation, or even reduction can be reached.Document " Romero A, Ballas N, Kahou S E, et al.Fitnets:Hints for
thin deep nets[J].arXiv preprint arXiv:1412.6550,2014. " a kind of middle multistage training side of proposition
Method, is first respectively trained multiple shallow-layer networks, is finally combined multiple shallow-layer networks, so as to realize a deep layer network.This
Sample, which is done, needs that artificially multiple network parameters are adjusted respectively, takes time and effort, and multiple shallow-layer networks are respectively trained can lose
The related information between network is lost, influence will be produced on the last performance of network.Document " Lee C Y, Xie S, Gallagher
P,et al.Deeply-Supervised Nets[C]//Proceedings of the Eighteenth
International Conference on Artificial Intelligence and Statistics.2015:562-
570 " then introduce multiple subsidiary classification devices in the hidden layer of depth convolutional neural networks, although this method can be to a certain extent
Gradient disappearance problem during deep layer network error reverse conduction is compensated, but the subsidiary classification device introduced also can be last to network
Precision produces influence.
The problem of network of more depth can not be trained never fundamentally solves, and the network structure proposed is still
Designed based on traditional convolutional neural networks, simply used various Optimization Skills in the training process, such as:At the beginning of more preferable network
Beginningization parameter, more efficient excitation function etc..
The content of the invention
In order to solve the above problems, instructed it is an object of the invention to provide a kind of improved parallel channel convolutional neural networks
Practice method.
In order to achieve the above object, the improved parallel channel convolutional neural networks training method that the present invention is provided includes pressing
The following steps that order is carried out:
1) direct-connected and two parallel channels of convolution are utilized respectively feature extraction is carried out to the data in convolutional neural networks, obtained
To direct channel eigenmatrix and convolutional channel eigenmatrix;
2) by step 1) obtained two eigenmatrixes merge, and it is input to maximum pond layer and average pond layer enters
Row Data Dimensionality Reduction;
3) repeat step 1), step 2), obtain final eigenmatrix;
4) by above-mentioned steps 3) obtained final eigenmatrix carries out global average pond and inputs full articulamentum being changed into one
Dimensional feature matrix, and one-dimensional characteristic matrix is classified using softmax graders and convolutional neural networks are trained,
Calculate the penalty values of this network training;
5) gradient calculation is carried out using error backpropagation algorithm, calculates each layer error term and weights gradient;
6) according to step 4) in gained penalty values judge whether network restrains, if do not restrained, according to step 5) middle acquisition
Weights gradient adjusts convolutional neural networks initiation parameter and is trained again, and network training result is exported if having restrained.
In step 1) in, described is utilized respectively direct-connected and two parallel channels of convolution to the data in convolutional neural networks
Feature extraction is carried out, obtaining the method for direct channel eigenmatrix and convolutional channel eigenmatrix is:First, data are distinguished defeated
Enter direct channel and convolutional channel;Then data are directly mapped as to direct channel eigenmatrix as defeated in direct channel
Go out, convolution operation is carried out to data using multiple convolutional layers on convolutional channel, the input of each convolutional layer is a upper convolution
Layer output, using last convolutional layer output matrix as convolutional channel eigenmatrix.
In step 2) in, it is described by step 1) obtained two eigenmatrixes are merged, and it is input to maximum pond
The method of layer and average pond layer progress Data Dimensionality Reduction is:First, by obtained by eigenmatrix obtained by direct channel and convolutional channel
Eigenmatrix is merged, that is, obtains the set of multiple eigenmatrixes;Then gained eigenmatrix is inputted into maximum pond respectively
Layer and average pond layer, in maximum pond layer, the maximum being worth in wave filter is taken using wave filter, and in average pond, layer uses filter
Ripple device takes the average value in wave filter.
In step 4) in, it is described by above-mentioned steps 3) obtained final eigenmatrix carries out global average pond and inputs
Full articulamentum is changed into one-dimensional characteristic matrix, and one-dimensional characteristic matrix is classified and refreshing to convolution using softmax graders
It is trained through network, calculating the method for the penalty values of this network training is:First, final eigenmatrix is carried out global flat
Equal pond, uses the average value that data in eigenmatrix are calculated with final eigenmatrix wave filter of the same size;Then, input
Each neuron carries out non-linear to the data in the eigenmatrix behind global average pond respectively in full articulamentum, full articulamentum
Conversion obtains one-dimensional characteristic matrix;Finally, one-dimensional characteristic Input matrix softmax graders are classified.
