CN109978003A - Image classification method based on intensive connection residual error network - Google Patents
Image classification method based on intensive connection residual error network Download PDFInfo
- Publication number
- CN109978003A CN109978003A CN201910135688.7A CN201910135688A CN109978003A CN 109978003 A CN109978003 A CN 109978003A CN 201910135688 A CN201910135688 A CN 201910135688A CN 109978003 A CN109978003 A CN 109978003A
- Authority
- CN
- China
- Prior art keywords
- residual error
- network
- intensive connection
- image
- error network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 32
- 239000000284 extract Substances 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 9
- 230000013016 learning Effects 0.000 claims description 7
- 239000000203 mixture Substances 0.000 claims description 5
- 210000002569 neuron Anatomy 0.000 claims description 3
- 238000012549 training Methods 0.000 abstract description 7
- 238000013135 deep learning Methods 0.000 abstract description 5
- 238000004364 calculation method Methods 0.000 abstract description 4
- 230000007547 defect Effects 0.000 abstract description 4
- 230000000694 effects Effects 0.000 abstract description 4
- 238000007796 conventional method Methods 0.000 abstract description 3
- 230000015654 memory Effects 0.000 description 8
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000003475 lamination Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image classification methods based on intensive connection residual error network, and this method carries out data prediction to input picture first, and expands image data amount；Then picture is input to intensive connection residual error network and extracts characteristics of image；The characteristics of image extracted is finally input to Softmax classifier, obtains picture classification result.This method overcomes the defect that conventional method effect is bad, common deep learning method calculation amount is too big, model training and operation are too high to hardware device requirement in image classification task, the advantages of by combining residual error network and intensive connection network, less computing resource is occupied while reaching higher discrimination.
Description
Technical field
The present invention relates to machine learning and technical field of machine vision, more particularly to one kind is based on intensive connection residual error network
Image classification method.
Background technique
Image classification is one of most active research topic in computer vision field, quickly and accurately identifies object
It has great significance to robot navigation, medical diagnosis, security protection, industrial detection etc..Extracting characteristics of image is in object identification
Most important is also most intractable work, and the quality of characteristics of image has been largely fixed the effect of identification.Compared to traditional figure
As feature extracting method, deep neural network can extract the stronger characteristics of image of ability to express, be more advantageous to image classification, because
Most of object identification frames are all based on deep learning now for this.But current most of deep learnings for image classification
Method calculation amount is very big, and the training time is long, has higher requirement to training equipment.Depth residual error network and intensive connection network are
The outstanding convolutional neural networks model emerged in recent years, depth residual error network is it is possible to prevente effectively from traditional convolutional neural networks mould
The gradient of type disappears or explodes and model degradation problem.Intensive connection network improves the utilization rate of parameter, and network can utilize
Less parameters amount reaches higher discrimination, but a large amount of intensive connections in network inevitably consume a large amount of GPU
Memory, the intensive connection network for calculating deep layer are higher to the hardware requirement of machine.
Summary of the invention
Technical problem to be solved by the invention is to provide a kind of image classification method based on intensive connection residual error network,
This method overcomes conventional depth learning method to be directed to the defect of image classification, passes through combination residual error network and intensive connection network
Advantage reaches higher discrimination using less parameter amount, reduces computing resource and occupies, effectively improves the discrimination of image.
In order to solve the above technical problems, including following step the present invention is based on the image classification method of intensive connection residual error network
It is rapid:
Step 1: carrying out data prediction to input picture, and expand image data amount；
Step 2: picture, which is input to intensive connection residual error network, extracts characteristics of image, the intensive residual error network that connects is by one
A miniature compact connection network and a residual error structure composition, in intensively connection residual error network, serial operation will be in network
It is input in residual error function again after the output characteristic pattern series connection of all convolutional layers；
Step 3: the characteristics of image extracted is input to Softmax classifier, picture classification result is obtained.
Further, all sample standard deviations are subtracted the mean value and variance of entire data set sample by the data prediction, so that
The convergence of model is faster；The expansion image data amount trains the random overturning of picture and rotation, with the multiplicity of expanding data
Property, so that model learning has more the feature represented.
