CN113505120B - Double-stage noise cleaning method for large-scale face data set - Google Patents

Double-stage noise cleaning method for large-scale face data set Download PDF

Info

Publication number
CN113505120B
CN113505120B CN202111061863.6A CN202111061863A CN113505120B CN 113505120 B CN113505120 B CN 113505120B CN 202111061863 A CN202111061863 A CN 202111061863A CN 113505120 B CN113505120 B CN 113505120B
Authority
CN
China
Prior art keywords
noise
data set
samples
training
closed
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
Application number
CN202111061863.6A
Other languages
Chinese (zh)
Other versions
CN113505120A (en
Inventor
龚勋
陈锐
吴世杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southwest Jiaotong University
Original Assignee
Southwest Jiaotong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Southwest Jiaotong University filed Critical Southwest Jiaotong University
Priority to CN202111061863.6A priority Critical patent/CN113505120B/en
Publication of CN113505120A publication Critical patent/CN113505120A/en
Application granted granted Critical
Publication of CN113505120B publication Critical patent/CN113505120B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/21—Design, administration or maintenance of databases
    • G06F16/215—Improving data quality; Data cleansing, e.g. de-duplication, removing invalid entries or correcting typographical errors
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Quality & Reliability (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a double-stage noise cleaning method for a large-scale face data set, which adopts a double-stage strategy to enable a network to spontaneously detect a closed set noise sample and an open set noise sample in a training process, and finally reuses the closed set noise sample and deletes the open set noise sample from a training set. After the method is adopted to clean the data set, compared with an unwashed data set, the effect of the trained model is greatly improved in lfw precision, more than 90% of noise can be correctly identified through statistics, and a very good noise identification effect is realized. The model training results on the cleaned data set are also greatly superior to the training results on the uncleaned data set, which is proved in common test sets such as LFW, Age-DB and cfp-fp.

