CN109948776A - A kind of confrontation network model picture tag generation method based on LBP - Google Patents
A kind of confrontation network model picture tag generation method based on LBP Download PDFInfo
- Publication number
- CN109948776A CN109948776A CN201910140998.8A CN201910140998A CN109948776A CN 109948776 A CN109948776 A CN 109948776A CN 201910140998 A CN201910140998 A CN 201910140998A CN 109948776 A CN109948776 A CN 109948776A
- Authority
- CN
- China
- Prior art keywords
- lbp
- picture
- texture pattern
- confrontation network
- model
- 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
Abstract
The confrontation network model picture tag generation method based on LBP that the invention discloses a kind of, the method comprises the following steps: to the picture of input, identification Objective extraction is carried out using production confrontation network model and carries out model modification, obtains initial differentiation result set.It is extracted based on texture pattern of the LBP algorithm to original image, obtains the texture pattern of picture is complete.Small noise largely isolated in texture pattern or pseudo- target are removed by calculating the coefficient of variation, obtains the picture texture pattern after denoising.Picture texture pattern after initial results collection and denoising is input to jointly in confrontation network arbiter, confrontation network texture continuity label training is carried out based on picture texture pattern, the label for obtaining detection target eliminates the initial randomness for differentiating result set well.
Description
Technical field
The present invention relates to computer vision research field, in particular to a kind of confrontation network model picture mark based on LBP
Sign generation method.
Background technique
With deepening continuously based on the field of target recognition research for fighting network and people are for network model target
The requirement of verification and measurement ratio is continuously improved, and the label generating method for fighting network is also increasingly subject to the attention of people.Fight the one of network
Typical realize of kind is constructed using tired fold of depth convolutional layer, and the semantic information of picture will necessarily be lost in the convolution process of arbiter,
And the arbiter based on convolutional layer can not resist to resisting sample attack, mainly model do not have rotational invariance, how structure
The arbiter model for making a strong robustness becomes the hot spot of research.
It is the important component for fighting network model that picture tag, which generates, directly affects confrontation network to picture target
The accuracy rate of identification.Currently, it is to generate letter that generating function, which influences confrontation one of them critically important reason of Network Recognition accuracy rate,
Exponential model does not have the characteristics of rotational invariance, can not be added according to the texture features of object in picture from what main modulation label generate
Weight coefficient, it is therefore desirable to which finding one kind being capable of a kind of confrontation network model picture tag generation method based on LBP.
LBP algorithm due to texture blending effect is good and have the advantages that rotational invariance and become people research coke
Point, the prior art are influenced very big by factors such as picture rotation and background changings, are caused during picture tag generates
Confrontation network model recognition accuracy is affected.
Summary of the invention
The shortcomings that it is a primary object of the present invention to overcome existing confrontation web tab generation technique and deficiency, provide one kind
Confrontation network model picture tag generation method based on LBP can effectively improve confrontation network model recognition accuracy.
In order to achieve the above object, the invention adopts the following technical scheme:
A kind of confrontation network model picture tag generation method based on LBP of the present invention, specifically includes that steps are as follows:
S1, the picture to input carry out identification Objective extraction using production confrontation network model, and carry out model more
Newly, initial differentiation result set is obtained;
S2, it is extracted based on texture pattern of the LBP algorithm to original image, obtains the texture pattern of picture is complete;
S3, small noise largely isolated in texture pattern or pseudo- target are removed by calculating the coefficient of variation, obtains denoising
Picture texture pattern after sound;
S4, the picture texture pattern after initial results collection and denoising is input to jointly in confrontation network arbiter, is based on
Picture texture pattern carries out confrontation network texture continuity label training, obtains the label of detection target, eliminates well just
The randomness for the differentiation result set that begins.