In step 5) in, described carries out gradient calculation using error backpropagation algorithm, calculates each layer error term and power
The method of value gradient is:First, last layer of penalty values are calculated and as the mistake of last layer according to softmax classifier results
Poor item;Then, each layer error term, the ith feature of l-th of convolutional layer are calculated using the chain rule of error backpropagation algorithm
The error term of matrix is:
M is l+1 layers of eigenmatrix number,The error term of j-th of eigenmatrix of the l+1 convolutional layer is represented,Represent to this layer of activation primitive derivation, J represents this layer of penalty values,Represent j-th of eigenmatrix of l layers to l+1
The connection weight of layer ith feature matrix;
Finally, formula is utilizedEach layer weights gradient is calculated, whereinRepresent l-1 layers i-th
Eigenmatrix.
In step 6) in, it is described according to step 4) in gained penalty values judge whether network restrains, if do not restrained, foundation
Step 5) in the weights gradient adjustment that obtains convolutional neural networks initiation parameter and be trained again, as restrained, export
The method of network training result is:First, classification results and actual value are compared into simultaneously calculating difference and is used as penalty values;Then will
Penalty values and the classification thresholds of realization setting are compared, and such as less than classification thresholds then judge network convergence, otherwise do not restrain;Most
Afterwards, such as convergence output web results, otherwise according to formulaW(t+1):=W (t)+V
(t+1) convolutional neural networks initiation parameter is adjusted, wherein t represents iterations, and V (t) is momentum term, and μ is factor of momentum, it
Determine contribution of the history weight correction to this weight amendment;η is learning rate;λ is weight attenuation coefficient, and W is represented
Convolutional neural networks initiation parameter.
The advantage for the improved parallel channel convolutional neural networks training method that the present invention is provided is:1) by direct-connected logical
The introducing in road can ensure the circulation of data in a network, gradient unstable difficulty when overcoming the training of deep layer convolutional neural networks
Topic, can train the network of deeper;2) it using maximum pondization and average pond, can tie up the eigenmatrix between feature extraction twice
Degree is consistent and can combine the advantage of two kinds of pond methods.
Brief description of the drawings
The improved parallel channel convolutional neural networks training method flow chart that Fig. 1 provides for the present invention.
Fig. 2 is characterized the parallel channel structure chart for extracting part;
Fig. 3 is double ponds layer schematic diagram;
Fig. 4 is that parallel channel convolutional neural networks error term calculates schematic diagram;
Fig. 5 is the performance comparision of the different ponds mode on CIFAR-10 data sets;
Fig. 6 trains accuracy with the change curve of iterations for the present invention on CIFAR-10 data sets.
Embodiment
The improved parallel channel convolutional neural networks that the present invention is provided are trained with specific embodiment below in conjunction with the accompanying drawings
Method is described in detail.
As shown in figure 1, the improved parallel channel convolutional neural networks training method that the present invention is provided includes entering in order
Capable the following steps:
1) direct-connected and two parallel channels of convolution are utilized respectively feature extraction is carried out to the data in convolutional neural networks, obtained
To direct channel eigenmatrix and convolutional channel eigenmatrix;
It will be magnified the data set that the small coloured image for being 32 × 32 constitutes by tens thousand of and be input in convolutional neural networks, this hair
It is bright to use the CIFAR-10 data sets being made up of the 60000 big small coloured image for 32 × 32, then it is utilized respectively direct-connected and volume
The data that two parallel channels of product are concentrated to data carry out feature extraction, and parallel channel structure refers to Fig. 2.In direct channel,
Direct channel eigenmatrix is extracted using mapping function Y=X;In convolutional channel, all convolution kernel sizes are that 3 × 3, step-length is
1 and weight term all only only used to carry out convolution to last layer output characteristic matrix, not using bias term, therefore l layers
Jth width convolution response characteristic matrixIt can be calculated by formula (1):
In formula, M represents the set of last layer output characteristic matrix;I-th of output characteristic matrix of l-1 layers of expression,The weights of l layers of jth width convolution response characteristic matrix and l-1 layers of the i-th width output characteristic matrix are represented, that is, need study
Convolution kernel;" ⊙ " represents that Hadamard product calculations, i.e. matrix corresponding element are multiplied and summed.
