CN113516207B - Long-tail distribution image classification method with noise label - Google Patents
Long-tail distribution image classification method with noise label Download PDFInfo
- Publication number
- CN113516207B CN113516207B CN202111059448.7A CN202111059448A CN113516207B CN 113516207 B CN113516207 B CN 113516207B CN 202111059448 A CN202111059448 A CN 202111059448A CN 113516207 B CN113516207 B CN 113516207B
- Authority
- CN
- China
- Prior art keywords
- data
- noise
- sample
- loss
- relaxation
- 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
- 238000009826 distribution Methods 0.000 title claims abstract description 41
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000012549 training Methods 0.000 claims abstract description 42
- 230000001419 dependent effect Effects 0.000 claims abstract description 24
- 238000012216 screening Methods 0.000 claims abstract description 11
- 238000005070 sampling Methods 0.000 claims description 15
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000010438 heat treatment Methods 0.000 claims 1
- 230000006870 function Effects 0.000 abstract description 7
- 230000000694 effects Effects 0.000 abstract description 6
- 230000007246 mechanism Effects 0.000 abstract description 2
- 230000007704 transition Effects 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 230000003416 augmentation Effects 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000013434 data augmentation Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000002040 relaxant effect Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- QVRVXSZKCXFBTE-UHFFFAOYSA-N n-[4-(6,7-dimethoxy-3,4-dihydro-1h-isoquinolin-2-yl)butyl]-2-(2-fluoroethoxy)-5-methylbenzamide Chemical compound C1C=2C=C(OC)C(OC)=CC=2CCN1CCCCNC(=O)C1=CC(C)=CC=C1OCCF QVRVXSZKCXFBTE-UHFFFAOYSA-N 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010845 search algorithm Methods 0.000 description 1
- 238000012358 sourcing Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a long-tail distribution image classification method with a noise label, which is used for learning through relaxation interval loss depending on a sample and assisting with an anti-noise data enhancement strategy and solving the problem of image classification with a long-tail characteristic and a noise label at the same time. According to the data noise characteristics, introducing a relaxation variable dependent on the sample when calculating the sample function interval to relax the interval constraint, and calculating the smooth relaxation loss dependent on the sample according to the sample interval classification; according to the data long tail characteristics, a data enhancement strategy adjusted in stages is implemented, samples are enhanced strongly and weakly respectively, and a sample screening mechanism based on relaxation loss is provided in the formal training stage for screening out noise data. The method is simple and convenient to implement, flexible in means, and capable of remarkably improving the classification effect on long-tail data, noise data and training data with the characteristics of the long-tail data and the noise data.
Description
Technical Field
The invention relates to the field of image classification, in particular to a method for classifying images under noise labels and long tail distribution data.
Background
In recent years, Convolutional Neural Networks (CNNs) have been widely used in the field of computer vision. In the case of a fixed amount of training data, the overfitting phenomenon is increasingly prominent due to the increase of the number of parameters, and the requirement for accurately labeling data is also increased in order to improve the overall performance. However, obtaining a large number of accurately labeled samples is often quite expensive. In this regard, non-expert crowd-sourcing or systematic tagging is a practical solution, however this easily leads to mislabeling of tags. Many reference datasets, such as ImageNet, CIFAR-10/-100, MNIST, QuickDraw, etc., contain 3% to 10% noise label samples. Existing research on noisy labels has generally focused on splitting correctly labeled and incorrectly labeled samples, but neglecting the distribution of the data. In the real world, data often presents the characteristic of long tail distribution, several main categories in the data set occupy the dominance, and the data of other categories is insufficient in quantity. Therefore, in the current image classification task based on the deep neural network, how to classify data with long tail features and noise labels simultaneously to reduce the influence of the noise labels under long tail distribution is very important in practical application.
