CN109858618A - The neural network and image classification method of a kind of convolutional Neural cell block, composition - Google Patents
The neural network and image classification method of a kind of convolutional Neural cell block, composition Download PDFInfo
- Publication number
- CN109858618A CN109858618A CN201910170897.5A CN201910170897A CN109858618A CN 109858618 A CN109858618 A CN 109858618A CN 201910170897 A CN201910170897 A CN 201910170897A CN 109858618 A CN109858618 A CN 109858618A
- Authority
- CN
- China
- Prior art keywords
- convolution kernel
- convolutional neural
- cell block
- convolution
- neural cell
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of convolutional Neural cell block, the neural network and image classification method of composition, convolutional Neural cell block, includingnThe convolution kernel of a different directions and onem×mConvolution kernel heap poststack connectionaA 1 × 1 convolution kernel is constituted；It further include the jump connection of identical transformation, whereinaIt is equal with the port number of input feature vector figure；The present invention decomposes convolution kernel, while guaranteeing that convolution kernel possesses big receptive field, reduces original number of parameters；It proposes to use diagonal convolution, directly acquires the correlation in the direction of former characteristic pattern diagonally opposing corner, the adaptability of spatial alternation is enhanced.
Description
Technical field
The present invention relates to the building of neural network and applied technical fields, and in particular to a kind of convolutional Neural cell block, structure
At neural network and image classification method.
Background technique
Existing computer vision field, convolutional neural networks are common tools；Convolutional neural networks are generally by continuous
Convolutional layer, active coating and pond layer heap are folded forms；Convolutional layer is made of multiple convolutional Neural units, receives upper one layer of input
Feature is calculated with the convolutional Neural unit of this layer, and output channel number is equal to the spy of this layer of convolutional Neural element number
Sign figure.Active coating is made of linear amending unit；It only include a linear amending unit inside a general active coating, to upper one
The characteristic pattern of layer output carries out Nonlinear Mapping；One pond layer is equally that the characteristic pattern exported to upper one layer carries out Chi Huaying
It penetrates.
Existing neural network generally comprises two kinds, the first resnet network, and Resnet is one kind proposed in 2016
Network frame, between the convolutional layer directly stacked originally, the identity mapping that joined jump connection is (identical to reflect for it
Penetrate) so that network itself only needs to be fitted the residual error of script network；Before this, network is by stacking convolutional layer, active coating
Nonlinear transformation is realized with pond layer.Second is inception third version structure, in inception v3 structure
In, the convolution kernel that 3 × 3 convolution kernel resolves into continuous 3 × 1 is superimposed by some structure with 1 × 3 convolution kernel.This portion
Separation structure is similar with structure of the invention.
But in existing convolutional neural networks, since there are many number of plies of network, model parameter be all it is very huge, in order to
The parameter of network is reduced, convolutional neural networks mainly use 3 × 3 and 1 × 1 convolution kernel；This makes the direct feeling of convolutional layer
Wild very little；If using the convolution kernel of big receptive field, and will lead to the increase of parameter amount, so as to cause it is computationally intensive, to space
Indeformable adaptability is poor.
Summary of the invention
The present invention, which provides one kind, can reduce the convolutional Neural cell block of number of parameters, the neural network of composition and image point
Class method.
The technical solution adopted by the present invention is that: a kind of convolutional Neural cell block, the convolution kernel and one including n different directions
The connection of convolution kernel heap poststack a 1 × 1 convolution kernel of a m × m is constituted；Further include identical transformation jump connection, wherein a with
The port number of input feature vector figure is equal.
Further, the convolution kernel of the different directions includes the oblique convolution kernel 1 in oblique convolution kernel b × 1 in a left side and a right side
×b；Left tiltedly convolution kernel is that all positions of b × b convolution kernel other than left diagonally opposing corner all remain 0；Right tiltedly convolution kernel is b
All positions of × b convolution kernel other than right diagonally opposing corner all remain 0.
Further, in the convolution kernel of the n different directions, each direction includes the direction volume of two vertical direction
Long-pending stacking.