Further, the intensive connection residual error network is indicated using following formula:
yl=yl-1+F(Hd(y0..., yd]), [W0..., Wd])
Wherein yl-1And ylRespectively first of network intensively connects outputting and inputting for residual error structure, [y0..., yd]
[y0..., yd] indicate all convolutional layers of intensive connection residual error network internal, [W0, Wd] indicate in intensive connection residual error network
The corresponding weight parameter of each convolutional layer in portion.
Further, in the picture classification operation of the Softmax classifier, the output of Softmax classifier is each
Sample belongs to the probability of each classification, and the expression formula of loss function is shown below:
Wherein, xiIndicate the input of i-th of neuron of output layer, θ is learning parameter.
It is since the present invention is based on the image classification methods of intensive connection residual error network to use above-mentioned technical proposal, i.e., our
Method carries out data prediction to input picture first, and expands image data amount；Then picture is input to intensive connection residual error
Network extracts characteristics of image；The characteristics of image extracted is finally input to Softmax classifier, obtains picture classification result.
This method overcomes in image classification task that conventional method effect is bad, common deep learning method calculation amount is too big, model instruction
Practice and run and too high defect is required to hardware device, the advantages of by combining residual error network and intensive connection network, reach compared with
Less computing resource is occupied while high discrimination.
Detailed description of the invention
The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:
Fig. 1 is residual error schematic network structure；
Fig. 2 is intensive connection schematic network structure；
Fig. 3 is the intensive connection residual error schematic network structure with two layers of convolution；
Fig. 4 occupies contrast schematic diagram with the intensive network internal storage that connect for common convolutional neural networks.
Specific embodiment
The present invention is based on the image classification methods of intensive connection residual error network to include the following steps:
Step 1: carrying out data prediction to input picture, and expand image data amount；
Step 2: picture, which is input to intensive connection residual error network, extracts characteristics of image, the intensive residual error network that connects is by one
A miniature compact connection network and a residual error structure composition, in intensively connection residual error network, serial operation will be in network
It is input in residual error function again after the output characteristic pattern series connection of all convolutional layers；
Usual depth residual error network is piled up by a series of residual error structures, residual error structure can following formula indicate:
yl=yl-1+F(yl-1, Wl)
Wherein yl-1Indicate the input of first of residual error structure, ylIndicate the output of first of residual error structure, WlIndicate first it is residual
The weight parameter of poor structure, F indicate that residual error function as long as making F ()=0, that is, completes in residual error network training process
One identical mapping yl=yl-1: obviously, go one determining function F ()=0 of fitting not true compared to some is approached with network
Fixed function yl=yl-1It is more easier, residual error network structure is as shown in Figure 1；
Intensive connection network has extremely strong ability in feature extraction and higher parameter utilization rate, the intensive core for connecting network
Thought thinks that each layer of convolution all receives the output of its all convolutional layer previous as input, intensive connection network structure such as Fig. 2
It is shown；L layers of input may be defined as in intensive connection network:
yl=Hl([y0, y1..., yl-1])
Wherein [y0, y1..., yl-1] indicate the 0th, 1 ..l-1 layer of output, H expression serial operation, by [y0, y1,
yl-1] be connected in series together after as l layers of input；The characteristic pattern of each layer of all layers of front of reception of network can not only be protected
It is transferred to card information smoothing deep layer, and the utilization rate of parameter can be improved；But deeper intensive connection network needs
More memories are wanted, because a large amount of intensive connections in network inevitably consume a large amount of GPU memory；
Therefore it is directed to the excessive problem of EMS memory occupation existing for intensive connection network, and sufficiently in conjunction with residual error network and intensively
Connect the advantage of network, this method proposes a kind of novel intensive connection residual error network structure, this intensively connect residual error structure by
One small-sized intensive connection network and a residual error structure composition, Fig. 3 are the intensive connection residual error with two layers of convolution
Network structure；
Compared with raw residual structure, the intensive input for connecting the summation layer (residual error operation) in residual error network structure is by original
The output of the last one convolutional layer of beginning structure becomes convolutional layer output all in residual error structure, can make so each residual
Convolutional layer in poor structure is fully used, and intensive connect residual error network (DRN) is by a series of intensive connection residual error structure
Composition；
Step 3: the characteristics of image extracted is input to softmax classifier, picture classification result is obtained.
