CN111738343B - Image labeling method based on semi-supervised learning - Google Patents
Image labeling method based on semi-supervised learning Download PDFInfo
- Publication number
- CN111738343B CN111738343B CN202010589985.1A CN202010589985A CN111738343B CN 111738343 B CN111738343 B CN 111738343B CN 202010589985 A CN202010589985 A CN 202010589985A CN 111738343 B CN111738343 B CN 111738343B
- Authority
- CN
- China
- Prior art keywords
- class
- sub
- voter
- marked
- different
- 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.)
- Active
Links
- 238000002372 labelling Methods 0.000 title claims abstract description 9
- 238000002156 mixing Methods 0.000 claims abstract description 9
- 238000013135 deep learning Methods 0.000 claims abstract description 8
- 238000000034 method Methods 0.000 claims description 22
- 238000012549 training Methods 0.000 claims description 17
- 238000013145 classification model Methods 0.000 claims description 9
- 230000007717 exclusion Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 2
- 238000010801 machine learning Methods 0.000 abstract description 3
- 238000013136 deep learning model Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000002401 inhibitory effect Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/259—Fusion by voting
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 an image labeling method based on semi-supervised learning, which is characterized in that different classifiers are designed for samples of different categories, partial labeled samples are utilized to train the classifier, the results of the different classifiers are voted, and the category with the highest accuracy is selected, so that unknown samples are labeled. In order to reduce the influence caused by error classification, the samples in each category obtained by the classifier and the samples in the labeled corresponding category are subjected to random linear mixing operation, so that the error classification result also contains the characteristics of the corresponding category, and a new idea is provided for the fields of deep learning and machine learning by semi-supervised learning.
Description
Technical Field
The invention belongs to the field of semi-supervised learning, and relates to a method for image labeling based on semi-supervised learning.
Background
In recent years, the introduction of deep learning technology and the maturation of machine learning technology make breakthrough progress in the field of computer vision, and the traditional problems in the field of computer vision, such as classification problems, detection problems and semantic segmentation problems, are better solved. However, most computer vision tasks are supervised learning tasks, which means that all input data needs to be manually marked, people need to acquire related pictures, and carefully mark the pictures by using related software, which is very labor and material consuming, in the field of object detection, such as the coco data set contains eighty classes, and seventy-five thousands of pictures in total, each of which needs to manually find out the objects to be marked one by one and mark the corresponding objects with rectangular frames, and the marking of such large data sets generally needs hundreds of people to spend several worship or even months. Errors can be avoided during manual annotation, errors in annotation can cause some influence on later training, and error correction is difficult, so that a method is needed to solve the problem of data set annotation, improve the annotation accuracy and save some labor cost. Semi-supervised learning is always a research hotspot in the field of machine learning, is a method for combining supervised learning and non-supervised learning, trains a model through a part of marked files and a large number of unmarked files, can greatly lighten the burden of staff, but the current semi-supervised learning method cannot enable the training result of the model to be the same as the supervised training result, so that the supervised learning is still the main stream in the field of computer vision, the semi-supervised learning process and the supervised learning have a common place, the semi-supervised learning needs to mark part of data, a small amount of marked data can be used for classifying and marking the unmarked data, and the model trained by a small amount of samples also has a good classification effect through the prior art, so the method for assisting the marking by using the semi-supervised learning method is provided.
Disclosure of Invention
In order to solve the problems, the technical scheme of the invention is a method for image labeling based on semi-supervised learning, which comprises the following steps:
The method comprises the following steps:
S10, adding background types: the target is divided into A class and B class, and the background class formed by random sampling of other classes than A or B class is introduced first;
S20, constructing a cross network classification model: marking M classified picture data, training a model between every two classes by using a deep learning network, so that M (M-1) models are all used, different networks are selected during training of the classes A, B in different sequences, the marked data are trained through the M (M-1) models, and M-1 independent classification models of one class are marked as a class learner, and M class learners are all used;
s30, forming a sub-voter: predicting the class of unlabeled data through M class learners, forming a voting subset by all results comprising a certain class, and sharing M sub-voters according to the number of the classes, wherein each sub-voter comprises 2M-2 groups of different prediction results, each group comprises prediction of a certain class, and when the overlapping condition of class Avs class B and class Bvs class A occurs, the training networks of the class Avs class B and the class Bvs class A are different, and the prediction is also different;
S40, voting according to a mutual exclusion voter: each sub-voter in the M sub-voters generates 2M-2 prediction probabilities for the same picture, a series of rules and thresholds are set, so long as the voting result generated by the sub-voter exceeds the thresholds, the sample is considered to belong to the category corresponding to the set, only one picture of the prediction category of the voting result is reserved, and the prediction label is marked.
