CN107122800A - A kind of Robust digital figure mask method based on the screening that predicts the outcome - Google Patents
A kind of Robust digital figure mask method based on the screening that predicts the outcome Download PDFInfo
- Publication number
- CN107122800A CN107122800A CN201710298619.9A CN201710298619A CN107122800A CN 107122800 A CN107122800 A CN 107122800A CN 201710298619 A CN201710298619 A CN 201710298619A CN 107122800 A CN107122800 A CN 107122800A
- Authority
- CN
- China
- Prior art keywords
- mrow
- outcome
- predicts
- msub
- screening
- 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
- 238000012216 screening Methods 0.000 title claims abstract description 11
- 238000000034 method Methods 0.000 title claims description 33
- 239000003550 marker Substances 0.000 claims abstract description 16
- 230000006870 function Effects 0.000 claims abstract description 12
- 238000010801 machine learning Methods 0.000 claims abstract description 9
- 238000012549 training Methods 0.000 claims description 3
- 241001269238 Data Species 0.000 claims description 2
- 238000012545 processing Methods 0.000 claims description 2
- 230000009467 reduction Effects 0.000 claims description 2
- 239000006185 dispersion Substances 0.000 claims 1
- 230000004069 differentiation Effects 0.000 abstract 1
- 238000002372 labelling Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000015556 catabolic process Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 230000035699 permeability Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012795 verification 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
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of robust machine learning method based on the screening that predicts the outcome, the digital picture annotation results more reliable for obtaining.Specifically, the present invention is using the classical thought in machine learning --- maximize interval principle, predicting the outcome for being obtained to digital picture to be marked under a variety of measuring similarities is screened, the result for being wherein spaced maximum is chosen as the final output that predicts the outcome, the mark to digital picture is completed.Predicting the outcome has a case that large-spacing avoids to predict the outcome in theory and is difficult to differentiate between, with good robustness.For explicitly counting period, the present invention weighs the differentiation degree predicted the outcome using machine learning classics loss function, so that the size being spaced.Wherein, loss function refers to the gap between the predictive marker (centrifugal pump) of predict the outcome (successive value) and candidate, and the loss is smaller, and to represent the interval predicted the outcome bigger.
Description
Technical field
The present invention relates to a kind of Robust digital figure mask method based on the screening that predicts the outcome, belong to machine learning techniques
Field.
Background technology
With all kinds of social network sites be widely current and digital product a large amount of popularizations, have the number of magnanimity all the time
Word image is produced and propagated.Related service is provided in such large-scale view data, it is also most difficult that one most crucial
Task be to allow the semanteme of computer automatic understanding image, and image labeling is then key technology therein.Automatic image annotation
Task be the visual signature based on digital picture to predict its semantic marker.Specifically, extracted first from digital picture
Visual signature represents these example images, is then based on these character representations, from the view data with semantic marker
A marking model is trained in set.After the corresponding character representation input marking model of the digital picture of mark to be predicted,
Model can just make prediction to their semantic marker.Traditional digital picture label technology is often semantic merely with having
The view data of mark, i.e., by the way of supervised learning.But substantial amounts of raw image data is there is on internet, these
Digital picture is very easy to obtain but itself does not have semantic marker (i.e. Unlabeled data).How to effectively utilize largely " just
It is suitable " Unlabeled data helps to improve the performance of marking model, with very strong realistic meaning, it has also become machine learning field
An important subject, this kind of method is referred to as semi-supervised learning method.