Disclosure of Invention
In order to solve the defects of the prior art and achieve the purpose of reducing the influence of noise labels under long tail distribution, the invention adopts the following technical scheme:
a method for classifying long-tail distribution images with noise labels comprises the following steps:
s1, according to the data noise characteristics, each sample image and the noise label thereofAt sample intervalsOn the basis of (2), introducing a relaxation variableForming sample relaxation intervals of noisy samples;
The sample interval isClass interval ofWhereinIs shown asA sampleIs marked with a labelIs a categoryI.e. specimenBelong to the categoryAnd the process of, accordingly,indicates all belong to the categoryA set of sequence numbers of samples of (1);
the sample relaxation intervals were:
wherein,indicating the sample image and its correct label,representing a prediction function for predicting to which class a sample image belongs,in order to be a sample space, the sample space,Nis the total number of samples and is,is composed ofA set of tags for each of the categories,the representation of the real number field is performed,is shown anddifferent noise labelsAnd x corresponding thereto, the largest value among the values obtained by the prediction function,,representing an optimal interval; the traditional DNN classification network usually follows a linear conversion layer after a feature extractor, but the strategy is easy to generate the situation that the classifier falls into linear inseparability when fitting data with noise, so the invention proposes a relaxation variableIntroducing slack variables with relaxed spacing constraintsSample relaxation interval ofThe tolerance of classification prediction results is increased;
according to the sample interval, the smooth relaxation Loss (Slack Loss) depending on the sample is calculated in a segmented mode;
S2, according to the data long tail characteristics, implementing the divisionData Augmentation strategy for stage adjustment for noisy Data setsEach group of samples in (1)For sample imageRespectively carrying out weak data enhancement and strong data enhancement to obtain corresponding weak enhancement data and strong enhancement data, dividing training into a preheating stage and a formal stage, considering the negative influence of a strong data enhancement method on a high noise rate data set, respectively calculating and adding relaxation loss in the training stage by using the weak enhancement data and the strong enhancement data, and calculating and adding the relaxation loss in the noise rateAndas the weight, in the preheating stage, directly calculating the relaxation loss of the weak enhancement data and the strong enhancement data; in the formal training stage, a group of sample images are screened and relaxed to be used as pure data according to the relaxation loss in the preheating stage, residual noise data are screened out, and the relaxation loss is calculated. The method of injecting strong data enhancement during the warm-up phase of training may improve performance for training of low noise data sets, but is counterproductive as the noise of the data set increases. Conversely, the weak data enhancement during the warm-up phase can greatly improve the performance of the high noise data training. Based on this summary, the present invention divides model training into two phases, adjusting the enhancement strategy at different phases.
Further, the relaxation loss in S1 is:
further onThe warm-up stage in S2, using weak enhancement data directlyAnd strong enhancement dataCalculating the relaxation loss as the noise rateAndas weights, the overall loss is calculated:
further, the formal training phase in S2 includes the following steps:
s21, screening out the slack loss according to the slack loss in the preheating stage、As weak enhancement dataAnd strong enhancement dataFront of minimum medium slack lossA partial sample image;
s22, according toWeak enhanced data after screeningFrom strong enhancement dataObtained by intermediate samplingAccording to the screened strong enhancement dataFrom weak enhancement of dataObtained by intermediate samplingScreening out the remaining noise data;
s23, obtaining、As correct sample image, at the noise rateAndas weight, calculating the overall loss, returning the loss, updating network parameters:
further, in the above S1, an optimum interval is set、For training data pointsSample intervalGreater than the optimum intervalTherefore, it needs to be pushed to the class boundary to make the data boundary more gradual; for the sample interval atData points within the interval,In the opposite direction, so that the data point has a certain probability of turning into the other side of the class boundary;、indicating for a categoryAndis not an exact formula but is stated to be inversely proportional to the number of samples corresponding to the class in view of the relationship between the two classesAndis/are as followsTo the power. Thereby setting the sample-dependent tolerance range.
Further, the relaxation variable in S1Will be uniformly distributedMultiplication byFrom which the slack variable is extractedI.e. by,Representing the noise rate, i.e., the probability of sample label error.
Further, for the setting of the long tail distribution data, the total number of samples isIn each category of training dataThe number of training samples isSatisfy the following requirementsThe ratio of the most sample number class to the least sample number class is used as an imbalance factor (imbalance factor)I.e. by。
Further, the sample image and its noise label in S1By transition matrix (transition matrix)Representation represents a noise label:
wherein,representing a sample imageThe corresponding category of the content file,is shown asnThe number of images of the sample is determined,representing categoriesIs classified into categoriesjThe probability of (a) of (b) being,. For the setting of the noise data, there are 2 cases, i.e., class-independent noise (class-dependent noise) and class-dependent noise (class-dependent noise). The category-independent noise assumes that the mislabeled samples are randomly and uniformly distributed, and the category-dependent noise focuses on the phenomenon of artificial labeling error caused by visual similarity. Both types of noise distributions may use a transition matrixAnd (4) showing.