Nonlinear transformation is carried out using ReLU activation primitives to each neuron on convolution response characteristic matrix, the is obtained
L layers of eigenmatrix are
2) by step 1) obtained two eigenmatrixes merge, and it is input to maximum pond layer and average pond layer enters
Row Data Dimensionality Reduction;
When carrying out eigenmatrix merging, the eigenmatrix and direct channel eigenmatrix that each two convolutional layer is obtained are done
One-accumulate, is designated as an accumulator module.For convenience of describing, two convolutional layers are designated as a convolution module, each two adds up
Module is designated as a structure soon, and adds convolution decay factor to eigenmatrix obtained by convolution module, then obtained by each block structure
Eigenmatrix is defined as follows:
Y=λ1Hn1(X,{W(n1)})+λ2Hn2[Hn1(X,{W(n1)}),{W(n2)}]+X (3)
In formula, X and Y represent input and the output data of block structure respectively;λ1And λ2As " convolution decay factor ", is one
The individual constant for needing to set in advance, is respectively provided with different size of convolution decay factor for each convolution module here;W(n1)And W(n2)The set of eigenmatrix weighting parameter when representing to add up twice respectively, is the parameter for needing to train;Hn1() and Hn2
() represents cumulative for the first time and second of cumulative equivalency transform function respectively.
Wherein, equivalency transform function H cumulative for the first timen1The expression formula of () is:
Hn1(X,{W(n1))=W(m2)⊙f(W(m1)⊙X) (4)
Second of cumulative equivalency transform function Hn2The expression formula of () is:
Hn2(Yn1,{W(n2))=W(m4)⊙f(W(m3)⊙Yn1) (5)
Describe for convenience, make Yn1=Hn1(X,{W(n1)), Yn2=Hn2(Yn1,{W(n2)), therefore formula (3) can also be described
Into following form:
Y=λ1Yn1+λ2Yn2+X (6)
Further to Data Dimensionality Reduction, to reduce deep layer network calculations amount, the eigenmatrix after above-mentioned merging is carried out maximum
Pondization and average pond, double pond modes are as shown in figure 3, two kinds of pond methods use size for filtering that 3 × 3, step-length is 2
Device;The wave filter in wherein maximum pond is that filtering selects maximum, can keeping characteristics matrix as far as possible conspicuousness information;Average
Pond be calculate wave filter in character pair average, can keeping characteristics matrix background information.Pass through double ponds, eigenmatrix
Number is doubled, and the dimension of single matrix halves.Experiment display, the pond mode in double ponds is than single on CIFAR-10 data sets
Pond mode can obtain higher classification accuracy, as shown in Figure 5.
3) repeat step 1), step 2), obtain final eigenmatrix;
Because color image size is 32 × 32 in data set, therefore feature extraction twice and Shuan Chi are carried out to eigenmatrix
Change, obtain final eigenmatrix, can so deepen network depth as far as possible.
4) by above-mentioned steps 3) obtained final eigenmatrix carries out global average pond and inputs full articulamentum being changed into one
Dimensional feature matrix, and one-dimensional characteristic matrix is classified using softmax graders and convolutional neural networks are trained,
Calculate the penalty values of this network training;
By above-mentioned steps 3) obtained final eigenmatrix carries out global average pond and inputs full articulamentum, can obtain feature
Dimension is (1 × 1) × q eigenmatrix, and it is classified using softmax graders, wherein, q is class categories number, for every
The set that class has r sample can be expressed as { (x(1),y(1)),(x(2),y(2)),...(x(r),y(r)), y(r)∈{1,2,…,
q}.Calculating the method for the penalty values of this network training is:First, Probability p that each classification j occurs (y=j | x) is calculated;So
After use hθ(x) q output probability is represented, then function is:
Wherein, hθ(x) output of convolutional neural networks is represented, i is sample sequence number, and θ is network parameter,For normalizing
Change the factor;
Finally use cross entropy as loss function counting loss value, its expression formula is:
Wherein l represents penalty values, 1 { y(i)=j } represent to work as y(i)=j durations are 1, are otherwise sample number for 0, r.