A kind of neural network using convolutional Neural cell block, convolution kernel, c convolution mind including sequentially connected m × m
Residual block that the convolution kernel for being d through cell block, step-length, e convolutional Neural cell block, step-length are d, e convolutional Neural cell block,
Step-length be f receptive field be f × f pond layer, full articulamentum.
A kind of image classification method using neural network, comprising the following steps:
Step 1: building neural network；
Step 2: training step 1 constructs obtained neural network；
Step 3: data enhancing processing being carried out to test set picture, is inputted the neural network obtained after step 2 training
In, the classification of image can be completed.
The beneficial effects of the present invention are:
(1) present invention decomposes convolution kernel, while guaranteeing that convolution kernel possesses big receptive field, reduces original ginseng
Number quantity；
(2) present invention proposes to use diagonal convolution, the correlation in the direction of former characteristic pattern diagonally opposing corner is directly acquired, to space
The adaptability of transformation enhances；
(3) present invention is higher for accuracy in image classification, the parameter amount stored needed for reducing.
Detailed description of the invention
Fig. 1 is the diagonally opposing corner convolution kernel used in the present invention, and a is left diagonally opposing corner convolution kernel, and b is right diagonally opposing corner convolution kernel.
Fig. 2 is the convolutional Neural unit block structure used in the embodiment of the present invention.
Fig. 3 is the convolutional Neural unit block structure used in the embodiment of the present invention.
Fig. 4 is the convolutional Neural unit block structure used in the embodiment of the present invention.
Fig. 5 is the convolutional Neural unit block structure used in the embodiment of the present invention.
Specific embodiment
The present invention will be further described in the following with reference to the drawings and specific embodiments.
The convolution kernel heap poststack of a kind of convolutional Neural cell block, convolution kernel and a m × m including n different directions connects
A 1 × 1 convolution kernel is connect to constitute；It further include the jump connection of identical transformation, wherein a is equal with the port number of input feature vector figure.
The convolution kernel of different directions includes oblique convolution kernel b × 1 in a left side and the oblique 1 × b of convolution kernel in a right side；Left oblique convolution kernel
All positions for being b × b convolution kernel other than left diagonally opposing corner all remain 0；Right tiltedly convolution kernel be b × b convolution kernel in addition to
All positions except right diagonally opposing corner all remain 0.
In the convolution kernel of n different directions, each direction includes the stacking of the direction convolution of two vertical direction.
Convolutional Neural unit of the invention is different from common convolutional Neural unit, the receptive field one of common convolutional Neural unit
As be it is rectangular, convolution unit proposed by the present invention is diagonal convolution unit, as shown in Figure 1, a is left oblique convolution kernel, b is right tiltedly volume
Product core.This convolution kernel can extract the correlation in one layer of characteristic pattern oblique angle direction；Required for one diagonal convolutional Neural unit
The parameter amount of occupancy is g × 1 × h, wherein g is the size of convolution kernel, and h is the quantity of convolution kernel；Parameter amount is equivalent to a g
The parameter amount of the normal convolution kernel of × 1 × h or 1 × g × h.The oblique convolution kernel in a left side that a is one 5 × 1 in Fig. 1, working method
It is consistent with normal convolution kernel, it is equivalent to all positions of normal 5 × 5 convolution kernel of convolution other than left diagonally opposing corner and all remains
0, b is same as above.
Fig. 2 is a kind of structure of convolutional Neural cell block of the present invention, for the improvement made on mono- block of original resnet；
Original mono- block of resnet is stacked by two 3 × 3 convolution kernels, along with the jump of an identical transformation connects；Fig. 2 is
Improvement has been made on this basis, and first layer changes the convolution kernel of four large scale different directions into, then plus one 3 × 3
Convolution kernel.It can get the relevance of different directions, the convolution kernel of large scale will increase receptive field, obtain more information.Then
They are stacked, compresses dimension with 1 × 1 convolution kernel, obtains the characteristic pattern of the dimension as input, and and input phase
Add；In order to reduce the parameter amount of network convolution kernel, the convolution nuclear volume of 5 branches of first layer is equally inputted into convolution nuclear volume
Half, after stacking, the quantity of 1 × 1 convolution kernel is equal to the port number of input feature vector figure, and guarantee outputs and inputs feature
The port number of figure is identical, is able to carry out direct addition.Since two 3 × 3 convolution kernels of original resnet network are equivalent to 5 × 5
Receptive field sets 5 × 1 for the scale of four different directions of first layer, another is set as 3 × 3 convolution kernels.