Although being marked for digital picture for task, it usually can be obtained using semi-supervised learning method than being learned using supervision
Learning method preferably marks performance, but still exist performance not robust the problem of, or even in some cases be not so good as using supervision
The performance of learning method.Specifically, for the data characteristics of digital picture, image labeling task generally uses similar close original
Reason --- similar instances have similar mark, build semi-supervised learning method.The key of this kind of method is portray example similar
Property.Therefore, the performance of learning method depend heavilys on the measuring similarity of example.Although existing many researchers propose respectively
How the measure of kind of various kinds, but in actual task, build reliable measuring similarity so that semi-supervised learning method
Performance is without prejudice, the problem of being still an opening.It is not any so far to grind for digital picture mark task
Study carefully achievement and show that a certain measuring similarity is particularly suitable for this generic task, this just result in user in actual applications and is supervised using half
Superintend and direct mode of learning to train the select permeability faced during marking model on many measuring similarities, if selection is improper, it will make
The mark ability of gained marking model is then caused damage by extreme influence to user, especially when the mark of mistake can be brought
During serious consequence, such as medical imaging mark task dispatching, then influence more very.Therefore, being badly in need of a kind of learning method of robust is used to count
Word image labeling task.
The content of the invention
Goal of the invention：When being used for digital picture mark task for current semi-supervised learning method, there is performance degradation
Problem, it is more reliable for obtaining the invention provides a kind of Robust digital figure mask method based on the screening that predicts the outcome
Digital picture annotation results.Specifically, the present invention is using the classical thought in machine learning --- interval thought is maximized, it is right
What digital picture to be marked was obtained under a variety of measuring similarities, which predict the outcome, is screened, and chooses the knot for being wherein spaced maximum
Fruit is as the final output that predicts the outcome, to complete the mark to digital picture.Predict the outcome with large-spacing to a certain degree
On avoid to predict the outcome and cannot distinguish between so that with higher reliability.In order to portray interval, the present invention uses loss function
Mode weigh the gap size predicted the outcome.Loss function is used to calculate predicted value (successive value) and corresponding predictive marker
Gap between (centrifugal pump), the loss is smaller, and to represent the interval that this group predicts the outcome bigger.
Technical scheme：A kind of Robust digital figure mask method based on the screening that predicts the outcome, it is main to include with next system
Row step：
(1) by using semi-supervised learning method and using a variety of measuring similarities, many of current image to be marked are obtained
Organize predicting the outcome for related semantic marker；
(2) to every group of predicted value, its gap size is calculated using classical loss function；
(3) one group of predicted value of loss reduction is chosen as finally predicting the outcome, and corresponding digital picture is labeled.
Brief description of the drawings
Fig. 1 is the workflow diagram of digital picture mark；
Fig. 2 is the workflow diagram of Robust Learning method；
Fig. 3 is to maximize the workflow diagram that result screening is predicted under the principle of interval.
Embodiment
With reference to specific embodiment, the present invention is furture elucidated, it should be understood that these embodiments are merely to illustrate the present invention
Rather than limitation the scope of the present invention, after the present invention has been read, various equivalences of the those skilled in the art to the present invention
The modification of form falls within the application appended claims limited range.
As shown in figure 1, the workflow diagram of digital picture automatic marking.Step 1 is to start action.Step 2, which is inputted, to be used for
The digital image set of training, wherein both including the digital image data with semantic marker, it is assumed that a shared l is individual, also includes
A large amount of unmarked digital image datas to be marked, it is assumed that a shared u is individual.In step 3, first to the image in data acquisition system
Feature is extracted, each image is represented by a characteristic vector, i.e. xi, i=1,2 ..., l, l+1 ..., l+u.Feature extraction
Method can use the method for being applied to generation view data feature in machine learning classical textbook, for example, first carry out image point
Cut, then color, Texture eigenvalue are extracted to each segmentation block；Then, it is with markd example images generative semantics marker bit
yi∈ { ± 1 }, i=1,2 ..., l, wherein, yi=+1 represents that corresponding view data carries positive related semantic marker, on the contrary
Then with negative related semantic marker.In following step 4, using this training data set, pass through present invention proposition
Mechanism, realize the learning method of robust, prediction be marked to the Unlabeled data in data acquisition system, step 5 is finally completed
The automatic marking work of shown digital picture.
Fig. 2 is the specific workflow of step 4 in Fig. 1.By being performed step 7, in step 8, semi-supervised is obtained
Data acquisition system needed for practisingStep 9 builds multiple measuring similarity G according to existing all examples1,
G2,...,Gm, the construction method of measuring similarity can be using the classical way in the association area textbooks such as machine learning, such as
Measure based on Euclidean distance or the measure based on COS distance etc..Measurement quantity m can consider calculating cost,
The factors such as data scale are specifically set, and method for measuring similarity covering common method used is caused under normal circumstances.