The step S1 as a preferred technical solution, specifically:
S11, construction Wo Sesitan production fight network WGAN model, and the WGAN model includes image composer and figure
As arbiter, and the loss function using intersection entropy function as both sides;
S12, network WGAN model is fought to the production constructed in S11, generator is by encoder and decoder two parts
It constitutes;
S13, network model generator is fought to the generator constructed in S12, encoder is made of eight layers of convolutional layer;
S14, network model generator is fought to the generator constructed in S12, decoder is made of eight layers of warp lamination;
S15, network model generator is fought to the generator constructed in S12, the output of encoder will continue to execute all the way
In addition the process of convolution is merged into decoder in the input of corresponding warp lamination by Concatenate operation all the way;
S16, network WGAN model is fought to the production constructed in S11, arbiter is made of four layers of convolutional layer;
S17, network WGAN model is fought to the production constructed in S11, to fixed generator G, optimal arbiterExpression formula is as follows:
Wherein Pdata(x) it is distributed for truthful data, PgIt (x) is generation data distribution;
S18, input raw image data collection are trained the production confrontation network WGAN model of step S11 building,
Obtain initial differentiation result set.
The step S2 as a preferred technical solution, specifically:
Assuming that pixel gray value is g at the c of picture target areac, corresponding eight neighborhood territory pixels point gray value sample point
Respectively gp, the pixel of each neighborhood and the point carry out binaryzation and compared by p=0 ..., P, obtain the two of one eight into
System string;Picture More General Form U (LBP is obtained by neighborhood comparisonP,R), if U (LBPP,R) < T then judges that the pixel belongs to inspection
It surveys target and its texture pattern is LBP (gc), otherwise the pixel belongs to background area, texture pattern P+1, wherein T
For the mode threshold being manually set according to scene, point g centered on PcSurrounding sample point number, R be sample point and central point away from
From, specific formula is as follows:
Represent gray value g at central pixel point ccWith its surrounding pixel point gray value gp, p=0 ..., p-shaped
At textural characteristics.
Described the step of picture More General Form is obtained by neighborhood comparison as a preferred technical solution, specifically: for
Neighborhood compares the circulation binary string to be formed, and calculating the signal transition times in string is the More General Form that pixel is formed, U
(LBPP,R) it is defined as follows:
Wherein g0For gcIt is positive right
The neighborhood territory pixel point of side, R and P are respectively sample point away from central point gcRadius and sample point number.
It is described to remove a large amount of orphans in texture pattern by calculating the coefficient of variation as a preferred technical solution, in step S3
The step of vertical small noise or pseudo- target specifically:
Assuming that texture blending result set is Rlbp;Noise model Poisson is established using the component that current texture extracts result set
It is distributed and counts its histogram Hist [Rlbp], to resulting histogram calculation correlation variance Var [R];Finally to resulting correlation
Variance is ranked up, and searches out maximum value, as coefficient of variation λ;Then determine to differentiate result set hair when coefficient of variation λ is greater than 1
Raw mutation;
λ=max (Var [Rlbp])/C
Wherein RlbpFor initial area judging result set, Var [] is related variance, and C is constant.
As a preferred technical solution, in step S3, statistic texture extracts result set histogram Hist [Rlbp] formula is such as
Under:
Wherein H, W are respectively the height and width of picture;
It is as follows to calculate resulting histogram correlation variance Var [R] formula:
Var[Rlbp]=E { [Xlbp-E(Xlbp)]2}
=E (Xlbp 2)-(E(Xlbp))2。
As a preferred technical solution, in step 4, confrontation network texture continuity label is carried out based on picture texture pattern
Training, the process for obtaining detection target labels are as follows:
Assuming that having extracted and being by the texture pattern of denoisingTruthful data be distributed as y~
pdata, then the input for fighting network arbiter is D (x, y), and the input for fighting network generator is G (x, z), fights the damage of network
It is as follows to lose function:
Wherein G is used for the generator function of descriptive model, and D is used for descriptive model arbiter function,With
Come describe confrontation network model loss function, ifThen the training of texture continuity label is completed, wherein ε
For the error coefficient of artificial settings.