Preferably, all sample standard deviations are subtracted the mean value and variance of entire data set sample by the data prediction, so that
The convergence of model is faster；The expansion image data amount trains the random overturning of picture and rotation, with the multiplicity of expanding data
Property, so that model learning has more the feature represented.
Preferably, the intensive connection residual error network is indicated using following formula:
yl=yl-1+F(Hd([y0..., yd]), [W0..., Wd])
Wherein yl-1And ylRespectively first of network intensively connects outputting and inputting for residual error structure, [y0..., yd]
[y0... .yd] indicate all convolutional layers of intensive connection residual error network internal, [W0..., Wd] indicate intensive connection residual error net
The corresponding weight parameter of each convolutional layer in network inside.
Preferably, in the picture classification operation of the Softmax classifier, the output of Softmax classifier is each
Sample belongs to the probability of each classification, and the expression formula of loss function is shown below:
Wherein, xiIndicate the input of i-th of neuron of output layer, θ is learning parameter.
This method is described in detail with reference to the accompanying drawing, the present embodiment proposes close first of all for verifying this method
The performance of collection connection residual error network, carries out image classification experiment in two benchmark datasets of CIFAR-10 and CIFAR-100；It connects
In the comparable situation of network parameter amount, compare raw residual network and with Bu Tong intensively connection residual error structure intensively connecting
Residual error network is connect in performances such as accuracy rate, the training speeds of CIFAR-10 data set；Finally to the intensive connection residual error net of proposition
Network and intensive connection network are compared in terms of performance and EMS memory occupation.
Firstly, in order to verify the performance of the intensive connection residual error network of this method proposition, in CIFAR-10 and CIFAR-100
Image classification experiment is carried out in the two benchmark datasets；For the experiment, designs a specific 2D and intensively connect residual error net
Network, the network include a convolution kernel having a size of 3*3, the initial convolutional layer that convolution kernel number is 32, are intensively connected with being followed by three kinds
Residual error structure is connect, these three convolution kernel sizes for intensively connecting residual error structure are all 3*3, and convolution kernel number is respectively 16,32 and
64, it is finally that an overall situation is averaged pond layer to feature progress dimensionality reduction, is then fed into full articulamentum and classifies.For CIFAR
The network structure of data set is as shown in table 1, and wherein N indicates the quantity of residual error structure, and K indicates intensive connection residual error inside configuration volume
The depth of the number of lamination, network is controlled by N and K, and such as when K is 2, N is 8, the intensive depth for connecting residual error network is 50
Layer；This method has used three kinds of different intensive connection residual error networks, and K is respectively 2,3,4, and is respectively designated as DRN-A, DRN-
B and DRN-C.
Table 1
Then, in the comparable situation of network parameter amount, compare raw residual network and there is Bu Tong intensively connection residual error
The intensive connection residual error network of structure is in performances such as accuracy rate, the training speeds of CIFAR-10 data set；Experimental result such as 2 institute of table
Show, wherein what intensively connection residual error network A, B, C respectively indicated the intensive connection residual error Web vector graphic is to have 2,3,4 layers of convolution
Intensive connection residual error structure, from the results shown in Table 2 in the comparable situation of parameter amount, it is intensive connect residual error network without
By being that accuracy rate or the speed of service are all substantially better than raw residual network, this is the result shows that intensively connect the ginseng of residual error network
Number utilization rate is higher.
Table 2
Finally, the intensive connection residual error network and intensive connection network to proposition compare in terms of performance and EMS memory occupation
Compared with the intensive network that connects has excellent ability in feature extraction, but due to the particularity of intensive connection structure, a large amount of series connection behaviour
Make so that intensively the EMS memory occupation amount of connection network increased significantly compared to common depth network, as shown in Figure 4；Since this is lacked
It falls into, the intensive network that connects needs better GPU that could run；The present embodiment is intensive using similar depth and close network-wide
Connection residual error network with intensively connect network and compares, table 3 is intensive connection residual error network and intensively connects the performance of network
Comparison result, wherein DenseNet-40-48 expression is the intensive connection network that one 40 layers of growth rate is 48, DRN-38-C-
2 indicate it is the 38 layers of intensive connection residual error network constituted using intensive connection residual error structure C, and by each layer of network
The convolution kernel number of convolution expands 2 times, to keep and network capacity similar in intensive connection network；It can from the result of table 3
Out, residual error network is intensively connected in the occupancy for reaching with significantly reducing memory while intensively connecting accuracy rate similar in network
Amount.