In step 10, using obtained multiple measuring similarities as input, semi-supervised learning method is performed, you can obtain multigroup pre-
Survey resultK=1,2 ..., m.Semi-supervised learning method can be using classical way such as half herein
Supervise SVMs, figure semi-supervised learning method etc..Because measuring similarity has a strong impact on the performance of semi-supervised learning method,
But the quality of measuring similarity can not be learnt in advance during task in actual applications, therefore, for from multiple similar
Degree measurement obtain it is multigroup predict the outcome, step 11 using maximize interval strategy screened to predicting the outcome, so as to obtain
Reliably predict the outcome.The continuous carry out discretization that predicts the outcome that step 12 is obtained to step 11, generates final mark.Specifically
For, can be to the f that predicts the outcome that selects in the case of having knowable to classification ratiokAccording to the positive and negative mark of classification ratio cut partition
Note.In the case that classification ratio is unknowable, positive and negative mark can be carried out according to formula (1) and is assigned.
Fig. 3 is the specific workflow of step 11 in Fig. 2.Step 14 is to start action.It is pre- that step 15 inputs multigroup candidate
Survey resultK=1,2 ..., m, the final prediction with largest interval is selected for subsequent step
As a result.The corresponding part of Unlabeled data in every group of predicted value of step 16 pairK=1,
2 ..., m carries out standardization processing according to formula (2), and wherein U={ l+1, l+2 ..., l+u } represents the subscript collection of Unlabeled data
Close.
Here standardized operation is necessary, because only that ensureing for the prediction knot on each group Unlabeled data that compares
Fruit is in same dimension, and gap size just has comparativity.The new Unlabeled data of m groups is obtained after standardization to predict the outcome zk=
[zK, 1, zK, 2..., zK, u], zk=1,2 ..., m is met Step 17
Choose classical large-spacing loss functionCalculate the loss that each group newly predicts the outcomeHere lose
It is smaller, it is spaced bigger.Conventional classics loss function can be using following several：
Loss function -1：
Loss function -2：
Loss function -3：
Loss function -4：
Loss function -5：
Wherein,It is that the mark estimated result that discretization is obtained, method are carried out according to accordingly predicting the outcome
With step 12.Then as shown in step 18, the predicted value f of correspondence least disadvantage is found outk, whereinHave most
The z of small losskCorresponding index, by fkIt is used as the final output that predicts the outcome.Select optimum prediction result fkAnd after exporting, it is whole
Individual workflow ends at step 19.In order to further illustrate the validity of technology, table 1 gives the reality in image labeling data
Test result verification.
The present invention is verified using image labeling reference data set pair effectiveness of the invention.Experiment is common using 5
Mark classification.For each mark classification, experiment stochastical sampling in all samples carries out the present invention and benchmark measure of supervision
Contrast experiment.Experimental result reports 50 random niceties of grading (means standard deviation) for repeating experiment.
The nicety of grading result of table 1 (numerical value is bigger, and performance is more excellent)
Claims (4)
1. a kind of Robust digital figure mask method based on the screening that predicts the outcome, it is characterised in that mainly include the following steps that：
(1) a variety of measuring similarities are used, related the multigroup of semantic marker of current image to be marked is obtained and predicts the outcome；
(2) every group is predicted the outcome, its gap size is calculated using loss function；
(3) one group of selection loss reduction (interval is maximum) is labeled as finally predicting the outcome to corresponding digital picture.
2. the Robust digital figure mask method as claimed in claim 1 based on the screening that predicts the outcome, it is characterised in that numeral
The step of automatic image annotation is：
The digital image set for training is inputted, wherein both including the digital image data with semantic marker, it is assumed that altogether
There are l, also comprising a large amount of unmarked digital image datas to be marked, it is assumed that a shared u is individual；First to the figure in data acquisition system
As extracting feature, each image is represented by a characteristic vector, i.e. xi, i=1,2 ..., l, l+1 ..., l+u；Then, it is
With markd example images generative semantics marker bit yi∈ { ± 1 }, i=1,2 ..., l, wherein, yi=+1 represents corresponding figure
As data carry positive related semantic marker, it is on the contrary then with bear related semantic marker；To the unmarked number in data acquisition system
According to prediction is marked.