Compared with the prior art, the invention has the following advantages and beneficial effects:
The present invention proposes to carry out identification Objective extraction using production confrontation network model to the picture of input and carries out mould
Type updates, and obtains initial differentiation result set.In conjunction with LBP textural characteristics to the insensitivity of rotation, to the texture mould of original image
Formula extracts, and obtains the texture pattern of picture is complete.Again by calculate the coefficient of variation remove in texture pattern it is largely isolated
Small noise or pseudo- target, obtain the picture texture pattern data after denoising.It finally will be after initial results collection and denoising
Picture texture pattern is input to jointly in confrontation network arbiter, carries out confrontation network texture continuity based on picture texture pattern
Label training, obtains the label of detection target.This method is efficiently against picture semantic caused by confrontation network model convolutional layer
Distortion eliminates the initial randomness for differentiating result set well.Due to the processes operation such as local binary patterns and coefficient of variation
Quickly, so the algorithm is with good performance simultaneously.
Detailed description of the invention
Fig. 1 is the flow chart generated for picture tag of the invention.
Fig. 2 is production confrontation network model overall structure figure of the invention.
Fig. 3 is production confrontation network generator and arbiter structure chart of the invention.
Fig. 4 (a)-Fig. 4 (d) is the lab diagram of confrontation network model of the invention.
Specific embodiment
Present invention will now be described in further detail with reference to the embodiments and the accompanying drawings, but embodiments of the present invention are unlimited
In this.
Embodiment
A kind of confrontation network model picture tag generation method based on LBP that the present invention develops, is in microcomputer
Under 16.04 environment of Ubuntu, using object-oriented design method and Software engineering standard, realized with Python.
Fig. 1 is the specific flow chart of the method for the invention.Illustrate some specific realities in the present invention by taking Fig. 1 as an example below
Existing process.Method of the invention is a kind of confrontation network model picture tag generation method based on LBP, the specific steps are that:
S1 carries out identification Objective extraction using production confrontation network model and carries out model modification to the picture of input,
Obtain initial differentiation result set;The specific method that step S1 is used for
S11, construction Wo Sesitan production fight network WGAN model, and model includes image composer and image discriminating
Device, and the loss function using intersection entropy function as both sides;Fig. 2 is the whole knot that constructed production fights network model
Composition, Fig. 3 are that constructed production fights the generator of network model and the specific structure of arbiter.
S12, network WGAN model is fought to the production constructed in S11, generator is by encoder and decoder two parts
It constitutes.
S13, network model generator being fought to the generator constructed in S12, encoder is made of eight layers of convolutional layer,
Structure is C64-C128-C256-C512-C512-C512-C512-C512.
S14, network model generator being fought to the generator constructed in S12, decoder is made of eight layers of warp lamination,
Its structure is CD512-CD512-CD512-CD512-CD256-CD128-CD64-CD3.
S15, network model generator is fought to the generator constructed in S12, the output of encoder will continue to execute all the way
In addition the process of convolution is merged into decoder in the input of corresponding warp lamination by Concatenate operation all the way.
S16, network WGAN model is fought to the production constructed in S11, arbiter is made of four layers of convolutional layer, structure
For C64-C128-C256-C512.
S17, network WGAN model is fought to the production constructed in S11, to fixed generator G, optimal arbiterExpression formula is as follows:
Wherein Pdata(x) it is distributed for truthful data, PgIt (x) is generation data distribution;
S18, input raw image data collection are trained the production confrontation network WGAN model of step S11 building,
Obtain initial differentiation result set.
