CN113283578A - Data denoising method based on marking risk control - Google Patents
Data denoising method based on marking risk control Download PDFInfo
- Publication number
- CN113283578A CN113283578A CN202110399544.XA CN202110399544A CN113283578A CN 113283578 A CN113283578 A CN 113283578A CN 202110399544 A CN202110399544 A CN 202110399544A CN 113283578 A CN113283578 A CN 113283578A
- Authority
- CN
- China
- Prior art keywords
- data
- risk
- neural networks
- training
- networks
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000012954 risk control Methods 0.000 title claims abstract description 14
- 238000013528 artificial neural network Methods 0.000 claims abstract description 47
- 239000003550 marker Substances 0.000 claims description 8
- 238000011478 gradient descent method Methods 0.000 claims description 5
- 239000013598 vector Substances 0.000 claims description 4
- 238000002372 labelling Methods 0.000 claims 2
- 238000013135 deep learning Methods 0.000 abstract description 3
- 230000007786 learning performance Effects 0.000 abstract description 3
- 238000009825 accumulation Methods 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000000593 degrading effect Effects 0.000 description 1
- 230000002542 deteriorative effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012358 sourcing Methods 0.000 description 1
- 230000002747 voluntary effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a data denoising method based on marking risk control, wherein the success of data deep learning usually depends on a large amount of accurately marked data, but a large amount of accurately marked data is usually difficult to collect in an actual scene. In order to reduce the influence of data marking noise on the performance of the neural networks, the method maintains that two neural networks mutually select data with small loss as low-risk data to update the peer-to-peer networks, and each network respectively filters out high-risk data and retrains the rest data. In order to solve the problem that learning performance is degraded due to the fact that two networks are more and more similar along with training, when the inconsistency of the two neural networks is stable, data are stopped being selected from each other, and the networks are updated until convergence is achieved by using the obtained low-risk data. Compared with the prior art, the deep neural network has stronger robustness.
Description
Technical Field
The invention relates to a data denoising method based on marker risk control, which can screen high-risk marker data to improve robustness and belongs to the technical field of computer artificial intelligent data analysis.
Background
In recent years, deep learning has been highly successful in various fields such as face recognition, automatic driving, machine translation, and the like. However, the performance of deep learning depends on a large amount of accurately labeled data, and in practical application, it is often difficult to collect a large amount of accurately labeled data, because obtaining accurate data labels requires a large amount of manpower and material resources. To address this problem, people often use crowd sourcing techniques to assign large amounts of unmarked data to voluntary users for marking, and the resulting markings tend to be noisy due to the level of the users being uneven. How to learn from data with labeled noise becomes a concern for many scholars.
Disclosure of Invention
The purpose of the invention is as follows: in practical application, training data of the deep neural network is often noisy, and training the deep neural network on the data with labeled noise can impair the classification performance of the deep neural network. In order to solve the problem, the invention provides a data denoising method based on marking risk control. The invention maintains that two neural networks mutually select data with small loss to learn by each other, and then each network discovers and removes high-risk data in the data and retrains the rest data. The difference between the two networks is taken care of during the training process to prevent the learning performance from degrading and thus improve the robustness.
The technical scheme is as follows: a data denoising method based on marking risk control comprises the following steps:
first, a data set is prepared
The data therein contains the marking noise. Randomly initializing two peer-to-peer deep neural networks in a data set
Training the T rounds respectively to obtain two neural networks f and g.
For a data set
Respectively calculating the cross entropy loss of each data by using the neural networks f and g, sorting the data by using the cross entropy loss and respectively selecting the data with small part loss
And
neural network f is
Go up training K round to get f ', f' at
The data in which the prediction result is inconsistent with the original mark is regarded as high-risk data
Remove high risk data and leave the data
(\ represents the difference set of the two sets,
indicating that the previously selected data has a small loss,
to represent
High risk data in, so
To represent
In (1) removing
The data remaining thereafter) is retrained. In the same way g is
Training K rounds to obtain g 'and g' in
The data of (2) are predicted, and the data of which the prediction result is inconsistent with the original mark are regarded as high-risk samples
Remove high risk data and leave the data
And g is retrained.