3. the Robust digital figure mask method as claimed in claim 2 based on the screening that predicts the outcome, it is characterised in that logarithm
The step of prediction is marked according to the Unlabeled data in set be：
Data needed for obtaining machine learning algorithm are representedBuild multiple measuring similarity G1,
G2,...,Gm, using obtained multiple measuring similarities as input, perform semi-supervised learning method, you can obtain multigroup pre-
Survey resultK=1,2 ..., m, to from multiple measuring similarities obtain it is multigroup predict the outcome,
Screened using interval strategy is maximized to predicting the outcome, so as to reliably be predicted the outcome；To predict the outcome carry out from
Dispersion, generates final mark.
4. the Robust digital figure mask method as claimed in claim 3 based on the screening that predicts the outcome, it is characterised in that input
Multigroup candidate prediction resultK=1,2 ..., m, to Unlabeled data correspondence in every group of predicted value
PartK=1,2 ..., m according to formula (2) carry out standardization processing, wherein U=l+1,
L+2 ..., l+u } represent the indexed set of Unlabeled data；
<mrow>
<msub>
<mi>z</mi>
<mrow>
<mi>k</mi>
<mo>,</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<mrow>
<msub>
<mi>f</mi>
<mrow>
<mi>k</mi>
<mo>,</mo>
<mi>l</mi>
<mo>+</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>-</mo>
<mi>min</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>f</mi>
<mrow>
<mi>k</mi>
<mo>,</mo>
<mi>U</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mi>max</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>f</mi>
<mrow>
<mi>k</mi>
<mo>,</mo>
<mi>U</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>min</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>f</mi>
<mrow>
<mi>k</mi>
<mo>,</mo>
<mi>U</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>&times;</mo>
<mn>2</mn>
<mo>-</mo>
<mn>1</mn>
<mo>,</mo>
<mo>&ForAll;</mo>
<mi>k</mi>
<mo>&Element;</mo>
<mrow>
<mo>{</mo>
<mrow>
<mn>1</mn>
<mo>,</mo>
<mn>2</mn>
<mo>,</mo>
<mn>...</mn>
<mo>,</mo>
<mi>m</mi>
</mrow>
<mo>}</mo>
</mrow>
<mo>,</mo>
<mo>&ForAll;</mo>
<mi>i</mi>
<mo>&Element;</mo>
<mrow>
<mo>{</mo>
<mrow>
<mn>1</mn>
<mo>,</mo>
<mn>2</mn>
<mo>,</mo>
<mn>...</mn>
<mo>,</mo>
<mi>u</mi>
</mrow>
<mo>}</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
The new Unlabeled data of the m groups that are obtained after standardization predicts the outcomeK=1,2 ..., m is met
min(zk)=- 1, max (zk)=1,Calculate the loss that each group newly predicts the outcomeK=1,
2,...,m；Find out the predicted value f of correspondence least disadvantage (largest interval)k, it is used as the final output that predicts the outcome.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710298619.9A CN107122800B (en) | 2017-04-27 | 2017-04-27 | Robust digital image labeling method based on prediction result screening |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710298619.9A CN107122800B (en) | 2017-04-27 | 2017-04-27 | Robust digital image labeling method based on prediction result screening |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107122800A true CN107122800A (en) | 2017-09-01 |
| CN107122800B CN107122800B (en) | 2020-09-18 |
Family
ID=59725131
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710298619.9A Active CN107122800B (en) | 2017-04-27 | 2017-04-27 | Robust digital image labeling method based on prediction result screening |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107122800B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034214A (en) * | 2018-07-09 | 2018-12-18 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating label |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7814040B1 (en) * | 2006-01-31 | 2010-10-12 | The Research Foundation Of State University Of New York | System and method for image annotation and multi-modal image retrieval using probabilistic semantic models |
| CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
| CN103116893A (en) * | 2013-03-15 | 2013-05-22 | 南京大学 | Digital image labeling method based on multi-exampling multi-marking learning |