S2, it is extracted based on texture pattern of the LBP algorithm to original image, obtains the texture pattern of picture is complete,
In extracted based on texture pattern of the LBP algorithm to original image, specifically:
Assuming that pixel gray value is g at the c of picture target areac, corresponding eight neighborhood territory pixels point gray value sample point
Respectively gp, the pixel of each neighborhood and the point carry out binaryzation and compared by p=0 ..., P, obtain the two of one eight into
System string;Picture More General Form U (LBP is obtained by neighborhood comparisonP,R), if U (LBPP,R) < T then judges that the pixel belongs to inspection
It surveys target and its texture pattern is LBP (gc), otherwise the pixel belongs to background area, texture pattern P+1, wherein T
For the mode threshold being manually set according to scene.Point g centered on PcSurrounding sample point number, R be sample point and central point away from
From.
Wherein,Represent gray value g at central pixel point ccWith its surrounding pixel point gray value gp, p=
0 ..., p-shaped at textural characteristics.
S3, small noise largely isolated in texture pattern or pseudo- target are removed by calculating the coefficient of variation, obtains denoising
Picture texture pattern after sound, specific steps are as follows:
Assuming that texture blending result set is Rlbp;Noise model Poisson is established using the component that current texture extracts result set
It is distributed and counts its histogram Hist [Rlbp], to resulting histogram calculation correlation variance Var [R];Finally to resulting correlation
Variance is ranked up, and searches out maximum value, as coefficient of variation λ;Then determine to differentiate result set hair when coefficient of variation λ is greater than 1
Raw mutation;
λ=max (Var [Rlbp])/C
Wherein RlbpFor initial area judging result set, Var [] is related variance, and C is constant
Statistic texture extracts result set histogram Hist [Rlbp] formula is as follows
Wherein H, W are respectively the height and width of picture
It is as follows to calculate resulting histogram correlation variance Var [R] formula
Var[Rlbp]=E { [Xlbp-E(Xlbp)]2}
=E (Xlbp 2)-(E(Xlbp))2。
S4, the picture texture pattern after initial results collection and denoising is input to jointly in confrontation network arbiter, is based on
Picture texture pattern carries out confrontation network texture continuity label training, obtains the label of detection target, eliminates well just
Begin to differentiate that the randomness of result set obtains wherein carrying out confrontation network texture continuity label training based on picture texture pattern
Target labels are detected, specifically:
S41, original image and picture texture pattern are input to jointly in confrontation network arbiter, carry out confrontation network
Texture continuity label training;
S42, S41 confrontation network texture continuity label is trained, fights network losses function such as following formula:
Wherein D is arbiter network approximating function, and G is generator network approximating function,Y~
pdata, p hereindataFor original image;
S43, to confrontation network losses function, fight network training loss function final goal formula:
For the purpose of be that the training objective that both makes is opposite.Arbiter D is optimized to given generator G, most
Bigization log-likelihood function to differentiate the source of G (z) He x, i.e. arbiter D can correct decision generate picture G (z) and true figure
Piece x;On the contrary, optimize generator G to given arbiter D, minimize log-likelihood function so that G (z) picture approaching to reality
The picture of the distribution of picture x, i.e. generator output can mix the spurious with the genuine.Fig. 4 (b) and Fig. 4 (d) is original image, and Fig. 4 (a) is
The picture tag that network model generates is fought, Fig. 4 (c) is the picture tag that the present invention generates.It can be seen that from Fig. 4 (a), (c)
The method of the present invention can eliminate picture noise and pseudo- target generates model label in conjunction with LBP texture pattern binaryzation characteristic
Accuracy rate influence.In addition, since the processes operation such as local binary patterns and coefficient of variation is quick, so the algorithm has simultaneously
There is good performance.
The above embodiment is a preferred embodiment of the present invention, but embodiments of the present invention are not by above-described embodiment
Limitation, other any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present invention,
It should be equivalent substitute mode, be included within the scope of the present invention.