Further, the sample image and its noise label in S1Sampled in noisy data setsCorresponding to the sample image and its correct labelSampled in a clean data setWhereinIs shown asnThe number of images of the sample is determined,representing a sample imageThe corresponding category of the content file,as to the number of samples,average sampling from potential distribution of data。
The invention has the advantages and beneficial effects that:
starting from the category correlation interval, the method introduces the relaxation variable depending on the sample, relaxes the interval constraint, and increases the tolerance of the classification prediction result, thereby bearing the risk of wrong classification caused by noise or unbalanced distribution; considering the negative influence of a strong data enhancement method on a high-noise-rate data set, the method calculates the relaxation loss in the training stage by using weak enhancement data and strong enhancement data respectively; the method for injecting strong data enhancement in the preheating stage of training can improve the performance of the training of low-noise data sets, but can be counterproductive when the noise of the data sets is increased, and on the contrary, the weak data enhancement in the preheating stage can greatly improve the performance of the training of high-noise data. Finally, the effect of noise signatures under long tail distributions is reduced.
Drawings
FIG. 1a is a graph of accuracy versus loss variation for noise sample learning on a CIFAR-10 data set.
FIG. 1b is a graph of accuracy versus loss variation for long tail distribution learning on a CIFAR-10 dataset.
Fig. 2c is a distribution diagram of class independent noise under a long tail distribution.
Fig. 2d is a distribution diagram of class-correlated noise under a long-tailed distribution.
FIG. 3 is a graph of sample dependence tolerance in the present invention.
Detailed Description
The following detailed description of embodiments of the invention refers to the accompanying drawings. It should be understood that the detailed description and specific examples, while indicating the present invention, are given by way of illustration and explanation only, not limitation.
Noise label learning has received a lot of attention in recent years and also has achieved surprising effects. However, existing deep neural networks DNN still have drawbacks in addressing noise labeling and long tail learning. As shown in fig. 1a, in which the noise ratio is symmetricalWhen DNN is used to fit the noise label, the fluctuation in validation accuracy accounts for the noise capacity of the model. As shown in fig. 1b, in which the imbalance factorThe application of DNN in long tail distribution learning also embodies similar characteristics, i.e. DNN fits the main class first, then gradually the tail class. From the above analysis, it can be found that the contradiction between the few-sample class learning and the noise fitting confounds the prediction of the neural network, giving the noise label learning band under the long tail distributionNew challenges arise.
Deep Neural Networks (DNNs), during training with noisy data, tend to remember common patterns first, and then fit noise samples step by step. A similar process occurs in class imbalance learning, where the network tends to fit the main class first, then over-fit the tail class step by step. In this regard, the present invention starts with class-aware intervals (class-dependent margin), and introduces a sample-dependent relaxation Variable (Slack Variable) for taking up the risk of misclassification due to noise or unbalanced distribution. In addition, the invention also provides an anti-noise data enhancement strategy.
1. Experimental setup and preparation
The invention mainly solves the problem of classifying data with long tail characteristics and noise labels simultaneously in an image classification task. Defining an input space as a set of input spaces,Is as followsnAn input image is input to the image processing device,Nis the total number of samples and is,the label sets of individual categories arePotential distribution of data. Training data pairsSample at,Is the total number of samples and is,representing an input imageA corresponding category. The image classification task designed in the present invention therefore needs to be aimed at by constrainingDeriving a prediction functionIndicating to inputTo proceed withCalculating and outputting classification resultsSo thatThe number of (a) is the largest, in short,predicting each input imageTo which category the prediction result belongs, and outputting the prediction result;The optimization goal is to maximize the correct number of prediction results,representing a real number domain.
For the setting of the long tail distribution data, the total number of samples is set asIn each category of training dataIs defined as the number of training samplesSatisfy the following requirements. The invention defines the sample number ratio between the most sample number category and the least sample number category as an unbalance factor (imbalance factor)I.e. by. As shown in fig. 2a, 2b, the distribution of the long tail data generally follows an exponential decay.