Table 3
This method overcome in image classification task conventional method effect is bad, common deep learning method calculation amount too
Greatly, model training and operation require too high defect to hardware device, by combining the excellent of residual error network and intensive connection network
Point occupies less computing resource, effectively reduces hardware cost while capable of reaching higher discrimination, improve image point
The application scenarios of class.
Claims (4)
1. a kind of image classification method based on intensive connection residual error network, it is characterised in that this method includes the following steps:
Step 1: carrying out data prediction to input picture, and expand image data amount；
Step 2: picture, which is input to intensive connection residual error network, extracts characteristics of image, intensive connection residual error network is small by one
Type intensively connects network and a residual error structure composition, and in intensively connection residual error network, serial operation will be all in network
It is input in residual error function again after the output characteristic pattern series connection of convolutional layer；
Step 3: the characteristics of image extracted is input to Softmax classifier, picture classification result is obtained.
2. the image classification method according to claim 1 based on intensive connection residual error network, it is characterised in that: the number
All sample standard deviations are subtracted the mean value and variance of entire data set sample by Data preprocess, so that the convergence of model is faster；The expansion
It fills image data amount to train the random overturning of picture and rotation, with the diversity of expanding data, so that model learning has more generation
The feature of table.
3. the image classification method according to claim 1 based on intensive connection residual error network, it is characterised in that: described close
Collection connection residual error network is indicated using following formula:
yl=yl-1+F(Hd([y0..., yd]), [W0..., Wd])
Wherein yl-1And ylRespectively first of network intensively connects outputting and inputting for residual error structure, [y0..., yd]
[y0..., yd] indicate all convolutional layers of intensive connection residual error network internal, [W0..., Wd] indicate intensive connection residual error net
The corresponding weight parameter of each convolutional layer in network inside.
4. the image classification method according to claim 1 based on intensive connection residual error network, it is characterised in that: described
In the picture classification operation of Softmax classifier, the output of Softmax classifier is that each sample belongs to the general of each classification
The expression formula of rate, loss function is shown below:
Wherein, xiIndicate the input of i-th of neuron of output layer, θ is learning parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910135688.7A CN109978003A (en) | 2019-02-21 | 2019-02-21 | Image classification method based on intensive connection residual error network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910135688.7A CN109978003A (en) | 2019-02-21 | 2019-02-21 | Image classification method based on intensive connection residual error network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109978003A true CN109978003A (en) | 2019-07-05 |
Family
ID=67077229
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910135688.7A Pending CN109978003A (en) | 2019-02-21 | 2019-02-21 | Image classification method based on intensive connection residual error network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109978003A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796716A (en) * | 2019-10-21 | 2020-02-14 | 东华理工大学 | Image coloring method based on multiple residual error networks and regularized transfer learning |
| CN111652054A (en) * | 2020-04-21 | 2020-09-11 | 北京迈格威科技有限公司 | Joint point detection method, posture recognition method and device |
| CN111861923A (en) * | 2020-07-21 | 2020-10-30 | 济南大学 | Target identification method and system based on lightweight residual error network image defogging |
| CN112183718A (en) * | 2020-08-31 | 2021-01-05 | 华为技术有限公司 | Deep learning training method and device for computing equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106991646A (en) * | 2017-03-28 | 2017-07-28 | 福建帝视信息科技有限公司 | A kind of image super-resolution method based on intensive connection network |
| CN107437096A (en) * | 2017-07-28 | 2017-12-05 | 北京大学 | Image classification method based on the efficient depth residual error network model of parameter |
| CN108596329A (en) * | 2018-05-11 | 2018-09-28 | 北方民族大学 | Threedimensional model sorting technique based on end-to-end Deep integrating learning network |
| CN108830242A (en) * | 2018-06-22 | 2018-11-16 | 北京航空航天大学 | SAR image targets in ocean classification and Detection method based on convolutional neural networks |