Claims (7)
1. a kind of confrontation network model picture tag generation method based on LBP, which is characterized in that specifically include that steps are as follows:
S1, the picture to input carry out identification Objective extraction using production confrontation network model, and carry out model modification, obtain
Take initial differentiation result set;
S2, it is extracted based on texture pattern of the LBP algorithm to original image, obtains the texture pattern of picture is complete;
S3, small noise largely isolated in texture pattern or pseudo- target are removed by calculating the coefficient of variation, obtain denoising it
Picture texture pattern afterwards;
S4, the picture texture pattern after initial results collection and denoising is input to jointly in confrontation network arbiter, is based on picture
Texture pattern carries out confrontation network texture continuity label training, obtains the label of detection target, eliminates initially sentence well
The randomness of other result set.
2. the confrontation network model picture tag generation method based on LBP according to claim 1, which is characterized in that described
Step S1 specifically:
S11, construction Wo Sesitan production fight network WGAN model, and the WGAN model includes that image composer and image are sentenced
Other device, and the loss function using intersection entropy function as both sides;
S12, network WGAN model is fought to the production constructed in S11, generator is made of encoder and decoder two parts;
S13, network model generator is fought to the generator constructed in S12, encoder is made of eight layers of convolutional layer;
S14, network model generator is fought to the generator constructed in S12, decoder is made of eight layers of warp lamination;
S15, network model generator is fought to the generator constructed in S12, the output of encoder will continue to execute convolution all the way
Process, in addition all the way by Concatenate operation be merged into decoder in the input of corresponding warp lamination;
S16, network WGAN model is fought to the production constructed in S11, arbiter is made of four layers of convolutional layer;
S17, network WGAN model is fought to the production constructed in S11, to fixed generator G, optimal arbiterTable
It is as follows up to formula:
Wherein Pdata(x) it is distributed for truthful data, PgIt (x) is generation data distribution;
S18, input raw image data collection are trained the production confrontation network WGAN model of step S11 building, obtain
It is initial to differentiate result set.
3. the confrontation network model picture tag generation method based on LBP according to claim 1, which is characterized in that described
Step S2 specifically:
Assuming that pixel gray value is g at the c of picture target areac, corresponding eight neighborhood territory pixels point gray value sample point is respectively
gp, the pixel of each neighborhood and the point are carried out binaryzation and compared, obtain one eight binary strings by p=0 ..., P;
Picture More General Form U (LBP is obtained by neighborhood comparisonP,R), if U (LBPP,R) < T then judges that the pixel belongs to detection mesh
It marks and its texture pattern is LBP (gc), otherwise the pixel belongs to background area, and texture pattern P+1, wherein T is root
According to the mode threshold that scene is manually set, point g centered on PcSurrounding sample point number, R are sample point at a distance from central point, tool
Body formula is as follows:
Represent gray value g at central pixel point ccWith its surrounding pixel point gray value gp, p=0 ..., p-shaped at
Textural characteristics.
4. the confrontation network model picture tag generation method based on LBP according to claim 3, which is characterized in that described
The step of obtaining picture More General Form by neighborhood comparison specifically: compare the circulation binary string to be formed for neighborhood, calculate
Signal transition times in string are the More General Form that pixel is formed, U (LBPP,R) it is defined as follows:
Wherein g0For gcThe neighborhood territory pixel point of front-right, R and P are respectively sample point away from central point gcRadius and sample point
Number.
5. the confrontation network model picture tag generation method based on LBP according to claim 1, which is characterized in that step
It is described specific by calculating the step of coefficient of variation removes the small noise largely isolated in texture pattern or pseudo- target in S3
Are as follows:
Assuming that texture blending result set is Rlbp;Noise model Poisson distribution is established using the component that current texture extracts result set
And count its histogram Hist [Rlbp], to resulting histogram calculation correlation variance Var [R];Finally to resulting related variance
It is ranked up, searches out maximum value, as coefficient of variation λ;Then determine to differentiate that result set occurs to dash forward when coefficient of variation λ is greater than 1
Become;
λ=max (Var [Rlbp])/C
Wherein RlbpFor initial area judging result set, Var [] is related variance, and C is constant.