| CN105184303A (en) * | 2015-04-23 | 2015-12-23 | 南京邮电大学 | Image marking method based on multi-mode deep learning |
-
2017
- 2017-04-27 CN CN201710298619.9A patent/CN107122800B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7814040B1 (en) * | 2006-01-31 | 2010-10-12 | The Research Foundation Of State University Of New York | System and method for image annotation and multi-modal image retrieval using probabilistic semantic models |
| CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
| CN103116893A (en) * | 2013-03-15 | 2013-05-22 | 南京大学 | Digital image labeling method based on multi-exampling multi-marking learning |
| CN105184303A (en) * | 2015-04-23 | 2015-12-23 | 南京邮电大学 | Image marking method based on multi-mode deep learning |
Non-Patent Citations (3)
| Title |
|---|
| LIANG PENG ET AL: "An empirical study of automatic image annotation through Multi-Instance Multi-Label Learning", 《2010 IEEE YOUTH CONFERENCE ON INFORMATION, COMPUTING AND TELECOMMUNICATIONS》 * |
| 史彩娟等: "基于Hessian半监督特征选择的网络图像标注", 《计算机应用研究》 * |
| 李艳洁: "图像自动标注算法研究", 《万方学位论文库》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034214A (en) * | 2018-07-09 | 2018-12-18 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating label |
| CN109034214B (en) * | 2018-07-09 | 2019-12-27 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating a mark |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107122800B (en) | 2020-09-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105808752B (en) | A kind of automatic image marking method based on CCA and 2PKNN | |
| CN109977780A (en) | A kind of detection and recognition methods of the diatom based on deep learning algorithm | |
| CN111461190A (en) | Deep convolutional neural network-based non-equilibrium ship classification method | |
| CN108564115A (en) | Semi-supervised polarization SAR terrain classification method based on full convolution GAN | |
| CN109936582A (en) | Construct the method and device based on the PU malicious traffic stream detection model learnt | |
| CN110598693A (en) | Ship plate identification method based on fast-RCNN | |
| CN111882446A (en) | Abnormal account detection method based on graph convolution network | |
| CN107833213A (en) | A kind of Weakly supervised object detecting method based on pseudo- true value adaptive method | |
| CN109284779A (en) | Object detection method based on deep full convolution network | |
| CN104751175B (en) | SAR image multiclass mark scene classification method based on Incremental support vector machine | |
| EP3033713A1 (en) | Recognition process of an object in a query image | |
| CN107330027A (en) | A kind of Weakly supervised depth station caption detection method | |
| CN110288013A (en) | A kind of defective labels recognition methods based on block segmentation and the multiple twin convolutional neural networks of input | |
| CN105930792A (en) | Human action classification method based on video local feature dictionary | |
| CN104750875A (en) | Machine error data classification method and system | |
| WO2015146113A1 (en) | Identification dictionary learning system, identification dictionary learning method, and recording medium | |
| CN102129565B (en) | Object detection method based on feature redundancy elimination AdaBoost classifier | |
| CN114299324A (en) | Pathological image classification method and system based on multi-scale domain confrontation network | |
| CN109523514A (en) | To the batch imaging quality assessment method of Inverse Synthetic Aperture Radar ISAR | |
| CN112016618A (en) | Measurement method for generalization capability of image semantic segmentation model | |
| CN105938547B (en) | A kind of papery Water Year Book digitizing solution | |
| CN104809468A (en) | Multi-view classification method based on indefinite kernels | |
| CN106548195A (en) | A kind of object detection method based on modified model HOG ULBP feature operators | |
| CN107122800A (en) | A kind of Robust digital figure mask method based on the screening that predicts the outcome | |
| Ghanmi et al. | Table detection in handwritten chemistry documents using conditional random fields |
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 |