5) gradient calculation is carried out using error backpropagation algorithm, calculates each layer error term and weights gradient;
When carrying out backpropagation calculation error, last layer of error amount is equal to penalty values, i.e. δ(0)=l, k-th cumulative
The error term of module is multiplied by connection weight between the two equal to the error term of+1 accumulator module of kth.Therefore k-th of accumulator module
Error term δ(k)For:
δ(k)=δ(n1)+δ(k+1) (9)
δ in formula(k+1)For the error term of+1 accumulator module of kth;δ(n1)For the error term of the accumulator module convolutional channel.Meter
The key of formula (9) is to calculate δ(n1), then can just obtain the error term of remaining accumulator module successively by formula (9):
In above formula, λ1And λ2The convolution decay factor of respectively two convolution modules, is disposed as the positive integer less than 1;
δ(m1)、δ(m2)、δ(m3)And δ(m4)It is followed successively by the error term of four convolutional layers in Fig. 4;δ(n1)And δ(n2)Respectively two convolutional channels
Error term.
Every layer in formula (10) of error term is substituted into successively can obtain the error term of convolutional channel:
In formula (11) 1. formula and 2. formula be respectively just equivalency transform functional expression (4) and formula (5) derivative with therefore can be by formula
(11) it is abbreviated as:
δ(n1)=[(λ2δ(k+1))*Hn1′(A(m0))+λ1δ(k+1)]*Hn2′(A(m2)) (12)
Formula (12) is substituted into formula (7), it is possible to calculate the mistake of all accumulator modules of binary channels convolutional neural networks one by one
Poor item.
According to BP chain rules and gradient calculation formulaCan draw k-th accumulator module last
Layer weights gradient be:
6) according to step 4) in gained penalty values judge whether network restrains, if do not restrained, according to step 5) middle acquisition
Weights gradient adjusts convolutional neural networks initiation parameter and is trained again, and network training result is exported if having restrained.
According to step 4) in loss function calculate penalty values, penalty values and classification thresholds are compared, such as less than divided
Class threshold value then network convergence;Otherwise according to step 5) in gained weights gradient according to formula (14), (15) update weights gradient, lay equal stress on
New training network.
W(t+1):=W (t)+V (t+1) (15)
T represents iterations in formula, and V (t) is momentum term, and μ is factor of momentum, and it determines history weight correction to this
The contribution of secondary weight amendment;η is learning rate;λ is weight attenuation coefficient.
The present invention is overcome the unstable problem of deep layer convolutional neural networks gradient, can use up by the introducing of direct channel
Network may be deepened, so as to improve classification accuracy.It was found from the test result on CIFAR-10 data sets, with convolutional Neural
The increase of network depth, classification accuracy rate is improved, and refers to Fig. 6.
Claims (6)
1. a kind of improved parallel channel convolutional neural networks training method, it is characterised in that:Described method is included in order
The following steps of progress:
1) direct-connected and two parallel channels of convolution are utilized respectively feature extraction is carried out to the data in convolutional neural networks, obtain straight
Communication channel eigenmatrix and convolutional channel eigenmatrix;
2) by step 1) obtained two eigenmatrixes merge, and it is input to maximum pond layer and average pond layer enters line number
According to dimensionality reduction;
3) repeat step 1), step 2), obtain final eigenmatrix;
4) by above-mentioned steps 3) obtained final eigenmatrix carries out global average pond and inputs full articulamentum being changed into one-dimensional spy
Matrix is levied, and one-dimensional characteristic matrix is classified using softmax graders and convolutional neural networks are trained, is calculated
The penalty values of this network training;
5) gradient calculation is carried out using error backpropagation algorithm, calculates each layer error term and weights gradient;
6) according to step 4) in gained penalty values judge whether network restrains, if do not restrained, according to step 5) weights of middle acquisition
Gradient adjusts convolutional neural networks initiation parameter and is trained again, and network training result is exported if having restrained.
2. improved parallel channel convolutional neural networks training method according to claim 1, it is characterised in that:In step
1) in, described direct-connected and two parallel channels of convolution that are utilized respectively carry out feature extraction to the data in convolutional neural networks,
Obtaining the method for direct channel eigenmatrix and convolutional channel eigenmatrix is:First, data are inputted respectively direct channel and
Convolutional channel;Then data are directly mapped as to direct channel eigenmatrix as output in direct channel, in convolutional channel
Upper to carry out convolution operation to data using multiple convolutional layers, the input of each convolutional layer is the output of a upper convolutional layer, will most
Latter convolutional layer output matrix as convolutional channel eigenmatrix.