6. the confrontation network model picture tag generation method based on LBP according to claim 5, which is characterized in that step
In S3, statistic texture extracts result set histogram Hist [Rlbp] formula is as follows:
Wherein H, W are respectively the height and width of picture;
It is as follows to calculate resulting histogram correlation variance Var [R] formula:
Var[Rlbp]=E { [Xlbp-E(Xlbp)]2}
=E (Xlbp 2)-(E(Xlbp))2。
7. the confrontation network model picture tag generation method based on LBP according to claim 1, which is characterized in that step 4
In, confrontation network texture continuity label training is carried out based on picture texture pattern, the process for obtaining detection target labels is as follows:
Assuming that having extracted and being by the texture pattern of denoisingTruthful data is distributed as y~pdata, then
The input for fighting network arbiter is D (x, y), and the input for fighting network generator is G (x, z), fights the loss function of network
It is as follows:
Wherein G is used for the generator function of descriptive model, and D is used for descriptive model arbiter function,For describing
The loss function of network model is fought, ifThen the training of texture continuity label is completed, and wherein ε is artificial
The error coefficient of setting.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910140998.8A CN109948776A (en) | 2019-02-26 | 2019-02-26 | A kind of confrontation network model picture tag generation method based on LBP |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910140998.8A CN109948776A (en) | 2019-02-26 | 2019-02-26 | A kind of confrontation network model picture tag generation method based on LBP |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109948776A true CN109948776A (en) | 2019-06-28 |
Family
ID=67006976
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910140998.8A Pending CN109948776A (en) | 2019-02-26 | 2019-02-26 | A kind of confrontation network model picture tag generation method based on LBP |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109948776A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110401488A (en) * | 2019-07-12 | 2019-11-01 | 北京邮电大学 | A kind of demodulation method and device |
CN110717960A (en) * | 2019-10-22 | 2020-01-21 | 北京建筑大学 | Method for generating building rubbish remote sensing image sample |
CN111798409A (en) * | 2020-05-19 | 2020-10-20 | 佛山市南海区广工大数控装备协同创新研究院 | Deep learning-based PCB defect data generation method |
CN112801297A (en) * | 2021-01-20 | 2021-05-14 | 哈尔滨工业大学 | Machine learning model adversity sample generation method based on conditional variation self-encoder |
CN112818159A (en) * | 2021-02-24 | 2021-05-18 | 上海交通大学 | Image description text generation method based on generation countermeasure network |
WO2022120532A1 (en) * | 2020-12-07 | 2022-06-16 | Huawei Technologies Co., Ltd. | Presentation attack detection |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106204586A (en) * | 2016-07-08 | 2016-12-07 | 华南农业大学 | A kind of based on the moving target detecting method under the complex scene followed the tracks of |
CN108520503A (en) * | 2018-04-13 | 2018-09-11 | 湘潭大学 | A method of based on self-encoding encoder and generating confrontation network restoration face Incomplete image |
CN108564611A (en) * | 2018-03-09 | 2018-09-21 | 天津大学 | A kind of monocular image depth estimation method generating confrontation network based on condition |
CN108764005A (en) * | 2018-01-31 | 2018-11-06 | 华侨大学 | A kind of high-spectrum remote sensing atural object space Spectral Characteristic extracting method and system |
CN109166126A (en) * | 2018-08-13 | 2019-01-08 | 苏州比格威医疗科技有限公司 | A method of paint crackle is divided on ICGA image based on condition production confrontation network |
-
2019
- 2019-02-26 CN CN201910140998.8A patent/CN109948776A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106204586A (en) * | 2016-07-08 | 2016-12-07 | 华南农业大学 | A kind of based on the moving target detecting method under the complex scene followed the tracks of |
CN108764005A (en) * | 2018-01-31 | 2018-11-06 | 华侨大学 | A kind of high-spectrum remote sensing atural object space Spectral Characteristic extracting method and system |