For the setting of the noise data, there are 2 cases, i.e., class-independent noise (class-dependent noise) and class-dependent noise (class-dependent noise). The category-independent noise assumes that the mislabeled samples are randomly and uniformly distributed, and the category-dependent noise focuses on the phenomenon of artificial labeling error caused by visual similarity. Both types of noise distribution may use a transition matrix (transition matrix)Is shown, each element thereinRepresentative categoriesIs classified into categoriesThe probability of (c). Corresponding correct sample and label thereofSampling in clean data setsSamples representing data and noise labels thereforSampling in noisy data setsWhereinAs to the number of samples,can be defined as:
as shown in FIGS. 2a and 2b, the class-independent noise, or symmetric noise, is assumed to be of a certain classAre evenly distributed over the other classes, i.e.,Representing the noise rate, i.e., the probability of sample label error; while class-dependent noise, or asymmetric noise, assumes a certain oneCategoriesAre all error labeled as another specific categoryI.e. by. As shown in fig. 2c and 2d, the setting of noise data under a long tail distribution.
2. Sample dependent tolerance range setting
Conventional DNN classification networks typically follow the feature extractor with a linear conversion layer, however this strategy tends to produce cases where the classifier falls into linearity inseparability when fitting to noisy data. Therefore, the present invention proposes a relaxation variableTo relax the interval constraint and increase the tolerance of classification. Relaxation variablesAt corresponding optimum intervals empiricallyTo be restricted, i.e.(ii) a For data with noisy samples, there is a noise due to each sampleThe probability of (noise rate) being a wrong sample, so here we will distribute uniformlyMultiplication by noise rateFrom which the slack variable is extractedI.e. by。
the relaxation interval increases the tolerance of the classification prediction results. As shown in FIG. 3, the optimum interval is set、For training data pointsInterval of functionGreater than the optimum intervalTherefore, it needs to be pushed to the class boundary to make the data boundary more gradual; for the function interval atData points within the interval,In the opposite direction, so that the data point has a certain probability of turning into the other side of the class boundary;、relative to that in the above theoretical calculationIn the present embodiment, the description is given for the categoryAndis not an exact formula but specifies that they are inversely proportional to the number of samples of the class in view of the relationship between the two classesAndis/are as followsTo the power.
3. Loss of slack space
4. weak data enhancement strategy and strong data enhancement strategy
The present invention relates to 2 data enhancement strategies, namely weak data enhancement and strong data enhancement. The implementation of weak data enhancement (weak augmentation) is simple random flip (flip) and crop (crop), while strong data enhancement (strong augmentation) uses the implementation of AutoAutoAutoAutoaugmentation and adopts a data enhancement strategy automatically selected by a search algorithm on ImageNet.
5. Anti-noise data enhancement strategy implemented in stages
The method of injecting strong data enhancement during the warm-up phase of training may improve performance for training of low noise data sets, but is counterproductive as the noise of the data set increases. Conversely, the weak data enhancement during the warm-up phase can greatly improve the performance of the high noise data training. Based on this summary, the present invention divides model training into two phases, adjusting the enhancement strategy at different phases. In the warm-up phase, weak enhancement data is directly usedAnd strong enhancement dataCalculating the loss, namely:
in the formal training stage, the proportion of the screening quantity to the total quantity of the samples is thatOf (2) a sampleThe remaining noise data is filtered out as "correct samples" for calculating the loss and update parameters, defining the loss as:
6. Training phase data screening strategy
The invention screens the data in the formal training phase and has the noise rate ofThe training data of (1) is selected as the proportion of the total number of samplesEffective sample ofAs "correct samples" for calculating the loss and update parameters, the remaining noise data is filtered out. The screening process firstly separatelyIs defined as:
namely, it isRespectively representing weakly enhanced dataAnd strong enhancement dataFront of minimum medium slack lossA portion of the sample. Then, according to the screened weak enhancement dataFrom strong enhancement dataObtained by intermediate sampling(ii) a Similarly, according to the screened strong enhancement dataFrom weak enhancement of dataObtained by intermediate samplingThe remaining noise data is filtered out.