In each training round, data that f and g have not been seen in the current round are calculated
And calculating the inconsistency, and if the inconsistency tends to be stable, or the number of training rounds reaches a preset maximum value N. The learning process in the previous paragraph is stopped. And low risk data obtained at the end
And
f and g are trained respectively until the neural networks converge, and finally two neural networks are obtained through training.
In the prediction stage, the user inputs the feature vectors of the data to be detected on the two modes to the two neural networks obtained by training respectively, the two neural networks return the prediction results of the data to be detected to the user respectively, and then the one with higher confidence coefficient is selected from the two prediction results and is output as the final mark of the data.
The data set
Is an image data set in which the image data contains a marker noise. In the prediction stage, the user inputs the feature vectors of the image data to be detected on two modes to the two neural networks obtained by training respectively, the two neural networks return the prediction results of the image data to be detected to the user respectively,and then the one with higher confidence is selected from the two prediction results to be output as the final mark of the image data.
Has the advantages that: compared with the prior art, the data denoising method based on the marker risk control maintains two neural networks, each network selects data with small loss as low-risk data through cross entropy loss to update the peer-to-peer networks, then each network discovers and deletes the data still with higher risk in the networks and trains the networks again, and meanwhile, the difference between the two networks is concerned in the training process to prevent the learning performance from deteriorating. The noise of the image data in the image classification task can have adverse effect on a subsequent neural network classification model, and the more serious the noise is, the more the performance of the neural network is damaged. Noise in a real image is a complex result of accumulation of various noise components, so that the task of image denoising is very difficult. Compared with the prior art, the method can better remove image noise and obtain training data with higher purity, so that the neural network has stronger robustness.
Drawings
FIG. 1 is a schematic diagram of the method of the present invention;
FIG. 2 is a flow chart of the method of the present invention.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
As shown in fig. 1, a data denoising method based on labeled risk control maintains that two neural networks mutually select data with small loss as low-risk data to update a peer-to-peer network based on a small loss criterion, each network finds and removes high-risk data therein and retrains the remaining data, the inconsistency of the two networks is concerned in the training process, if the inconsistency tends to be stable or the number of learning rounds reaches a preset maximum number of learning rounds, the mutual selection of data is stopped, and the two networks are respectively trained on the low-risk data obtained in the last round until convergence, thereby controlling risk and improving robustness.
The image data denoising method based on the marking risk control comprises the following steps:
step 100 of preparing an image dataset
Wherein the images should have the same dimensions with the presence of marking noise.
Step 101, determining a network architecture, such as VGG, ResNet, EfficientNet and the like, according to the requirements of an image classification task, randomly initializing two peer-to-peer deep neural networks, and using a gradient descent method to perform image data set
Is trained T times over all image data to obtain two neural networks f and g.
Step 102, for an image dataset
For each image data, the neural networks f and g calculate the cross entropy loss separately. Ordering the image data using cross-entropy loss, selecting small loss samples of the top R (t) ratio to construct an image data set
And
r (t) is a parameter for controlling the proportion of the selected image data. l represents the loss function used in training the neural network, typically the cross entropy loss.
103, the two neural networks respectively screen out high-risk image data in the small loss image data selected by the peer-to-peer network and retrain the remaining image data, and the specific steps are as follows:
step 1031, f in the image dataset
Training a K wheel on all image data by using a gradient descent method to obtain f'; g in the image dataset
Training K rounds using gradient descent method on all image data to obtain g'.
Step 1032, will
Is input to the f' acquisition prediction flag. Image data in which the prediction flag does not coincide with the original flag is regarded as high-risk image data
Will be provided with
Each image data of (a) is input to the g' obtaining prediction flag. Image data in which the prediction flag does not coincide with the original flag is regarded as high-risk image data
Representing an image training data, wherein x represents an image feature,
which is a label of the collected image, because of the presence of the marking noise,
not necessarily its true mark y. y isf′And yg′Representing the prediction labels given by the neural networks f 'and g', respectively, for image x.
In step 1033, the remaining image data set
Training again on f; in the remaining image data set
And g is retrained.
And 104, calculating the inconsistency of the two neural networks on the image data which is not used for updating the neural networks, stopping mutual learning if the inconsistency reaches a stable state or the number of training rounds reaches a preset maximum value N, and returning to the step 102 to continue training if the inconsistency reaches the stable state or the number of training rounds reaches the preset maximum value N.