3. improved parallel channel convolutional neural networks training method according to claim 1, it is characterised in that:In step
2) it is described by step 1 in) obtained two eigenmatrixes are merged, and it is input to maximum pond layer and average pond layer
Carrying out the method for Data Dimensionality Reduction is:First, eigenmatrix obtained by eigenmatrix obtained by direct channel and convolutional channel is closed
And, that is, obtain the set of multiple eigenmatrixes;Then gained eigenmatrix is inputted into maximum pond layer and average pond layer respectively,
In maximum pond layer, the maximum being worth in wave filter is taken using wave filter, layer is taken in wave filter using wave filter in average pond
Average value.
4. improved parallel channel convolutional neural networks training method according to claim 1, it is characterised in that:In step
4) it is described by above-mentioned steps 3 in) obtained final eigenmatrix carries out global average pond and inputs full articulamentum being changed into one
Dimensional feature matrix, and one-dimensional characteristic matrix is classified using softmax graders and convolutional neural networks are trained,
Calculating the method for the penalty values of this network training is:First, global average pond is carried out to final eigenmatrix, using with most
Whole eigenmatrix wave filter of the same size calculates the average value of data in eigenmatrix;Then, full articulamentum, full connection are inputted
Each neuron carries out nonlinear transformation to the data in the eigenmatrix behind global average pond respectively and obtains one-dimensional spy in layer
Levy matrix;Finally, one-dimensional characteristic Input matrix softmax graders are classified.
5. improved parallel channel convolutional neural networks training method according to claim 1, it is characterised in that:In step
5) in, described carries out gradient calculation using error backpropagation algorithm, calculates the method for each layer error term and weights gradient and is:
First, last layer of penalty values are calculated and as the error term of last layer according to softmax classifier results;Then, utilize
The chain rule of error backpropagation algorithm calculates each layer error term, the error term of the ith feature matrix of l-th of convolutional layer
For:
M is l+1 layers of eigenmatrix number,The error term of j-th of eigenmatrix of the l+1 convolutional layer is represented,
Represent to this layer of activation primitive derivation, J represents this layer of penalty values,Represent j-th of eigenmatrix of l layers to i-th of l+1 layers
The connection weight of eigenmatrix;
Finally, formula is utilizedEach layer weights gradient is calculated, whereinRepresent l-1 layers of ith feature
Matrix.
6. improved parallel channel convolutional neural networks training method according to claim 1, it is characterised in that:In step
6) it is described according to step 4 in) in gained penalty values judge whether network restrains, if do not restrained, according to step 5) middle acquisition
Weights gradient adjusts convolutional neural networks initiation parameter and is trained again, as restrained then output network training result
Method is:First, classification results and actual value are compared into simultaneously calculating difference and is used as penalty values;Then penalty values and realization are set
Fixed classification thresholds are compared, and such as less than classification thresholds then judge network convergence, otherwise do not restrain;Finally, such as convergence output net
Network result, otherwise according to formulaW(t+1):=W (t)+V (t+1) adjusts convolution
Neutral net initiation parameter, wherein t represent iterations, and V (t) is momentum term, and μ is factor of momentum, and it determines that history is weighed
Rebuild contribution of the positive quantity to this weight amendment;η is learning rate;λ is weight attenuation coefficient, and W represents convolutional neural networks
Initiation parameter.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710247556.4A CN107092960A (en) | 2017-04-17 | 2017-04-17 | A kind of improved parallel channel convolutional neural networks training method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710247556.4A CN107092960A (en) | 2017-04-17 | 2017-04-17 | A kind of improved parallel channel convolutional neural networks training method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107092960A true CN107092960A (en) | 2017-08-25 |