CN108564611A (en) * | 2018-03-09 | 2018-09-21 | 天津大学 | A kind of monocular image depth estimation method generating confrontation network based on condition |
CN108520503A (en) * | 2018-04-13 | 2018-09-11 | 湘潭大学 | A method of based on self-encoding encoder and generating confrontation network restoration face Incomplete image |
CN109166126A (en) * | 2018-08-13 | 2019-01-08 | 苏州比格威医疗科技有限公司 | A method of paint crackle is divided on ICGA image based on condition production confrontation network |
Non-Patent Citations (2)
Title |
---|
TIMO OJALA 等: ""Multiresolution Gray-Scale and Rotation Invariant Texture Classification with Local Binary Patterns"", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
余思泉 等: ""基于对抗生成网络的纹理合成方法"", 《红外与激光工程》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110401488A (en) * | 2019-07-12 | 2019-11-01 | 北京邮电大学 | A kind of demodulation method and device |
CN110401488B (en) * | 2019-07-12 | 2021-02-05 | 北京邮电大学 | Demodulation method and device |
CN110717960A (en) * | 2019-10-22 | 2020-01-21 | 北京建筑大学 | Method for generating building rubbish remote sensing image sample |
CN111798409A (en) * | 2020-05-19 | 2020-10-20 | 佛山市南海区广工大数控装备协同创新研究院 | Deep learning-based PCB defect data generation method |
WO2022120532A1 (en) * | 2020-12-07 | 2022-06-16 | Huawei Technologies Co., Ltd. | Presentation attack detection |
CN112801297A (en) * | 2021-01-20 | 2021-05-14 | 哈尔滨工业大学 | Machine learning model adversity sample generation method based on conditional variation self-encoder |
CN112801297B (en) * | 2021-01-20 | 2021-11-16 | 哈尔滨工业大学 | Machine learning model adversity sample generation method based on conditional variation self-encoder |
CN112818159A (en) * | 2021-02-24 | 2021-05-18 | 上海交通大学 | Image description text generation method based on generation countermeasure network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109948776A (en) | A kind of confrontation network model picture tag generation method based on LBP | |
Tu et al. | Edge-guided non-local fully convolutional network for salient object detection | |
CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
CN106997597B (en) | It is a kind of based on have supervision conspicuousness detection method for tracking target | |
CN108346159A (en) | A kind of visual target tracking method based on tracking-study-detection | |
CN108520530A (en) | Method for tracking target based on long memory network in short-term | |
CN111242027B (en) | Unsupervised learning scene feature rapid extraction method fusing semantic information | |
CN103996018A (en) | Human-face identification method based on 4DLBP | |
CN103413120A (en) | Tracking method based on integral and partial recognition of object | |
CN107564035B (en) | Video tracking method based on important area identification and matching | |
CN109598684A (en) | In conjunction with the correlation filtering tracking of twin network | |
CN108268823A (en) | Target recognition methods and device again | |
CN110569855A (en) | Long-time target tracking algorithm based on correlation filtering and feature point matching fusion | |
CN102945553A (en) | Remote sensing image partition method based on automatic difference clustering algorithm | |
CN106203373B (en) | A kind of human face in-vivo detection method based on deep vision bag of words | |
CN110059730A (en) | A kind of thyroid nodule ultrasound image classification method based on capsule network | |
CN109409227A (en) | A kind of finger vena plot quality appraisal procedure and its device based on multichannel CNN | |
CN105373810A (en) | Method and system for building action recognition model | |
CN108830882A (en) | Video abnormal behaviour real-time detection method | |
CN106127766A (en) | Based on Space Coupling relation and the method for tracking target of historical models | |
CN108846845B (en) | SAR image segmentation method based on thumbnail and hierarchical fuzzy clustering | |
CN109902581A (en) | It is a kind of based on multistep weighting single sample portion block face identification method | |
CN113033345A (en) | V2V video face recognition method based on public feature subspace | |
CN104517300A (en) | Vision judgment tracking method based on statistical characteristic | |
Qin et al. | Multi-scaling detection of singular points based on fully convolutional networks in fingerprint images |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190628 |