And 105, respectively training the two neural networks on the low-risk image data obtained in the last round until convergence, so as to obtain two neural networks f and g.
And step 106, inputting the image data to be detected into f and g respectively to obtain prediction marks, and outputting the prediction mark with higher confidence coefficient as a final mark.
Claims (7)
1. A data denoising method based on labeling risk control is characterized in that two neural networks are maintained, data with small loss are mutually selected to serve as low-risk data to be updated to peer-to-peer networks based on a small loss criterion, each network finds and removes high-risk data in the high-risk data and retrains the high-risk data on the rest data, the inconsistency of the two networks is concerned in the training process, if the inconsistency tends to be stable or the number of learning rounds reaches a preset maximum number of learning rounds, the mutual selection of samples is stopped, the two networks are respectively trained on the low-risk samples obtained in the last round until convergence, the two neural networks are obtained, and the neural networks are used for denoising the data to be processed.
2. The method of claim 1, wherein the data set is prepared first
Wherein the data contains a marker noise; randomly initializing two peer-to-peer deep neural networks in a data set
Training T wheels respectively to obtain two neural networks f and g;
for a data set
Respectively calculating the cross entropy loss of each data by using the neural networks f and g, sorting the data by using the cross entropy loss and respectively selecting the data with small part loss
And
neural network f is
Go up training K round to get f ', f' at
The data in which the prediction result is inconsistent with the original mark is regarded as high-risk data
Remove high risk data and leave the data
Training again on f; in the same way g is
Training K rounds to obtain g 'and g' in
The data of (2) are predicted, and the data of which the prediction result is inconsistent with the original mark are regarded as high-risk samples
Remove high risk data and leave the data
G is retrained;
in each training round, data that f and g have not been seen in the current round are calculated
Calculating the inconsistency, and if the inconsistency tends to be stable, or the number of training rounds reaches a preset maximum value N; stopping the learning process in the previous paragraph; and low risk data obtained at the end
And
f and g are trained respectively until the neural networks converge, and finally two neural networks are obtained through training.
3. The data denoising method based on labeling risk control as claimed in claim 1, wherein in the prediction stage, the user inputs the feature vector of the data to be tested into two neural networks obtained by training, the two neural networks respectively return the prediction results of the data to be tested to the user, and then the one with higher confidence coefficient is selected from the two prediction results and used as the final label output of the data.
4. The method of claim 1, wherein the data set is denoised based on marker risk control
Is an image data set in which the image data contains a marker noise.
5. The method of claim 4, wherein in the prediction stage, the user inputs the feature vectors of the image data to be detected in the two modalities into the two neural networks obtained by training, the two neural networks respectively return the prediction results of the image data to be detected to the user, and then the one with higher confidence is selected from the two prediction results as the final label of the image data for output.
6. The method of claim 1, wherein the data denoising method based on marker risk control is applied to a data set
Respectively calculating the cross entropy loss of each data by using the neural networks f and g, sorting the data by using the cross entropy loss, and respectively selecting small loss samples with the first R (t) proportion to construct a data set
And
r (t) is a parameter for controlling the proportion of the selected data.