Specifically, the sample-dependent relaxation interval loss learning method comprises the following steps:
step 1: according to the data noise characteristics, each sample and the noise label thereofAfter calculating the sample function interval (functional margin)When a sample dependent relaxation Variable (Slack Variable) is introducedAnd calculating the sample-dependent smooth relaxation Loss (Slack Loss) according to the sample interval by sections by relaxing the interval constraint。
Data samples and noise signatures thereofSampling in noisy data setsCorresponding to correct sample and label thereofSampling in clean data setsWhereinAs to the number of samples,a noise rate representing the probability of sample label error,average sampling from potential distribution of data,In order to input the space, the input device is provided with a display,is composed ofA set of labels for each category.
at the same time, classifyIs defined asWhereinIs shown asSample (serial numberSample (1)Is marked with a labelIs a categoryI.e. specimenBelong to the categoryAnd the process of, accordingly,indicates all belong to the categoryA set of sequence numbers of samples of (1).
Relaxed interval for a particular noise sampleAt sample intervalsOn the basis of (2), introducing a relaxation variableThe relaxation interval is defined as:
relaxation variance of sampleAt corresponding optimum intervals empiricallyTo be restricted, i.e.(ii) a For noise rate ofWill be evenly distributedMultiplication by noise rateFrom which the slack variable is extractedI.e. by。
step 2: according to the Data long tail characteristic, a Data Augmentation strategy (Data Augmentation) adjusted in stages is implemented. And respectively carrying out strong data enhancement and weak data enhancement on the sample. In the preheating stage, directly calculating relaxation loss; in the formal training phase, a mechanism is provided to screen small loss samples as clean data, to screen out noisy data, and to calculate slack loss.
For noisy data setsEach group of samples in (1)To input ofRespectively carrying out weak data enhancement and strong data enhancement to obtain corresponding weak enhancement dataAnd strong enhancement data。
Considering the negative effect of strong data enhancement method on high noise rate data set, the invention uses the relaxation loss of training stage with weak enhancement data respectivelyAnd strong enhancement dataAre calculated and added to obtain the noise ratioAndas weights, the penalty is defined as:
The training is divided into a preheating stage and a formal stage, and the loss and the parameters are calculated and updated in the following modes respectively:
2.1: in the warm-up phase, weak enhancement data is directly usedAnd strong enhancement dataCalculating the loss, namely:
2.2: in the formal training stage, according to the relaxation loss of the samples, the proportion of the screening quantity to the total amount of the samples isOf (2) a sampleThe remaining noise data is filtered out as "correct samples" for calculating the loss and update parameters, defining the loss as:
in the formal training stage, the sample is screenedIs first defined separatelyEnhancing data for weaknessesAnd strong enhancement dataFront of minimum medium slack lossA portion of the sample; then, according to the screened weak enhancement dataFrom strong enhancement dataObtained by intermediate sampling(ii) a Similarly, according to the screened strong enhancement dataFrom weak enhancement of dataObtained by intermediate samplingThe remaining noise data is filtered out. Using the obtained、And calculating the loss by using the formula, returning the loss and updating the network parameters.
As shown in Table 1, on the CIFAR-10 and CIFAR-100 data sets with noise, using ResNet34 as a general network framework, noise ratios were set to the class-independent noise and the class-dependent noise, respectivelyAndcompared with Bootstrap, Forward, GCE, SCE and other methods. For class independent noise, the relaxation penalty proposed herein outweighs all other approaches. For class-dependent noise, the relaxation penalty is inIs slightly better than other methods, but inThe accuracy is not high. To this end, we give a reasonable explanation: the relaxation variables introduced by the invention add random disturbance to the sample label distribution, and the random disturbance may have negative effects because the category-related noise is accurate in the non-corresponding categories. And when the noise rate is small () Regularization adjustment of relaxation loss may balance this negative effect.
As shown in Table 2, on CIFAR-10 and CIFAR-100 datasets with long tail distributions, using ResNet34 as a general network framework, an imbalance factor is setAnd Focal localMixup, CE-DRW, CE-DRS, LDAM-DRW, BBN, etc. It can be seen that when the data is extremely unbalanced, the accuracy of the classification result of the invention is much higher than that of other methods. When in useThe performance of relaxation loss is somewhat inadequate for reasons similar to the interpretation of noise learning in the previous section.