-
2019
- 2019-02-21 CN CN201910135688.7A patent/CN109978003A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106991646A (en) * | 2017-03-28 | 2017-07-28 | 福建帝视信息科技有限公司 | A kind of image super-resolution method based on intensive connection network |
| CN107437096A (en) * | 2017-07-28 | 2017-12-05 | 北京大学 | Image classification method based on the efficient depth residual error network model of parameter |
| CN108596329A (en) * | 2018-05-11 | 2018-09-28 | 北方民族大学 | Threedimensional model sorting technique based on end-to-end Deep integrating learning network |
| CN108830242A (en) * | 2018-06-22 | 2018-11-16 | 北京航空航天大学 | SAR image targets in ocean classification and Detection method based on convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| TIANZHONG SONG 等: "Residual network with dense block", 《JOURNAL OF ELECTRONIC IMAGING》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796716A (en) * | 2019-10-21 | 2020-02-14 | 东华理工大学 | Image coloring method based on multiple residual error networks and regularized transfer learning |
| CN110796716B (en) * | 2019-10-21 | 2023-04-28 | 东华理工大学 | Image coloring method based on multiple residual error network and regularized transfer learning |
| CN111652054A (en) * | 2020-04-21 | 2020-09-11 | 北京迈格威科技有限公司 | Joint point detection method, posture recognition method and device |
| CN111652054B (en) * | 2020-04-21 | 2023-11-03 | 北京迈格威科技有限公司 | Joint point detection method, gesture recognition method and device |
| CN111861923A (en) * | 2020-07-21 | 2020-10-30 | 济南大学 | Target identification method and system based on lightweight residual error network image defogging |
| CN112183718A (en) * | 2020-08-31 | 2021-01-05 | 华为技术有限公司 | Deep learning training method and device for computing equipment |
| WO2022042713A1 (en) * | 2020-08-31 | 2022-03-03 | 华为技术有限公司 | Deep learning training method and apparatus for use in computing device |
| CN112183718B (en) * | 2020-08-31 | 2023-10-10 | 华为技术有限公司 | Deep learning training method and device for computing equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111325155B (en) | Video motion recognition method based on residual difference type 3D CNN and multi-mode feature fusion strategy | |
| CN109978003A (en) | Image classification method based on intensive connection residual error network | |
| CN111369563B (en) | Semantic segmentation method based on pyramid void convolutional network | |
| CN110082821B (en) | Label-frame-free microseism signal detection method and device | |
| CN112784929B (en) | Small sample image classification method and device based on double-element group expansion | |
| CN110781928B (en) | Image similarity learning method for extracting multi-resolution features of image | |
| CN111046900A (en) | Semi-supervised generation confrontation network image classification method based on local manifold regularization | |
| CN107644221A (en) | Convolutional neural networks traffic sign recognition method based on compression of parameters | |
| CN107292267A (en) | Photo fraud convolutional neural networks training method and human face in-vivo detection method | |
| CN113762138B (en) | Identification method, device, computer equipment and storage medium for fake face pictures | |
| CN109359527B (en) | Hair region extraction method and system based on neural network | |
| CN107301396A (en) | Video fraud convolutional neural networks training method and human face in-vivo detection method | |
| CN111612799A (en) | Face data pair-oriented incomplete reticulate pattern face repairing method and system and storage medium | |
| CN110826462A (en) | Human body behavior identification method of non-local double-current convolutional neural network model | |
| CN111242181B (en) | RGB-D saliency object detector based on image semantics and detail | |
| CN111178312B (en) | Face expression recognition method based on multi-task feature learning network | |
| CN105844291A (en) | Characteristic fusion method based on kernel typical correlation analysis | |
| CN112990316A (en) | Hyperspectral remote sensing image classification method and system based on multi-saliency feature fusion | |
| CN112884758A (en) | Defective insulator sample generation method and system based on style migration method | |
| CN112560828A (en) | Lightweight mask face recognition method, system, storage medium and equipment | |
| CN113724354A (en) | Reference image color style-based gray level image coloring method | |
| CN114330516A (en) | Small sample logo image classification based on multi-graph guided neural network model | |
| Chen et al. | Dlfmnet: End-to-end detection and localization of face manipulation using multi-domain features | |
| CN113688715A (en) | Facial expression recognition method and system | |
| CN110414586A (en) | Antifalsification label based on deep learning tests fake method, device, equipment and medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190705 |
|
| RJ01 | Rejection of invention patent application after publication |