Fig. 3 is the structure of another convolutional Neural cell block proposed on the basis of Fig. 2, by the basis of Fig. 2
One layer the last one branched extensions at two 3 × 3 convolution kernels stacking；This structure can be extracted preferably in characteristic pattern
The correlation of other forms, parameter setting are identical as Fig. 2.
Fig. 4 is another convolutional Neural unit block structure proposed by the present invention, in order to preferably extract four kinds of differences
The correlation in direction；The convolution kernel of different directions is stacked；First layer splits into five branches, point of four different directions
Branch and one 3 × 3 convolution kernel extract the spatial coherence in addition to rectilinear direction.Each branch becomes two Vertical Squares
To convolution stacking, to extract information from different perspectives；Branching into 1 × 3 lateral convolution kernel and 3 × 1 such as first
The stacking of vertical convolution kernel；Second branches into the stacking of 3 × 1 vertical convolution kernel and 1 × 3 lateral convolution kernel；Third
Branch into the stacking of 1 × 3 oblique convolution kernel in the right side and 3 × 1 oblique convolution kernel in a left side；4th branch into 1 × 3 a left side tiltedly convolution kernel and
The stacking of the oblique convolution kernel in 3 × 1 right side.Due to the convolution kernel that first layer is 3 × 1, the convolution kernel of the second layer is set as 3 × 3, can be with
Guarantee to obtain and original identical receptive field.It is identical as convolutional Neural cellular construction shown in Fig. 2 in parameter；In view of parameter amount
The reason of, the port number of first layer convolution kernel is set as the half of input picture, and the parameter amount of the second layer is set as special with input
The port number for levying figure is consistent, facilitates input feature vector figure and its results added.
Fig. 5 is to be made that improvement to the bottleneck of resnet structure, and the bottleneck structure of resnet was by three layer 1 × 1,3 originally
× 3,1 × 1 convolution kernel, which stacks, to be constituted, and the jump for being connected directly from input to output along with one connects；Structure in Fig. 5
For an intermediate layer network is optimized；3 × 3 convolution kernel is decomposed into four direction convolution kernel one 3 × 3 convolution,
For obtaining the feature correlation in addition to rectilinear direction.In order to reduce parameter amount, by the convolution kernel of first layer and the second layer
Quantity all reduces half, is further reduced parameter amount.
Neural network is designed according to above-mentioned four kinds of neural network units, convolution kernel, c volume including sequentially connected m × m
Accumulate the residual block, e convolutional Neural list that neural unit block, the convolution kernel that step-length is d, e convolutional Neural cell block, step-length are d
First block, step-length be f receptive field be f × f pond layer, full articulamentum.
The neural network designed through the invention can carry out image classification, comprising the following steps:
Step 1: building neural network；Neural network can be constructed using python, tensorflow or keras；
Step 2: training step 1 constructs obtained neural network；
Step 3: data enhancing processing being carried out to test set picture, is inputted the neural network obtained after step 2 training
In, the classification of image can be completed.
The convolutional Neural cell block that the present invention designs can be replaced any resnet block, by taking resnet32 as an example into
Row image classification, concrete scheme are as follows:
S1: being 16 by one layer of port number after image input, and the convolution kernel that receptive field is 3 × 3, obtaining port number is 16,
Scale characteristic pattern big as original image；Parameter amount is consistent with resent network parameter amount, sets port number by experience
It is 16.
S2: being connected through 10 blocks as shown in Figure 2, and obtaining port number is 16, the characteristic pattern of Scale invariant.
S3: not using pond layer, and the convolution kernel for being 2 by step-length reaches the identical purpose of He Chihua；While in order to guarantee
It obtains degree enough convolution kernel port number is double, becomes 32；This step does not have convolutional Neural unit designed by the invention, adopts
With residual block identical with resnet32.