7. The method for denoising data based on labeled risk control according to claim 2, wherein each network finds and removes high risk data therein and retrains on the remaining data, comprising the specific steps of:
step 1031, f in data set
Training a K round by using a gradient descent method to obtain f'; g in the data set
Training a K round by using a gradient descent method to obtain g';
step 1032, will
Each data input to f' gets a predictive flag; image data in which the prediction flag does not coincide with the original flag is regarded as high-risk data
Will be provided with
Each data input to g' gets a predictive flag; data in which the predicted tag is inconsistent with the original tag is considered high risk data
Step 1033, in the remaining data set
Training again on f; in the remaining data set
And g is retrained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110399544.XA CN113283578A (en) | 2021-04-14 | 2021-04-14 | Data denoising method based on marking risk control |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110399544.XA CN113283578A (en) | 2021-04-14 | 2021-04-14 | Data denoising method based on marking risk control |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113283578A true CN113283578A (en) | 2021-08-20 |
Family
ID=77276660
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110399544.XA Pending CN113283578A (en) | 2021-04-14 | 2021-04-14 | Data denoising method based on marking risk control |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113283578A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114330439A (en) * | 2021-12-28 | 2022-04-12 | 盐城工学院 | Bearing diagnosis method based on convolutional neural network |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740057A (en) * | 2018-12-28 | 2019-05-10 | 武汉大学 | A kind of strength neural network and information recommendation method of knowledge based extraction |
| CN110110780A (en) * | 2019-04-30 | 2019-08-09 | 南开大学 | A kind of picture classification method based on confrontation neural network and magnanimity noise data |
| CN110310199A (en) * | 2019-06-27 | 2019-10-08 | 上海上湖信息技术有限公司 | Borrow or lend money construction method, system and the debt-credit Risk Forecast Method of risk forecast model |
| EP3582142A1 (en) * | 2018-06-15 | 2019-12-18 | Université de Liège | Image classification using neural networks |
| US20200034693A1 (en) * | 2018-07-27 | 2020-01-30 | Samsung Electronics Co., Ltd. | Method for detecting defects in semiconductor device |
| CN111160474A (en) * | 2019-12-30 | 2020-05-15 | 合肥工业大学 | Image identification method based on deep course learning |
| CN111339934A (en) * | 2020-02-25 | 2020-06-26 | 河海大学常州校区 | Human head detection method integrating image preprocessing and deep learning target detection |
| CN111861909A (en) * | 2020-06-29 | 2020-10-30 | 南京理工大学 | Network fine-grained image denoising and classifying method |
| CN111931637A (en) * | 2020-08-07 | 2020-11-13 | 华南理工大学 | Cross-modal pedestrian re-identification method and system based on double-current convolutional neural network |
| CN111985520A (en) * | 2020-05-15 | 2020-11-24 | 南京智谷人工智能研究院有限公司 | Multi-mode classification method based on graph convolution neural network |
| CN112101328A (en) * | 2020-11-19 | 2020-12-18 | 四川新网银行股份有限公司 | Method for identifying and processing label noise in deep learning |
| CN112529024A (en) * | 2019-09-17 | 2021-03-19 | 株式会社理光 | Sample data generation method and device and computer readable storage medium |
| WO2021064856A1 (en) * | 2019-10-01 | 2021-04-08 | 日本電気株式会社 | Robust learning device, robust learning method, program, and storage device |
| WO2021064787A1 (en) * | 2019-09-30 | 2021-04-08 | 日本電気株式会社 | Learning system, learning device, and learning method |
-
2021
- 2021-04-14 CN CN202110399544.XA patent/CN113283578A/en active Pending
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3582142A1 (en) * | 2018-06-15 | 2019-12-18 | Université de Liège | Image classification using neural networks |
| US20200034693A1 (en) * | 2018-07-27 | 2020-01-30 | Samsung Electronics Co., Ltd. | Method for detecting defects in semiconductor device |
| CN109740057A (en) * | 2018-12-28 | 2019-05-10 | 武汉大学 | A kind of strength neural network and information recommendation method of knowledge based extraction |
| CN110110780A (en) * | 2019-04-30 | 2019-08-09 | 南开大学 | A kind of picture classification method based on confrontation neural network and magnanimity noise data |
| CN110310199A (en) * | 2019-06-27 | 2019-10-08 | 上海上湖信息技术有限公司 | Borrow or lend money construction method, system and the debt-credit Risk Forecast Method of risk forecast model |
| CN112529024A (en) * | 2019-09-17 | 2021-03-19 | 株式会社理光 | Sample data generation method and device and computer readable storage medium |
| WO2021064787A1 (en) * | 2019-09-30 | 2021-04-08 | 日本電気株式会社 | Learning system, learning device, and learning method |