Family
ID=59637594
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710247556.4A Pending CN107092960A (en) | 2017-04-17 | 2017-04-17 | A kind of improved parallel channel convolutional neural networks training method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107092960A (en) |
Cited By (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107729872A (en) * | 2017-11-02 | 2018-02-23 | 北方工业大学 | Facial expression recognition method and device based on deep learning |
CN108182456A (en) * | 2018-01-23 | 2018-06-19 | 哈工大机器人(合肥)国际创新研究院 | A kind of target detection model and its training method based on deep learning |
CN108460457A (en) * | 2018-03-30 | 2018-08-28 | 苏州纳智天地智能科技有限公司 | A kind of more asynchronous training methods of card hybrid parallel of multimachine towards convolutional neural networks |
CN108596330A (en) * | 2018-05-16 | 2018-09-28 | 中国人民解放军陆军工程大学 | A kind of full convolutional neural networks of Concurrent Feature and its construction method |
CN108846445A (en) * | 2018-06-26 | 2018-11-20 | 清华大学 | A kind of convolutional neural networks filter technology of prunning branches based on similarity-based learning |
CN108961245A (en) * | 2018-07-06 | 2018-12-07 | 西安电子科技大学 | Picture quality classification method based on binary channels depth parallel-convolution network |
CN109102017A (en) * | 2018-08-09 | 2018-12-28 | 百度在线网络技术(北京)有限公司 | Neural network model processing method, device, equipment and readable storage medium storing program for executing |
CN109144099A (en) * | 2018-08-28 | 2019-01-04 | 北京航空航天大学 | Unmanned aerial vehicle group action scheme fast evaluation method based on convolutional neural networks |
CN109635790A (en) * | 2019-01-28 | 2019-04-16 | 杭州电子科技大学 | A kind of pedestrian's abnormal behaviour recognition methods based on 3D convolution |
CN109739979A (en) * | 2018-12-11 | 2019-05-10 | 中科恒运股份有限公司 | Tuning method, tuning device and the terminal of neural network |
CN109993301A (en) * | 2017-12-29 | 2019-07-09 | 北京中科寒武纪科技有限公司 | Neural metwork training device and Related product |
CN110007959A (en) * | 2017-11-03 | 2019-07-12 | 畅想科技有限公司 | Hard-wired stratification mantissa bit length for deep neural network selects |
CN110020652A (en) * | 2019-01-07 | 2019-07-16 | 新而锐电子科技(上海)有限公司 | The dividing method of Tunnel Lining Cracks image |
CN110222559A (en) * | 2019-04-24 | 2019-09-10 | 深圳市微纳集成电路与系统应用研究院 | Smog image detecting method and device based on convolutional neural networks |
CN110276345A (en) * | 2019-06-05 | 2019-09-24 | 北京字节跳动网络技术有限公司 | Convolutional neural networks model training method, device and computer readable storage medium |
CN110555125A (en) * | 2018-05-14 | 2019-12-10 | 桂林远望智能通信科技有限公司 | Vehicle retrieval method based on local features |
CN110580523A (en) * | 2018-06-07 | 2019-12-17 | 清华大学 | Error calibration method and device for analog neural network processor |
CN110766681A (en) * | 2019-10-28 | 2020-02-07 | 福建帝视信息科技有限公司 | Bamboo strip surface defect detection method based on triple loss network |
CN110942106A (en) * | 2019-12-13 | 2020-03-31 | 东华大学 | Pooling convolutional neural network image classification method based on square average |
CN110956342A (en) * | 2020-01-02 | 2020-04-03 | 中国民航大学 | CliqueNet flight delay prediction method based on attention mechanism |
CN111027630A (en) * | 2019-12-13 | 2020-04-17 | 安徽理工大学 | Image classification method based on convolutional neural network |
CN111291771A (en) * | 2018-12-06 | 2020-06-16 | 西安宇视信息科技有限公司 | Method and device for optimizing characteristics of pooling layer |
CN111967574A (en) * | 2020-07-20 | 2020-11-20 | 华南理工大学 | Convolutional neural network training method based on tensor singular value delimitation |
CN112005532A (en) * | 2017-11-08 | 2020-11-27 | 爱维士软件有限责任公司 | Malware classification of executable files over convolutional networks |
CN112020721A (en) * | 2018-06-15 | 2020-12-01 | 富士通株式会社 | Training method and device for classification neural network for semantic segmentation, and electronic equipment |
CN112801929A (en) * | 2021-04-09 | 2021-05-14 | 宝略科技(浙江)有限公司 | Local background semantic information enhancement method for building change detection |
CN113011567A (en) * | 2021-03-31 | 2021-06-22 | 深圳精智达技术股份有限公司 | Training method and device of convolutional neural network model |