TABLE 1
Different methods classify the results in accuracy (%) on the noisy data set CIFAR-10/100, with the highest accuracy being marked in bold and the second accuracy being marked in oblique bold. Wherein the Slack Loss method uses the Slack Loss (Slack Loss) in the present invention as a Loss function, but does not use a data enhancement strategy; the Slack Loss + method uses the relaxation Loss and data enhancement strategy of the present invention, i.e., the complete method given by the present invention.
TABLE 2
The classification result accuracy (%) of the different methods on the long tail data set CIFAR-10/100, the highest accuracy being marked in bold, the second accuracy being marked in oblique bold. Wherein the Slack Loss method uses the Slack Loss (Slack Loss) in the present invention as a Loss function.
The above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.
Claims (10)
1. A long-tail distribution image classification method with a noise label is characterized by comprising the following steps:
s1, according to the data noise characteristics, the sample image and the noise label thereofAt sample intervalsOn the basis of (2), introducing a relaxation variableForming sample relaxation intervals of noisy samples;
The sample interval isClass interval ofWhereinIs shown asA sampleIs marked with a labelIs a category,Indicates all belong to the categoryA set of sequence numbers of samples of (1);
the sample relaxation intervals were:
wherein,indicating the sample image and its correct label,representing a prediction function for predicting to which class a sample image belongs,in order to be a sample space, the sample space,Nis the total number of samples and is,is composed ofA set of tags for each of the categories,the representation of the real number field is performed,is shown anddifferent noise labelsAnd x corresponding thereto, the largest value among the values obtained by the prediction function,,representing an optimal interval;
S2, according to the data long tail characteristic, the data enhancement strategy adjusted by stages is used for the sample imageRespectively performing weak data enhancement and strong data enhancement to obtain corresponding weak enhancement data and strong enhancement data, dividing training into a preheating stage and a formal stage, and directly calculating relaxation losses of the weak enhancement data and the strong enhancement data in the preheating stage; in the formal training stage, a group of sample images are screened and relaxed to be used as pure data according to the relaxation loss in the preheating stage, residual noise data are screened out, and the relaxation loss is calculated.
3. the method for classifying long tail distribution images with noise labels as claimed in claim 1, wherein the pre-heating stage in S2 directly uses weak enhancement dataAnd strong enhancement dataCalculating the relaxation loss as the noise rateAndas weights, the overall loss is calculated:
4. the method for classifying long-tail distribution images with noise labels as claimed in claim 1, wherein the formal training phase in S2 includes the following steps:
s21, screening out the slack loss according to the slack loss in the preheating stage、As weak enhancement dataAnd strong enhancement dataFront of minimum medium slack lossA partial sample image;
s22, according to the screened weak enhancement dataFrom strong enhancement dataObtained by intermediate samplingAccording to the screened strong enhancement dataFrom weak enhancement of dataObtained by intermediate samplingScreening out the remaining noise data;
s23, obtaining、As correct sample image, at the noise rateAndas weights, the overall loss is calculated:
6. the method for classifying long tail distribution images with noise labels as claimed in claim 1, wherein in S1, an optimal interval is set、For training data pointsSample intervalGreater than the optimum intervalPushing it to the class boundary, making the data boundary more gradual; for the sample interval atData points within the interval,In the opposite direction, so that the data point has a certain probability of turning into the other side of the class boundary;、indicating for a categoryAndis inversely proportional to the number of samples corresponding to the classAndis/are as followsTo the power.
7. The method for classifying long tail distribution images with noise labels as claimed in claim 1, wherein the relaxation variables in S1Will be uniformly distributedMultiplication byFrom which the slack variable is extractedI.e. by,Representing the noise rate, i.e., the probability of sample label error.