S4: by 9 convolutional Neural cell blocks as shown in Figure 2, the characteristic pattern that port number is 32 is obtained.
S5: the resnet residual block for being 2 by step-length, while port number is double again, becomes 64.
S6: by 9 convolutional Neural cell blocks as shown in Figure 2, the characteristic pattern that port number is 64 is obtained.
S7: being 8 by step-length, the pond layer that receptive field is 8 × 8.
S8: passing through full articulamentum, and classification prediction can be completed.
Convolutional neural networks of the present invention reduce 1/18th than existing resnet32 in parameter amount, but classify effect
Fruit is similar to resnet32；For example, by using convolutional Neural cell block shown in Fig. 2 compared with resnet block parameter amount, it is assumed that input
Port number is n, and parameter amount required for a resnet block is 2 × 3 × 3 × n=18n, and convolutional Neural cell block shown in Fig. 2
The parameter that first layer needs is 4 × 5 × n × 0.5+3 × 3 × n × 0.5=14.5n；Parameter required for the second layer be 1 × 1 ×
2.5n=2.5n adds together for 17n；Compared with former resent block, reduce 1/18th.
The above-mentioned convolutional neural networks that present invention design obtains are subjected to image classification；It is carried out with cifar10 data set real
It tests, on Training strategy, using stochastic gradient descent method SGD；Initial learning rate is 0.1,250 rounds of training；At the 81st
When round, learning rate is revised as 0.01；When the 121st round, learning rate is become 0.001；When the 181st round, it will learn
Habit rate becomes 0.0001.When training, momentum parameter, momentum 0.9 are set.Loss function uses cross entropy loss function.Simultaneously
In order to reduce over-fitting, data enhancing has been carried out to the picture of cifar10 data set, including random flip horizontal, level, has been hung down
Straight small translation etc..The accuracy that 93.09% is finally obtained on cifar10 data set, than using the correct of resent32
Rate improves 0.5%, and the parameter amount stored needed for reducing.
Diagonal convolution proposed by the present invention can directly acquire the correlation in former characteristic pattern diagonally opposing corner direction；Using the present invention
Neural network constructed by the convolutional Neural cell block of design promotes the adaptability of spatial alternation.It can be in different directions
Convolution kernel decomposes square convolution kernel, in the quantity for guaranteeing to reduce parameter in the case where possessing big receptive field.
Claims (5)
1. a kind of convolutional Neural cell block, which is characterized in that the convolution kernel of convolution kernel and a m × m including n different directions
The convolution kernel that heap poststack connects a 1 × 1 is constituted；It further include the logical of the jump connection of identical transformation, wherein a and input feature vector figure
Road number is equal.
2. a kind of convolutional Neural cell block according to claim 1, which is characterized in that the convolution kernel packet of the different directions
Include oblique convolution kernel b × 1 in a left side and the oblique 1 × b of convolution kernel in a right side；Left tiltedly convolution kernel be b × b convolution kernel in addition to left diagonally opposing corner it
Outer all positions all remain 0；Right tiltedly convolution kernel is that all positions of b × b convolution kernel other than right diagonally opposing corner are whole
Remain 0.
3. a kind of convolutional Neural cell block according to claim 1, which is characterized in that the convolution of the n different directions
In core, each direction includes the stacking of the direction convolution of two vertical direction.
4. a kind of neural network using convolutional Neural cell block as described in claim 1, which is characterized in that including being sequentially connected
The convolution kernel of m × m, c convolutional Neural cell block, step-length be d convolution kernel, e convolutional Neural cell block, step-length be d's
Residual block, e convolutional Neural cell block, step-length be f receptive field be f × f pond layer, full articulamentum.