US11455802B2 (en) | 2018-03-29 | 2022-09-27 | Beijing Bytedance Network Technology Co. Ltd. | Video feature extraction method and device |
CN109840584B (en) * | 2017-11-24 | 2023-04-18 | 腾讯科技(深圳)有限公司 | Image data classification method and device based on convolutional neural network model |
-
2017
- 2017-04-17 CN CN201710247556.4A patent/CN107092960A/en active Pending
Cited By (40)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107729872A (en) * | 2017-11-02 | 2018-02-23 | 北方工业大学 | Facial expression recognition method and device based on deep learning |
CN110007959A (en) * | 2017-11-03 | 2019-07-12 | 畅想科技有限公司 | Hard-wired stratification mantissa bit length for deep neural network selects |
CN112005532A (en) * | 2017-11-08 | 2020-11-27 | 爱维士软件有限责任公司 | Malware classification of executable files over convolutional networks |
CN109840584B (en) * | 2017-11-24 | 2023-04-18 | 腾讯科技(深圳)有限公司 | Image data classification method and device based on convolutional neural network model |
CN109993301B (en) * | 2017-12-29 | 2020-05-19 | 中科寒武纪科技股份有限公司 | Neural network training device and related product |
CN109993301A (en) * | 2017-12-29 | 2019-07-09 | 北京中科寒武纪科技有限公司 | Neural metwork training device and Related product |
CN108182456A (en) * | 2018-01-23 | 2018-06-19 | 哈工大机器人(合肥)国际创新研究院 | A kind of target detection model and its training method based on deep learning |
CN108182456B (en) * | 2018-01-23 | 2022-03-18 | 哈工大机器人(合肥)国际创新研究院 | Target detection model based on deep learning and training method thereof |
US11455802B2 (en) | 2018-03-29 | 2022-09-27 | Beijing Bytedance Network Technology Co. Ltd. | Video feature extraction method and device |
CN108460457A (en) * | 2018-03-30 | 2018-08-28 | 苏州纳智天地智能科技有限公司 | A kind of more asynchronous training methods of card hybrid parallel of multimachine towards convolutional neural networks |
CN110555125A (en) * | 2018-05-14 | 2019-12-10 | 桂林远望智能通信科技有限公司 | Vehicle retrieval method based on local features |
CN108596330B (en) * | 2018-05-16 | 2022-03-15 | 中国人民解放军陆军工程大学 | Parallel characteristic full-convolution neural network device and construction method thereof |
CN108596330A (en) * | 2018-05-16 | 2018-09-28 | 中国人民解放军陆军工程大学 | A kind of full convolutional neural networks of Concurrent Feature and its construction method |
CN110580523A (en) * | 2018-06-07 | 2019-12-17 | 清华大学 | Error calibration method and device for analog neural network processor |
CN112020721A (en) * | 2018-06-15 | 2020-12-01 | 富士通株式会社 | Training method and device for classification neural network for semantic segmentation, and electronic equipment |
CN108846445A (en) * | 2018-06-26 | 2018-11-20 | 清华大学 | A kind of convolutional neural networks filter technology of prunning branches based on similarity-based learning |
CN108846445B (en) * | 2018-06-26 | 2021-11-26 | 清华大学 | Image processing method |
CN108961245A (en) * | 2018-07-06 | 2018-12-07 | 西安电子科技大学 | Picture quality classification method based on binary channels depth parallel-convolution network |
CN109102017A (en) * | 2018-08-09 | 2018-12-28 | 百度在线网络技术(北京)有限公司 | Neural network model processing method, device, equipment and readable storage medium storing program for executing |
CN109102017B (en) * | 2018-08-09 | 2021-08-03 | 百度在线网络技术(北京)有限公司 | Neural network model processing method, device, equipment and readable storage medium |
CN109144099A (en) * | 2018-08-28 | 2019-01-04 | 北京航空航天大学 | Unmanned aerial vehicle group action scheme fast evaluation method based on convolutional neural networks |
CN111291771A (en) * | 2018-12-06 | 2020-06-16 | 西安宇视信息科技有限公司 | Method and device for optimizing characteristics of pooling layer |
CN111291771B (en) * | 2018-12-06 | 2024-04-02 | 西安宇视信息科技有限公司 | Method and device for optimizing pooling layer characteristics |
CN109739979A (en) * | 2018-12-11 | 2019-05-10 | 中科恒运股份有限公司 | Tuning method, tuning device and the terminal of neural network |
CN110020652A (en) * | 2019-01-07 | 2019-07-16 | 新而锐电子科技(上海)有限公司 | The dividing method of Tunnel Lining Cracks image |
CN109635790A (en) * | 2019-01-28 | 2019-04-16 | 杭州电子科技大学 | A kind of pedestrian's abnormal behaviour recognition methods based on 3D convolution |
CN110222559A (en) * | 2019-04-24 | 2019-09-10 | 深圳市微纳集成电路与系统应用研究院 | Smog image detecting method and device based on convolutional neural networks |
CN110276345A (en) * | 2019-06-05 | 2019-09-24 | 北京字节跳动网络技术有限公司 | Convolutional neural networks model training method, device and computer readable storage medium |