9. The method for classifying long tail distribution images with noise labels as claimed in claim 1, wherein the sample image and its noise label in S1By means of a transfer matrixRepresents the noise label:
10. the method for classifying long tail distribution images with noise labels as claimed in claim 1, wherein the sample image in S1And its noise labelSampled in noisy data setsCorresponding to the sample image and its correct labelSampled in a clean data setWhereinIs shown asnThe number of images of the sample is determined,representing a sample imageThe corresponding category of the content file,as to the number of samples,average sampling from potential distribution of data。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111059448.7A CN113516207B (en) | 2021-09-10 | 2021-09-10 | Long-tail distribution image classification method with noise label |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111059448.7A CN113516207B (en) | 2021-09-10 | 2021-09-10 | Long-tail distribution image classification method with noise label |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113516207A CN113516207A (en) | 2021-10-19 |
CN113516207B true CN113516207B (en) | 2022-01-25 |
Family
ID=78063294
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111059448.7A Active CN113516207B (en) | 2021-09-10 | 2021-09-10 | Long-tail distribution image classification method with noise label |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113516207B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113989905A (en) * | 2021-11-16 | 2022-01-28 | 广东履安实业有限公司 | Training of face recognition model, face recognition method and related device |
CN113869463B (en) * | 2021-12-02 | 2022-04-15 | 之江实验室 | Long tail noise learning method based on cross enhancement matching |
CN114519850A (en) * | 2022-04-20 | 2022-05-20 | 宁波博登智能科技有限公司 | Marking system and method for automatic target detection of two-dimensional image |
CN114863193B (en) * | 2022-07-07 | 2022-12-02 | 之江实验室 | Long-tail learning image classification and training method and device based on mixed batch normalization |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102945372B (en) * | 2012-10-18 | 2015-06-24 | 浙江大学 | Classifying method based on multi-label constraint support vector machine |
CN109543693B (en) * | 2018-11-28 | 2021-05-07 | 中国人民解放军国防科技大学 | Weak labeling data noise reduction method based on regularization label propagation |
CN111737552A (en) * | 2020-06-04 | 2020-10-02 | 中国科学院自动化研究所 | Method, device and equipment for extracting training information model and acquiring knowledge graph |
CN111832627B (en) * | 2020-06-19 | 2022-08-05 | 华中科技大学 | Image classification model training method, classification method and system for suppressing label noise |
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 |
-
2021
- 2021-09-10 CN CN202111059448.7A patent/CN113516207B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN113516207A (en) | 2021-10-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113516207B (en) | Long-tail distribution image classification method with noise label | |
Wang et al. | Esrgan: Enhanced super-resolution generative adversarial networks | |
CN111738301B (en) | Long-tail distribution image data identification method based on double-channel learning | |
CN106779064A (en) | Deep neural network self-training method based on data characteristics | |
CN112101544A (en) | Training method and device of neural network suitable for long-tail distributed data set | |
CN109857871A (en) | A kind of customer relationship discovery method based on social networks magnanimity context data | |
CN104143081A (en) | Smile recognition system and method based on mouth features | |
CN107564007B (en) | Scene segmentation correction method and system fusing global information | |
CN116503676B (en) | Picture classification method and system based on knowledge distillation small sample increment learning | |
CN104809706A (en) | Single lens computational imaging method based on gentle image color change priori | |
CN105306296A (en) | Data filter processing method based on LTE (Long Term Evolution) signaling | |
Al-Amaren et al. | RHN: A residual holistic neural network for edge detection | |
CN113837959A (en) | Image denoising model training method, image denoising method and image denoising system | |
CN113011337A (en) | Chinese character library generation method and system based on deep meta learning | |
CN115810191A (en) | Pathological cell classification method based on multi-attention fusion and high-precision segmentation network | |
Xue et al. | Research on edge detection operator of a convolutional neural network | |
Fernandez-Fernandez et al. | Quick, stat!: A statistical analysis of the quick, draw! dataset | |
CN114492631A (en) | Spatial attention calculation method based on channel attention | |
CN116488974B (en) | Light modulation identification method and system combined with attention mechanism | |
Li et al. | Learning domain-aware detection head with prompt tuning | |
Zhou et al. | Ec-darts: Inducing equalized and consistent optimization into darts | |
CN117421657A (en) | Sampling and learning method and system for noisy labels based on oversampling strategy | |
CN117095217A (en) | Multi-stage comparative knowledge distillation process | |
CN111008940A (en) | Image enhancement method and device | |
CN115392344A (en) | Long tail identification method for strong and weak double-branch network with difficult sample perception |
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 |