S50, correcting the error label based on random linear mixing: the newly marked samples are mixed with the original same-label samples in random proportion, so that interference caused by the erroneously marked samples to network training is restrained.
Preferably, the voting comprises the steps of:
S41, calculating a weighted average of accuracy by each sub-voter, wherein the accuracy of each classification network is different, and calculating a corresponding weight coefficient according to the accuracy, so that the predicted result of the model with high accuracy has larger influence on the final annotation;
S42, each sub-voter set scores the picture, and counts the number N of sub-voters exceeding a threshold;
s43, only the data with n=1 is reserved and labeled with a predictive label.
In the voting process, when the probability of two sub-voters appearing exceeds a threshold value, the representative sample may belong to two different labels, and the sub-voters judge that the voting fails and discard the sample.
Preferably, the random linear mixture fuses similar data, no new label is generated, and the calculation formula is as follows:
Wherein x i is a marked picture of the same category as x j, x j is a marked file output by the model, beta is a random parameter, and the value range is 0-1.
Preferably, the threshold is 0.95.
The beneficial effects of the invention are as follows: aiming at the condition that a small amount of marked data and a large amount of unmarked data exist in the image marking, the invention provides a method for predicting the large amount of unmarked data by adopting semi-supervised integrated learning, thereby realizing accurate category prediction of the large amount of unmarked data, providing a random linear mixing method according to the process of learning the image data by a deep learning network, and realizing more accurate image marking.
Drawings
FIG. 1 is a flow chart of steps of a method for image annotation based on semi-supervised learning according to an embodiment of the present invention;
Fig. 2 is a schematic diagram of a method for image labeling based on semi-supervised learning according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
On the contrary, the invention is intended to cover any alternatives, modifications, equivalents, and variations as may be included within the spirit and scope of the invention as defined by the appended claims. Further, in the following detailed description of the present invention, certain specific details are set forth in order to provide a better understanding of the present invention. The present invention will be fully understood by those skilled in the art without the details described herein.
Referring to fig. 1, a flowchart of steps of a method for image labeling based on semi-supervised learning according to an embodiment of the present invention includes the following steps:
S10, adding background types: the target is divided into A class and B class, and the background class formed by random sampling of other classes than A or B class is introduced first;
First, a background class is introduced for every two kinds of basic classifiers to resist the predicted interference problem of unknown class samples. In this way, the two classification models of the base learner are converted into three classification models, the proportion of the three classes A-B-background classes is 1:1:1, and the background class pictures are randomly extracted from other classes in equal proportion;
S20, constructing a cross network classification model: marking M classified picture data, training a model between every two classes by using a deep learning network, so that M (M-1) models are all used, different networks are selected during training of the classes A, B in different sequences, the marked data are trained through the M (M-1) models, and M-1 independent classification models of one class are marked as a class learner, and M class learners are all used;
Taking M as a specific embodiment, 20 is taken as a specific embodiment, and the data labeling system is constructed based on ensemble learning, so that weak classifiers trained by deep learning between every two different classes A-B are integrated. Finally, the pictures are marked on the basis of the integrated and learned models. The 20 classified picture datase:Sub>A are marked, ase:Sub>A model is trained between every two classes by using ase:Sub>A deep learning network, so 380 models (20X 19) are combined, weak classifiers participating in integration are independent as much as possible, the class A is trained by using the same network when the class B-A is repeated in order to avoid A-B, the class A is indexed, when the index A is larger than B, the deep learning model PnasNet (batch_size=64) is selected, and when the index X is smaller than Y, the deep learning model SeNet (batch_size=32) is selected. Different batch_size may make the characteristics learned by the network inconsistent, so different batch_size is selected for both cases. Training a small amount of marked data through 380 models, and recording 19 independent classification models of a certain class as a class learner, so that 20 class learners are all provided
S30, forming a sub-voter: predicting the class of unlabeled data through M class learners, forming a voting subset sum of all results comprising a certain class, wherein according to the number of the classes, M voting subset sums are shared, each voting subset sum comprises 2M-2 groups of different prediction results, each group comprises prediction of a certain determined class, and when the overlapping condition of class Avs class B and class Bvs class A occurs, the training networks of the class Avs class B and the class Bvs class A are different, and the prediction is also different;
A large amount of unlabeled data is passed through 20 classics to predict categories. All results comprising a certain class are formed into a sub-voter. So there are a total of 20 sub-voters, based on the number of categories, each sub-voter containing 38 different sets of predictions, each of the 38 sets of predictions containing predictions for a particular category. As described above, when the overlapping cases of "class Avs class B" and "class Bvs class a" occur, the networks trained by the two are different, so the predictions are also different.