| WO2021064856A1 (en) * | 2019-10-01 | 2021-04-08 | 日本電気株式会社 | Robust learning device, robust learning method, program, and storage device |
| CN111160474A (en) * | 2019-12-30 | 2020-05-15 | 合肥工业大学 | Image identification method based on deep course learning |
| CN111339934A (en) * | 2020-02-25 | 2020-06-26 | 河海大学常州校区 | Human head detection method integrating image preprocessing and deep learning target detection |
| CN111985520A (en) * | 2020-05-15 | 2020-11-24 | 南京智谷人工智能研究院有限公司 | Multi-mode classification method based on graph convolution neural network |
| CN111861909A (en) * | 2020-06-29 | 2020-10-30 | 南京理工大学 | Network fine-grained image denoising and classifying method |
| CN111931637A (en) * | 2020-08-07 | 2020-11-13 | 华南理工大学 | Cross-modal pedestrian re-identification method and system based on double-current convolutional neural network |
| CN112101328A (en) * | 2020-11-19 | 2020-12-18 | 四川新网银行股份有限公司 | Method for identifying and processing label noise in deep learning |
Non-Patent Citations (6)
| Title |
|---|
| J. ZHANG 等: "Improving Crowdsourced Label Quality Using Noise Correction", 《 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS》, vol. 29, no. 5, pages 1675 - 1688, XP011681456, DOI: 10.1109/TNNLS.2017.2677468 * |
| WANG WEI 等: "Learnability of Multi-Instance Multi-Label Learning", 《CHINESE SCIENCE BULLETIN》, vol. 57, no. 19, pages 2488, XP035075248, DOI: 10.1007/s11434-012-5133-z * |
| WANG, WEI 等: "Adaptive Switching Anisotropic Diffusion Model for Universal Noise Removal", 《 PROCEEDINGS OF THE 10TH WORLD CONGRESS ON INTELLIGENT CONTROL AND AUTOMATION (WCICA 2012)》, pages 4803 - 4808 * |
| ZHENGWEN ZHANG 等: "Making Deep Neural Networks Robust to Label Noise: A Reweighting Loss and Data Filtration", 《 2019 IEEE 4TH INTERNATIONAL CONFERENCE ON SIGNAL AND IMAGE PROCESSING (ICSIP)》, pages 289 - 293 * |
| 周杭驰 等: "基于深度学习的图像分类标注研究", 《中国优秀硕士学位论文全文数据库信息科技辑》, no. 2021 * |
| 郭翔宇 等: "一种改进的协同训练算法:Compatible Co-training", 《南京大学学报(自然科学)》, vol. 52, no. 04, pages 662 - 671 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114330439A (en) * | 2021-12-28 | 2022-04-12 | 盐城工学院 | Bearing diagnosis method based on convolutional neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108829826B (en) | Image retrieval method based on deep learning and semantic segmentation | |
| CN109299274B (en) | Natural scene text detection method based on full convolution neural network | |
| CN108288051B (en) | Pedestrian re-recognition model training method and device, electronic equipment and storage medium | |
| CN108335303B (en) | Multi-scale palm skeleton segmentation method applied to palm X-ray film | |
| CN111079847B (en) | Remote sensing image automatic labeling method based on deep learning | |
| JP2018097807A (en) | Learning device | |
| CN107480723B (en) | Texture Recognition based on partial binary threshold learning network | |
| CN110728694B (en) | Long-time visual target tracking method based on continuous learning | |
| CN111126470B (en) | Image data iterative cluster analysis method based on depth measurement learning | |
| CN108734200B (en) | Human target visual detection method and device based on BING (building information network) features | |
| CN111239137B (en) | Grain quality detection method based on transfer learning and adaptive deep convolution neural network | |
| CN112101364B (en) | Semantic segmentation method based on parameter importance increment learning | |
| CN114048822A (en) | Attention mechanism feature fusion segmentation method for image | |
| CN109872326B (en) | Contour detection method based on deep reinforced network jump connection | |
| KR20220116270A (en) | Learning processing apparatus and method | |
| CN112581483B (en) | Self-learning-based plant leaf vein segmentation method and device | |
| CN113283578A (en) | Data denoising method based on marking risk control | |
| CN117115614B (en) | Object identification method, device, equipment and storage medium for outdoor image | |
| US11776292B2 (en) | Object identification device and object identification method | |
| CN108428234B (en) | Interactive segmentation performance optimization method based on image segmentation result evaluation | |
| CN109255794B (en) | Standard part depth full convolution characteristic edge detection method | |
| CN117011515A (en) | Interactive image segmentation model based on attention mechanism and segmentation method thereof | |
| CN114708307B (en) | Target tracking method, system, storage medium and device based on correlation filter | |
| CN116310466A (en) | Small sample image classification method based on local irrelevant area screening graph neural network | |
| CN114758135A (en) | Unsupervised image semantic segmentation method based on attention mechanism |
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 |