5. a kind of image classification method using neural network as claimed in claim 4, which comprises the following steps:
Step 1: building neural network；
Step 2: training step 1 constructs obtained neural network；
Step 3: data enhancing processing being carried out to test set picture, is inputted in the neural network obtained after step 2 training, i.e.,
The classification of achievable image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910170897.5A CN109858618B (en) | 2019-03-07 | 2019-03-07 | Convolutional neural unit block, neural network formed by convolutional neural unit block and image classification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910170897.5A CN109858618B (en) | 2019-03-07 | 2019-03-07 | Convolutional neural unit block, neural network formed by convolutional neural unit block and image classification method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109858618A true CN109858618A (en) | 2019-06-07 |
| CN109858618B CN109858618B (en) | 2020-04-14 |
Family
ID=66900186
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910170897.5A Active CN109858618B (en) | 2019-03-07 | 2019-03-07 | Convolutional neural unit block, neural network formed by convolutional neural unit block and image classification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109858618B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796250A (en) * | 2019-10-11 | 2020-02-14 | 浪潮电子信息产业股份有限公司 | Convolution processing method and system applied to convolutional neural network and related components |
| CN112785663A (en) * | 2021-03-17 | 2021-05-11 | 西北工业大学 | Image classification network compression method based on arbitrary shape convolution kernel |
| CN114677568A (en) * | 2022-05-30 | 2022-06-28 | 山东极视角科技有限公司 | Linear target detection method, module and system based on neural network |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105374007A (en) * | 2015-12-02 | 2016-03-02 | 华侨大学 | Generation method and generation device of pencil drawing fusing skeleton strokes and textural features |
| CN105957059A (en) * | 2016-04-20 | 2016-09-21 | 广州视源电子科技股份有限公司 | Electronic component missing detection method and system |
| CN106710589A (en) * | 2016-12-28 | 2017-05-24 | 百度在线网络技术（北京）有限公司 | Artificial intelligence-based speech feature extraction method and device |
| CN107830846A (en) * | 2017-09-30 | 2018-03-23 | 杭州艾航科技有限公司 | One kind utilizes unmanned plane and convolutional neural networks measurement communication tower aerial angle method |
| CN108062551A (en) * | 2017-06-28 | 2018-05-22 | 浙江大学 | A kind of figure Feature Extraction System based on adjacency matrix, figure categorizing system and method |
| US20180174031A1 (en) * | 2016-10-10 | 2018-06-21 | Gyrfalcon Technology Inc. | Implementation Of ResNet In A CNN Based Digital Integrated Circuit |
| CN108305261A (en) * | 2017-08-11 | 2018-07-20 | 腾讯科技（深圳）有限公司 | Picture segmentation method, apparatus, storage medium and computer equipment |
| CN108520275A (en) * | 2017-06-28 | 2018-09-11 | 浙江大学 | A kind of regular system of link information based on adjacency matrix, figure Feature Extraction System, figure categorizing system and method |
| CN108701249A (en) * | 2016-01-25 | 2018-10-23 | 渊慧科技有限公司 | Image is generated using neural network |
| CN108985443A (en) * | 2018-07-04 | 2018-12-11 | 北京旷视科技有限公司 | Action identification method and its neural network generation method, device and electronic equipment |
| US20180373981A1 (en) * | 2017-06-21 | 2018-12-27 | TuSimple | Method and device for optimizing neural network |
-
2019
- 2019-03-07 CN CN201910170897.5A patent/CN109858618B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105374007A (en) * | 2015-12-02 | 2016-03-02 | 华侨大学 | Generation method and generation device of pencil drawing fusing skeleton strokes and textural features |
| CN108701249A (en) * | 2016-01-25 | 2018-10-23 | 渊慧科技有限公司 | Image is generated using neural network |
| CN105957059A (en) * | 2016-04-20 | 2016-09-21 | 广州视源电子科技股份有限公司 | Electronic component missing detection method and system |
| US20180174031A1 (en) * | 2016-10-10 | 2018-06-21 | Gyrfalcon Technology Inc. | Implementation Of ResNet In A CNN Based Digital Integrated Circuit |
| CN106710589A (en) * | 2016-12-28 | 2017-05-24 | 百度在线网络技术（北京）有限公司 | Artificial intelligence-based speech feature extraction method and device |
| US20180373981A1 (en) * | 2017-06-21 | 2018-12-27 | TuSimple | Method and device for optimizing neural network |
| CN108062551A (en) * | 2017-06-28 | 2018-05-22 | 浙江大学 | A kind of figure Feature Extraction System based on adjacency matrix, figure categorizing system and method |