CN110766681A (en) * | 2019-10-28 | 2020-02-07 | 福建帝视信息科技有限公司 | Bamboo strip surface defect detection method based on triple loss network |
CN110766681B (en) * | 2019-10-28 | 2023-04-14 | 福建帝视信息科技有限公司 | Bamboo strip surface defect detection method based on triple loss network |
CN111027630A (en) * | 2019-12-13 | 2020-04-17 | 安徽理工大学 | Image classification method based on convolutional neural network |
CN110942106A (en) * | 2019-12-13 | 2020-03-31 | 东华大学 | Pooling convolutional neural network image classification method based on square average |
CN111027630B (en) * | 2019-12-13 | 2023-04-07 | 安徽理工大学 | Image classification method based on convolutional neural network |
CN110942106B (en) * | 2019-12-13 | 2023-11-07 | 东华大学 | Pooled convolutional neural network image classification method based on square average |
CN110956342A (en) * | 2020-01-02 | 2020-04-03 | 中国民航大学 | CliqueNet flight delay prediction method based on attention mechanism |
CN111967574B (en) * | 2020-07-20 | 2024-01-23 | 华南理工大学 | Tensor singular value delimitation-based convolutional neural network training method |
CN111967574A (en) * | 2020-07-20 | 2020-11-20 | 华南理工大学 | Convolutional neural network training method based on tensor singular value delimitation |
CN113011567B (en) * | 2021-03-31 | 2023-01-31 | 深圳精智达技术股份有限公司 | Training method and device of convolutional neural network model |
CN113011567A (en) * | 2021-03-31 | 2021-06-22 | 深圳精智达技术股份有限公司 | Training method and device of convolutional neural network model |
CN112801929A (en) * | 2021-04-09 | 2021-05-14 | 宝略科技(浙江)有限公司 | Local background semantic information enhancement method for building change detection |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107092960A (en) | A kind of improved parallel channel convolutional neural networks training method | |
CN107247989A (en) | A kind of neural network training method and device | |
CN108009594B (en) | A kind of image-recognizing method based on change grouping convolution | |
CN108416755A (en) | A kind of image de-noising method and system based on deep learning | |
CN106991440A (en) | A kind of image classification algorithms of the convolutional neural networks based on spatial pyramid | |
CN108764471A (en) | The neural network cross-layer pruning method of feature based redundancy analysis | |
CN107563567A (en) | Core extreme learning machine Flood Forecasting Method based on sparse own coding | |
CN107437096A (en) | Image classification method based on the efficient depth residual error network model of parameter | |
CN106934352A (en) | A kind of video presentation method based on two-way fractal net work and LSTM | |
CN110245665A (en) | Image, semantic dividing method based on attention mechanism | |
CN106650789A (en) | Image description generation method based on depth LSTM network | |
CN109948029A (en) | Based on the adaptive depth hashing image searching method of neural network | |
CN106951395A (en) | Towards the parallel convolution operations method and device of compression convolutional neural networks | |
CN106971160A (en) | Winter jujube disease recognition method based on depth convolutional neural networks and disease geo-radar image | |
CN109543502A (en) | A kind of semantic segmentation method based on the multiple dimensioned neural network of depth | |
CN107609638A (en) | A kind of method based on line decoder and interpolation sampling optimization convolutional neural networks | |
CN104050507B (en) | Hyperspectral image classification method based on multilayer neural network | |
CN110263863A (en) | Fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2 | |
CN106780482A (en) | A kind of classification method of medical image | |
CN107392224A (en) | A kind of crop disease recognizer based on triple channel convolutional neural networks | |
CN108921298A (en) | Intensified learning multiple agent is linked up and decision-making technique | |
CN109817276A (en) | A kind of secondary protein structure prediction method based on deep neural network | |
CN110321967A (en) | Image classification innovatory algorithm based on convolutional neural networks | |
CN106951960A (en) | A kind of learning method of neutral net and the neutral net | |
CN106339753A (en) | Method for effectively enhancing robustness of convolutional neural network |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170825 |