S40, voting according to a mutual exclusion voter: each sub-voter in the M sub-voters generates 2M-2 prediction probabilities for the same picture, a series of rules and thresholds are set, and as long as the voting result generated by the sub-voter exceeds the thresholds, the sample is considered to belong to the category corresponding to the sub-voter set, only one picture of the prediction category of the voting result is reserved, and the prediction label is marked;
Each sub-voter set contains 38 prediction probabilities for the same picture, each sub-voter calculates a weighted average of accuracy, each classification network has different accuracy, and corresponding weight coefficient is calculated according to the accuracy, so that the predicted result of the model with high accuracy has larger influence on final annotation; then, each sub-voter set scores the picture, and counts the number N of sub-voters exceeding a threshold value of 0.95. Finally, only data with N=1 is reserved, and a prediction label is marked on the data to complete mutual exclusion voting prediction;
S50, correcting the error label based on random linear mixing: mixing the newly marked sample with the original same-label sample in random proportion, and inhibiting interference brought by the erroneously marked sample on network training;
The invention provides an RLB data mixing enhancement algorithm (Random linear blending), wherein the newly marked sample is mixed with the original same-tag sample in a random proportion, so that the interference of the erroneously marked sample on the network training is inhibited.
The voting comprises the following steps:
S41, calculating a weighted average of accuracy by each sub-voter, wherein the accuracy of each classification network is different, and calculating a corresponding weight coefficient according to the accuracy, so that the predicted result of the model with high accuracy has larger influence on the final annotation;
S42, each sub-voter set scores the picture, and counts the number N of sub-voters exceeding a threshold;
s43, only the data with n=1 is reserved and labeled with a predictive label.
In the voting process, when the probability of two sub-voters appearing exceeds a threshold value, the representative sample may belong to two different labels, and the sub-voters judge that the voting fails and discard the sample.
The random linear mixing fuses similar data, no new label is generated, and the calculation formula is as follows:
Wherein x i is a marked picture of the same category as x j, x j is a marked file output by the model, beta is a random parameter, and the value range is 0-1. Referring to fig. 2, a specific effect of the operation of the method of the present invention is shown.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.
Claims (4)
1. The image labeling method based on semi-supervised learning is characterized by comprising the following steps of:
S10, adding background types: the target is divided into A class and B class, and the background class formed by random sampling of other classes than A or B class is introduced first;
S20, constructing a cross network classification model: marking M classified picture data, training a model between every two classes by using a deep learning network, so that M (M-1) models are all used, different networks are selected during training of the classes A, B in different sequences, the marked data are trained through the M (M-1) models, and M-1 independent classification models of one class are marked as a class learner, and M class learners are all used;
S30, forming a sub-voter: predicting the class of unlabeled data through M class learners, forming all results comprising a certain class into one sub-voter, forming a total of M sub-voters according to the number of the classes, wherein each sub-voter comprises 2M-2 groups of different prediction results, each group comprises prediction of a certain determined class, and when the overlapping condition of class Avs class B and class Bvs class A occurs, the training networks of the class Avs class B and the class Bvs class A are different, and the prediction is different;
S40, voting according to a mutual exclusion voter: each sub-voter in the M sub-voters generates 2M-2 prediction probabilities for the same picture, a series of rules and thresholds are set, so long as the voting result generated by the sub-voter exceeds the thresholds, the sample is considered to belong to the category corresponding to the set, only one picture of the prediction category of the voting result is reserved, and the prediction label is marked;
s50, correcting the error label based on random linear mixing: the newly marked samples are mixed with the original same-label samples in random proportion, so that interference caused by the erroneously marked samples to network training is restrained.
2. The method of claim 1, wherein the voting comprises the steps of:
S41, calculating a weighted average of accuracy by each sub-voter, wherein the accuracy of each classification network is different, and calculating a corresponding weight coefficient according to the accuracy, so that the predicted result of the model with high accuracy has larger influence on the final annotation;
S42, each sub-voter set scores the picture, and counts the number N of sub-voters exceeding a threshold;
s43, only reserving data with N=1, and marking a predictive label for the data;
In the voting process, when the probability of occurrence of two sub-voters exceeds a threshold value, the representative sample may belong to two different labels, the voter judges that the voting fails, and the sample is abandoned.
3. The method of claim 1, wherein the random linear mixture fuses homogeneous data without generating new tags, and the calculation formula is:
Wherein x i is a marked picture of the same category as x j, x j is a marked file output by the model, beta is a random parameter, and the value range is 0-1.
4. The method of claim 1, wherein the threshold is 0.95.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010589985.1A CN111738343B (en) | 2020-06-24 | 2020-06-24 | Image labeling method based on semi-supervised learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010589985.1A CN111738343B (en) | 2020-06-24 | 2020-06-24 | Image labeling method based on semi-supervised learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111738343A CN111738343A (en) | 2020-10-02 |
| CN111738343B true CN111738343B (en) | 2024-07-26 |
Family
ID=72650996
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010589985.1A Active CN111738343B (en) | 2020-06-24 | 2020-06-24 | Image labeling method based on semi-supervised learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111738343B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113254641B (en) * | 2021-05-27 | 2021-11-16 | 中国电子科技集团公司第十五研究所 | Information data fusion method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107644235A (en) * | 2017-10-24 | 2018-01-30 | 广西师范大学 | Image automatic annotation method based on semi-supervised learning |
| CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11468358B2 (en) * | 2017-11-30 | 2022-10-11 | Palo Alto Networks (Israel Analytics) Ltd. | Framework for semi-supervised learning when no labeled data is given |
-
2020
- 2020-06-24 CN CN202010589985.1A patent/CN111738343B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107644235A (en) * | 2017-10-24 | 2018-01-30 | 广西师范大学 | Image automatic annotation method based on semi-supervised learning |
| CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111738343A (en) | 2020-10-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109741332A (en) | A kind of image segmentation and mask method of man-machine coordination | |
| Vasudevan et al. | When does dough become a bagel? analyzing the remaining mistakes on imagenet | |
| CN111444966A (en) | Media information classification method and device | |
| WO2021073390A1 (en) | Data screening method and apparatus, device and computer-readable storage medium | |
| CN108932724B (en) | Automatic system auditing method based on multi-person collaborative image annotation | |
| CN110310012B (en) | Data analysis method, device, equipment and computer readable storage medium | |
| CN111738343B (en) | Image labeling method based on semi-supervised learning | |
| CN110929169A (en) | Position recommendation method based on improved Canopy clustering collaborative filtering algorithm | |
| CN111797935B (en) | Semi-supervised depth network picture classification method based on group intelligence | |
| CN111339258B (en) | University computer basic exercise recommendation method based on knowledge graph | |
| CN110457155B (en) | Sample class label correction method and device and electronic equipment | |
| CN111290953B (en) | Method and device for analyzing test logs | |
| CN112464966B (en) | Robustness estimating method, data processing method, and information processing apparatus | |
| CN107247996A (en) | A kind of Active Learning Method applied to different distributed data environment | |
| CN116306969A (en) | Federal learning method and system based on self-supervision learning | |
| CN113988044B (en) | Method for judging error question reason type | |
| CN113128556B (en) | Deep learning test case sequencing method based on mutation analysis | |
| CN109063732A (en) | Image ranking method and system based on feature interaction and multi-task learning | |
| Jiang et al. | A classification algorithm based on weighted ML-kNN for multi-label data | |
| CN110427973B (en) | Classification method for ambiguity-oriented annotation samples | |
| CN110175531B (en) | Attitude-based examinee position positioning method | |
| CN113705720A (en) | Method for reducing weighted training deviation by applying weight correction in machine learning | |
| CN113934922A (en) | Intelligent recommendation method, device, equipment and computer storage medium | |
| CN111626409B (en) | Data generation method for image quality detection | |
| KR102604756B1 (en) | Server, system, method and program providing essay scoring management service |
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 |