| CN108520275A (en) * | 2017-06-28 | 2018-09-11 | 浙江大学 | A kind of regular system of link information based on adjacency matrix, figure Feature Extraction System, figure categorizing system and method |
| CN108305261A (en) * | 2017-08-11 | 2018-07-20 | 腾讯科技（深圳）有限公司 | Picture segmentation method, apparatus, storage medium and computer equipment |
| CN107830846A (en) * | 2017-09-30 | 2018-03-23 | 杭州艾航科技有限公司 | One kind utilizes unmanned plane and convolutional neural networks measurement communication tower aerial angle method |
| CN108985443A (en) * | 2018-07-04 | 2018-12-11 | 北京旷视科技有限公司 | Action identification method and its neural network generation method, device and electronic equipment |
Non-Patent Citations (1)
| Title |
|---|
| FRANC¸OIS CHOLLET: "Xception: Deep Learning with Depthwise Separable Convolutions", 《IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION 2017》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796250A (en) * | 2019-10-11 | 2020-02-14 | 浪潮电子信息产业股份有限公司 | Convolution processing method and system applied to convolutional neural network and related components |
| WO2021068376A1 (en) * | 2019-10-11 | 2021-04-15 | 浪潮电子信息产业股份有限公司 | Convolution processing method and system applied to convolutional neural network, and related components |
| CN112785663A (en) * | 2021-03-17 | 2021-05-11 | 西北工业大学 | Image classification network compression method based on arbitrary shape convolution kernel |
| CN112785663B (en) * | 2021-03-17 | 2024-05-10 | 西北工业大学 | Image classification network compression method based on convolution kernel of arbitrary shape |
| CN114677568A (en) * | 2022-05-30 | 2022-06-28 | 山东极视角科技有限公司 | Linear target detection method, module and system based on neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109858618B (en) | 2020-04-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108876792B (en) | Semantic segmentation method, device and system and storage medium | |
| CN109858618A (en) | The neural network and image classification method of a kind of convolutional Neural cell block, composition | |
| CN108062551A (en) | A kind of figure Feature Extraction System based on adjacency matrix, figure categorizing system and method | |
| CN108399362A (en) | A kind of rapid pedestrian detection method and device | |
| CN110136170A (en) | A kind of remote sensing image building change detecting method based on convolutional neural networks | |
| CN108520275A (en) | A kind of regular system of link information based on adjacency matrix, figure Feature Extraction System, figure categorizing system and method | |
| CN108009594B (en) | A kind of image-recognizing method based on change grouping convolution | |
| CN113159143B (en) | Infrared and visible light image fusion method and device based on jump connection convolution layer | |
| CN105981050B (en) | For extracting the method and system of face characteristic from the data of facial image | |
| CN113469074B (en) | Remote sensing image change detection method and system based on twin attention fusion network | |
| CN108416327A (en) | A kind of object detection method, device, computer equipment and readable storage medium storing program for executing | |
| CN112184554A (en) | Remote sensing image fusion method based on residual mixed expansion convolution | |
| CN110163813A (en) | A kind of image rain removing method, device, readable storage medium storing program for executing and terminal device | |
| CN107292336A (en) | A kind of Classification of Polarimetric SAR Image method based on DCGAN | |
| CN107341440A (en) | Indoor RGB D scene image recognition methods based on multitask measurement Multiple Kernel Learning | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN108596264A (en) | A kind of community discovery method based on deep learning | |
| CN109087375A (en) | Image cavity fill method based on deep learning | |
| CN113255788B (en) | Method and system for generating confrontation network face correction based on two-stage mask guidance | |
| CN112307982A (en) | Human behavior recognition method based on staggered attention-enhancing network | |
| CN113378721B (en) | Symmetrical and local discrimination-based face correction method and system for generating countermeasure | |
| CN106503659A (en) | Action identification method based on sparse coding tensor resolution | |
| CN113239870B (en) | Identity constraint-based face correction method and system for generating countermeasure network | |
| CN111882516B (en) | Image quality evaluation method based on visual saliency and deep neural network | |
| CN107679539A (en) | A kind of single convolutional neural networks local message wild based on local sensing and global information integration method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |