CN115564960B - Network image label denoising method combining sample selection and label correction - Google Patents
Network image label denoising method combining sample selection and label correction Download PDFInfo
- Publication number
- CN115564960B CN115564960B CN202211408454.3A CN202211408454A CN115564960B CN 115564960 B CN115564960 B CN 115564960B CN 202211408454 A CN202211408454 A CN 202211408454A CN 115564960 B CN115564960 B CN 115564960B
- Authority
- CN
- China
- Prior art keywords
- sample
- samples
- reusable
- network
- training
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/24—Aligning, centring, orientation detection or correction of the image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a network image label denoising method combining sample selection and label correction, which comprises the following steps: s1, firstly, selecting a clean sample according to cosine similarity between the sample and a category center; s2, selecting a reusable sample from the rest samples through sample uncertainty dynamic state and correcting; s3, finally, updating the network by using the clean sample and the corrected reusable sample; according to the method, after the clean samples are selected according to the cosine similarity between the samples and the category centers, the reusable samples are dynamically selected from the rest samples according to the uncertainty of the samples and are corrected, and finally the clean samples and the corrected reusable samples are used for updating the network, so that the utilization rate of the samples is improved, and meanwhile, the fine-grained classification performance is improved.
Description
Technical Field
The invention relates to the technical field of network label denoising, in particular to a network image label denoising method combining sample selection and label correction.
Background
For the noise problem, besides improving the accuracy of sample selection by reducing the sample coincidence rate between classes, another idea is to further reduce the influence of noise labels on the neural network by combining noise sample selection and loss correction. The method based on sample selection is to select clean samples for subsequent training by a certain method, wherein part of noise samples discarded by the sample selection method are internal noise, the samples are called reusable samples, and the true labels of the samples are still in a data set. Therefore, the utilization rate of the sample can be effectively improved by reusing the part of the sample, and the problem to be solved is a problem which needs to be solved for the classification of the fine-grained images which lack the data set.
Disclosure of Invention
The invention aims to provide a network image label denoising method combining sample selection and label correction, so as to solve the problems in the background technology.
In order to achieve the purpose, the invention provides the following technical scheme: a network image label denoising method combining sample selection and label correction comprises the following steps:
s1, firstly, selecting a clean sample according to cosine similarity between the sample and a category center;
s2, selecting a reusable sample from the rest samples through sample uncertainty dynamic state and correcting;
and S3, finally, updating the network by using the clean sample and the corrected reusable sample.
Further, in S1, the features of the picture are normalized in the Softmax layer, and the output process of the Softmax layer may be represented as:
after normalization, using a hyper-parameter s to scale cosine values, the Softmax output under the L2 constraint after feature normalization is calculated as follows:
Further, after normalization, the features are characterized by angles on a hypersphereDegree distribution, parameters of the last fully-connected layerFor the center of each class generated by pre-training, the output of the network full-connection layer is the cosine distance between the picture feature and the center of each class(ii) a Recording the cosine similarity of each picture with its corresponding class center:
is the ith sample and its class centerSorting H, taking the example with large cosine similarity in the training of each batch, sending the example into a peer-to-peer network, and carrying out the next training; the selection formula is as follows:
wherein the content of the first and second substances,for a correctable drop rate, D is the set of samples, and Dr is the reusable samples.
Further, the clean sample Dc is selected in S1, and the remaining samples can be divided into two types, namely a reusable sample Dr and a noise set Dn which need to be discarded in subsequent training;
when the sample isPrediction uncertainty ofThe sample belongs to the reusable sample set Dr if the following condition is satisfied:
whereinIs a sampleIs not determined, andto representMedian of uncertainty of medium samples, the uncertainty of each sample is measured by cross entropy:
according toRecord the sampleThe category j with the largest number of predicted times and the number m,is a sampleProbability of being predicted as j:
the uncertainty of the prediction is the smallest when the n predictions are all the same, and the time is the smallest when the n predictions are all the same,(ii) a The uncertainty is the largest when the n predictions are different, and the time is the moment,And n is 10.
Further, in S3, during the previous n training sessions, the output of the Softmax layer is smoothed, and is propagated backwards with the following loss:
Further, after n times of training, the reusable sample Dr is selected using the formula (6.6) and usedUpdating the network:
j is the category with the largest predicted times in the continuous n prediction processes.
Compared with the prior art, the invention has the beneficial effects that: according to the method, after the clean samples are selected according to the cosine similarity between the samples and the category centers, the reusable samples are dynamically selected from the rest samples according to the uncertainty of the samples and are corrected, and finally the clean samples and the corrected reusable samples are used together to update the network, so that the utilization rate of the samples is improved, and meanwhile, the fine-grained classification performance is improved.
Drawings
FIG. 1 is a schematic diagram of the front half of the CSSLC frame structure of the present invention;
FIG. 2 is a rear half of a CSSLC frame body structure of the present invention;
figure 3 is a diagram of the steps of a CSSLC method of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.
Referring to fig. 1, fig. 2 and fig. 3, the invention is a combination Sample selection and label Correction combined network image label denoising method (combination Sample selection with Loss Correction), referred to as CSSLC for short, different from the single Sample selection method and Loss Correction method, the method performs Loss Correction on a part of reusable samples on the basis of Sample selection, which can greatly improve the Sample utilization rate and the image classification performance;
firstly, a sample set is setThe division into three sets: clean sample set Dc, reusable sample set Dr and noise set Dn, sample setWhereinIs the (i) th training sample,is thatThe label of (1); for the reusable sample set Dr,is not a sampleThe true label of (1) is recorded asIn the next step, distinguishing a clean sample set Dc, reusing the sample set Dr and a noise set Dn, and performing loss correction on the reusable sample set Dr and then sending the reusable sample set Dr into a network for training;
on the premise of choosing a clean sample based on sample selection, a reusable sample is dynamically chosen again through the uncertainty of the sample for the noise sample which is to be discarded, and loss correction is carried out on the reusable sample. Since the higher the uncertainty for a sample, the more likely it is a noise sample, and the lower the uncertainty, the more likely it is a reusable sample.
In this embodiment, the conventional sample selection method first calculates the loss of the sample, then selects the sample according to the small loss, and selects the sample according to the cosine similarity between the sample and the class center, then calculates the loss, and selects the available sample before calculating the loss, and calculates the loss by using the samples.
Based on a simple observation, the network will fit a simple clean sample first, and the cosine similarity between the simple clean sample and the class center will be lower than that of the noise sample, so the clean sample is selected directly according to the cosine similarity between the sample and the class center.
The goal of Softmax is to maximize the probability of correct classification as much as possible, so it ignores some pictures that are difficult to resolve, i.e. low quality pictures, and preferentially fits high quality pictures; in order to increase the utilization rate of the picture, the features of the picture are normalized at the Softmax layer, so that the hard example gains more attention from the network, and the final output process of the Softmax layer can be expressed as:
after normalization, using a hyper-parameter s to scale cosine values, the Softmax output under the L2 constraint after feature normalization is calculated as follows:
wherein the content of the first and second substances,and withExpressing the ith sample and its label, after normalization, the characteristics are distributed in angle on the hypersphere, and the parameters of the last full connection layerFor the center of each class generated by pre-training, the output of the network full-connection layer is the cosine distance between the picture characteristic and the center of each classRecording the cosine similarity of each picture and the corresponding class center:
is the ith sample and its class centerThe cosine distances of the training data are sorted, the instances with large cosine similarity in each batch of training are sent to a peer-to-peer network, and the next step of training is carried outThe selection formula is as follows:
wherein, the first and the second end of the pipe are connected with each other,for a correctable drop rate, D is the sample set, dr is the reusable sample, and the selected picture is fed into the peer-to-peer network update network.
In this embodiment, after the clean sample Dc is selected, the remaining samples may be divided into two types, where one type of label is in the data set, and through training, the network predicts the correct label of this type of sample, and through correcting the label of this type of sample, the network may still continue to learn from this type of sample set, which is called reusable sample Dr, and another type of label is not in the data set, which is called noise set Dn, and needs to be discarded in the subsequent training.
When a reusable sample is fed into the network, the network will tend to give a definite prediction (which is not consistent with the label given by the data set) after training, and when a noisy sample is fed into the network, the network will give an uncertain prediction, entropy is used to measure the uncertainty of the sample, and the reusable sample is selected.
When the sample isPrediction uncertainty ofThe sample belongs to the reusable sample set Dr if the following condition is satisfied:
whereinIs a sampleIs not determined, andto representMedian of uncertainty of samples, cross entropy is used to measure the uncertainty of each sample:
according toRecord the sampleThe category j with the largest number of predicted times and the number m,is a sampleProbability of being predicted as j:
the uncertainty of the prediction is the smallest when the n predictions are all the same, and the time is the smallest when the n predictions are all the same,(ii) a The uncertainty is greatest when the n predictions are all different, when,And n is 10.
In this embodiment, a BCNN network is used for training, in the training, clean samples Dc are first selected, label smoothing is helpful for the network to learn in noisy data, in the previous n training processes, the output of the Softmax layer is smoothed, and back propagation is performed with the following loss:
TrainingAfter n times, reusable samples Dr are sorted out using equation (6.6) and usedUpdating the network:
j is the category with the largest predicted times in the continuous n prediction processes.
The algorithm flow of the invention is as follows:
inputting:
training set D
Mini-batch training set Dm
Clean sample set Dc
Reusable sample set Dr
Total number of training times Tmax
Number of pre-training times Tk
Number of iterations Nmax
And (3) outputting: updating the network h
Randomly initializing network parameters Dc = D, dr = D
The clean sample Dc is selected according to equation (6.5)
Updating the network h according to equation (6.11)
else:
Selecting a clean sample Dc according to formula (6.5)
Selecting reusable samples Dr according to formula (6.6)
Updating the network h according to equation (6.14)
Generally speaking, the invention mainly combines the sample selection with the loss correction, and provides a new method for selecting reusable samples, so that the method improves the sample utilization rate and improves the fine-grained classification performance.
Although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments described in the foregoing embodiments, or equivalents may be substituted for elements thereof.
Claims (5)
1. A network image label denoising method combining sample selection and label correction is characterized by comprising the following steps:
s1, firstly, selecting a clean sample according to cosine similarity between the sample and a category center;
s2, selecting a reusable sample from the rest samples through sample uncertainty dynamic state and correcting the reusable sample;
s3, finally, updating the network by using the clean sample and the corrected reusable sample;
in S1, the features of the picture are normalized in the Softmax layer, and the output process of the Softmax layer may be represented as:
w i x i =||w i ||||x i ||cosθ i =cosθ i (6.2)
after normalization, using a hyper-parameter s to scale cosine values, the Softmax output under the L2 constraint after feature normalization is calculated as follows:
wherein x is i And y i Represents the ith sample and its label;
after normalization, the characteristics are distributed on the hypersphere in an angle, and the parameter w of the last full-connection layer j For the center of each class generated by pre-training, the output of the network full-link layer is the cosine distance cos theta between the picture feature and the center of each class j (ii) a Recording the cosine similarity of each picture with its corresponding class center:
is the ith sample and its class center y i The cosine distances of the training data are used for sorting H, and the instances with high cosine similarity are taken from the training of each batch and sent to a peer-to-peer network for the next training; the selection formula is as follows:
where τ is a correctable discard rate, D is a set of samples, and Dr is a reusable sample.
2. The method for denoising network image labels combining sample selection and label correction as claimed in claim 1, wherein the clean samples Dc are selected in S1, and the remaining samples can be divided into two types, which are respectively reusable samples Dr and noise sets Dn, and need to be discarded in the subsequent training;
when the sample x i Prediction uncertainty of f (x) i ) The sample belongs to the reusable sample set Dr if the following condition is satisfied:
wherein f (x) i ) Is a sample x i And midf (x) i ) Is shown (D) r ∩D n ) Median of uncertainty of samples, cross entropy is used to measure the uncertainty of each sample:
3. the method for denoising network image labels combining sample selection and label correction as claimed in claim 2, wherein each sample x is recorded i Prediction Pre of the last 10 times i Prediction is updated as training progresses:
Pre i ={prec 1 ,prec 2 ,...,prec n } (6.8)
according to Pre i Record sample x i Class j with the largest predicted number of times and number m, p i Is a sample x i Probability of being predicted as j:
p i =m/n (6.9)
of the n predictions, where p is the same, the uncertainty is minimal when all n predictions are the same i =1,f(x i ) =0; the uncertainty is greatest when n predictions are different, when p is the time of prediction i =1/n,f(x i ) And (4) taking 10 as n, wherein the = log 1/n.
4. The method for denoising network image labels combining sample selection and label correction as claimed in claim 3, wherein in S3, the output of Softmax layer is smoothed in the previous n training processes, and the backward propagation is performed with the following loss:
where α is the tag smoothing factor for the data set.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211408454.3A CN115564960B (en) | 2022-11-10 | 2022-11-10 | Network image label denoising method combining sample selection and label correction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211408454.3A CN115564960B (en) | 2022-11-10 | 2022-11-10 | Network image label denoising method combining sample selection and label correction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN115564960A CN115564960A (en) | 2023-01-03 |
CN115564960B true CN115564960B (en) | 2023-03-03 |
Family
ID=84769821
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211408454.3A Active CN115564960B (en) | 2022-11-10 | 2022-11-10 | Network image label denoising method combining sample selection and label correction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115564960B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111414942A (en) * | 2020-03-06 | 2020-07-14 | 重庆邮电大学 | Remote sensing image classification method based on active learning and convolutional neural network |
CN113657449A (en) * | 2021-07-15 | 2021-11-16 | 北京工业大学 | Traditional Chinese medicine tongue picture greasy classification method containing noise labeling data |
CN113657561A (en) * | 2021-10-20 | 2021-11-16 | 之江实验室 | Semi-supervised night image classification method based on multi-task decoupling learning |
CN114169442A (en) * | 2021-12-08 | 2022-03-11 | 中国电子科技集团公司第五十四研究所 | Remote sensing image small sample scene classification method based on double prototype network |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114492574A (en) * | 2021-12-22 | 2022-05-13 | 中国矿业大学 | Pseudo label loss unsupervised countermeasure domain adaptive picture classification method based on Gaussian uniform mixing model |
CN114897049A (en) * | 2022-04-06 | 2022-08-12 | 济南融瓴科技发展有限公司 | Label noise monitoring method based on meta-learning |
CN115170813A (en) * | 2022-06-30 | 2022-10-11 | 南京理工大学 | Network supervision fine-grained image identification method based on partial label learning |
-
2022
- 2022-11-10 CN CN202211408454.3A patent/CN115564960B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111414942A (en) * | 2020-03-06 | 2020-07-14 | 重庆邮电大学 | Remote sensing image classification method based on active learning and convolutional neural network |
CN113657449A (en) * | 2021-07-15 | 2021-11-16 | 北京工业大学 | Traditional Chinese medicine tongue picture greasy classification method containing noise labeling data |
CN113657561A (en) * | 2021-10-20 | 2021-11-16 | 之江实验室 | Semi-supervised night image classification method based on multi-task decoupling learning |
CN114169442A (en) * | 2021-12-08 | 2022-03-11 | 中国电子科技集团公司第五十四研究所 | Remote sensing image small sample scene classification method based on double prototype network |
Also Published As
Publication number | Publication date |
---|---|
CN115564960A (en) | 2023-01-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110532880B (en) | Sample screening and expression recognition method, neural network, device and storage medium | |
CN109214353B (en) | Training method and device for rapid detection of face image based on pruning model | |
CN111506773B (en) | Video duplicate removal method based on unsupervised depth twin network | |
CN113688949B (en) | Network image data set denoising method based on dual-network joint label correction | |
CN113627479B (en) | Graph data anomaly detection method based on semi-supervised learning | |
CN114743037A (en) | Deep medical image clustering method based on multi-scale structure learning | |
CN107229945A (en) | A kind of depth clustering method based on competition learning | |
CN114330598A (en) | Multi-source heterogeneous data fusion method and system based on fuzzy C-means clustering algorithm | |
CN116258978A (en) | Target detection method for weak annotation of remote sensing image in natural protection area | |
CN114881125A (en) | Label noisy image classification method based on graph consistency and semi-supervised model | |
CN115051929A (en) | Network fault prediction method and device based on self-supervision target perception neural network | |
CN115564960B (en) | Network image label denoising method combining sample selection and label correction | |
CN114842371A (en) | Unsupervised video anomaly detection method | |
CN113192627A (en) | Patient and disease bipartite graph-based readmission prediction method and system | |
CN116662832A (en) | Training sample selection method based on clustering and active learning | |
CN111008940A (en) | Image enhancement method and device | |
CN115578568A (en) | Noise correction algorithm driven by small-scale reliable data set | |
CN115984682A (en) | Fish quantity estimation method based on Unet and BP neural network | |
Cai et al. | SSS-Net: A shadowed-sets-based semi-supervised sample selection network for classification on noise labeled images | |
CN115017988A (en) | Competitive clustering method for state anomaly diagnosis | |
CN114677535A (en) | Training method of domain-adaptive image classification network, image classification method and device | |
CN110188219B (en) | Depth-enhanced redundancy-removing hash method for image retrieval | |
CN116778968B (en) | Heart sound classifying method based on depth separable convolution and attention mechanism | |
Pei et al. | Evidential Multi-Source-Free Unsupervised Domain Adaptation | |
CN111291602A (en) | Video detection method and device, electronic equipment and computer readable storage medium |
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 |