Description

Double-stage noise cleaning method for large-scale face data set
Technical Field
The invention relates to the technical field of noise cleaning of face data sets, in particular to a two-stage noise cleaning method of a large-scale face data set.
Background
Noise in a human face refers to samples with false labels that are generated when the face data set is collected and produced. The noise in the face data set includes open set noise and closed set noise. Closed set noise is also called label flipping, e.g., samples that are originally of class a are labeled with class B; the open set noise samples, which do not belong to any of the classes in the training set themselves, are labeled to a class in the training set.
Noisy data is an inevitable phenomenon in large-scale face data sets. The size of face data sets that are popular today has reached the tens of millions. If all the face data is manually cleaned, it takes extremely time and money, and noise cannot be completely eliminated. The noise samples can seriously damage the performance of the trained model, and therefore, a cleaning method which can provide effective and high recognition rate aiming at large-scale human face data noise is urgently needed.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a method for cleaning the large-scale face data set by using the two-stage noise, which adopts a two-stage strategy to enable a network to spontaneously detect closed set noise samples and open set noise samples in the training process, finally reuses the closed set noise samples and deletes the open set noise samples from the training set, thereby solving the problems in the background technology.
In order to achieve the purpose, the invention provides the following technical scheme: a method for cleaning double-stage noise of a large-scale face data set comprises the following steps:
s1, constructing an initial face data set D1;
s2, detecting the closed set noise samples and the corresponding real categories in the data set D1, and outputting a closed set noise list file containing closed set noise;
s3, completing the reuse of closed set noise to generate a data set D2;
s4, training is continued by taking the data set D2 as input, and an open set noise list file containing open set noise is output;
and S5, deleting the open set noise in the data set D2 according to the open set noise list file in the step S4, and finally generating a cleaned clean data set.
Preferably, the initial face data set D1 of step S1 is a labeled data set after face detection and alignment.
Preferably, the closed set noise samples and the corresponding true classes in the detection data set D1 in the step S2 are specifically: the data set D1 was input into a ResNet50 network and trained to detect closed-set noise using the BoundryF 1 Loss function.
Preferably, the formula of the boundaryF1 Loss function is as follows:
Figure 404675DEST_PATH_IMAGE001
Figure 774345DEST_PATH_IMAGE002
；
wherein the content of the first and second substances,
Figure 881978DEST_PATH_IMAGE003
indicating the number of samples in a batch,
Figure 480450DEST_PATH_IMAGE004
indicating the second within a batch
Figure 747352DEST_PATH_IMAGE004
The number of the samples is one,
Figure 773077DEST_PATH_IMAGE005
and
Figure 624796DEST_PATH_IMAGE006
all represent a category of a sample label, and
Figure 276227DEST_PATH_IMAGE007
and the first
Figure 148368DEST_PATH_IMAGE004
The labels of the individual samples correspond to each other,
Figure 469627DEST_PATH_IMAGE008
representing the total number of classes in the training set,
Figure 942066DEST_PATH_IMAGE009
in order to be a scaling factor, the scaling factor,
Figure 541062DEST_PATH_IMAGE010
represents all and
Figure 267710DEST_PATH_IMAGE004
the samples are of a different class and,
Figure 291029DEST_PATH_IMAGE011
in order to be a penalty term,
Figure 985185DEST_PATH_IMAGE012
is shown as
Figure 463571DEST_PATH_IMAGE004
The feature vector of each sample and
Figure 700517DEST_PATH_IMAGE007
the included angle between class center feature vectors of the classes;
Figure 756722DEST_PATH_IMAGE013
is shown as
Figure 688906DEST_PATH_IMAGE004
The feature vector of each sample and
Figure 361196DEST_PATH_IMAGE006
the included angle between class center feature vectors of the classes is represented as geodesic distance on the normalized hypersphere.
Preferably, the reuse of the closed set noise in step S3 is completed to generate the data set D2, specifically: according to the closed set noise list file output in the step S2, the closed set noise samples are moved to the corresponding category directory in the data set D1, the reuse of the closed set noise is completed, and the data set D2 is generated.
Preferably, the outputting of the open set noise list file containing the open set noise in step S4 is specifically: inputting the data set D2 into a ResNet50 network for continuous training, using a NoiseDropLoss loss function for training, saving the vector product of the feature layer and the weight center layer of each sample in the training process, and enabling the vector value corresponding to the class to which the sample belongs to be equal to
Figure 311703DEST_PATH_IMAGE014
Using a fixed size queue to keep the updates
Figure 676825DEST_PATH_IMAGE014
And discarding the last
Figure 830726DEST_PATH_IMAGE014
As training progresses, the final cohort is saved, the set of data is fitted by a gaussian mixture model, and the children are matchedThe number of the models is set to be 3, three parameters of Gaussian distribution are obtained, the average value of the second Gaussian distribution is used as a judgment threshold, samples lower than the threshold are regarded as open set noise samples, and the open set noise samples are output as open set noise list files.
Preferably, the probability function of the gaussian mixture model is as follows:
Figure 434270DEST_PATH_IMAGE015
；
wherein the content of the first and second substances,
Figure 114650DEST_PATH_IMAGE016
represents the data of one observation, and the data of one observation,
Figure 260461DEST_PATH_IMAGE017
all of the parameters of the representative model are,
Figure 416505DEST_PATH_IMAGE018
represents the first
Figure 774805DEST_PATH_IMAGE018
The sub-models are,
Figure 575271DEST_PATH_IMAGE019
represents the number of the sub-models,
Figure 155898DEST_PATH_IMAGE020
represents that the observed data belongs to
Figure 284391DEST_PATH_IMAGE018
The probability of a sub-model,
Figure 695650DEST_PATH_IMAGE021
represents the first
Figure 960409DEST_PATH_IMAGE018
A gaussian distribution density function for the submodels.
The noisledroploss function is as follows:
Figure 431711DEST_PATH_IMAGE022
Figure 172134DEST_PATH_IMAGE023
wherein the content of the first and second substances,Lrepresents the loss function for a batch of samples,Tis a modulation function that makes decisions based on d, assigning different values to different samples.
The invention has the beneficial effects that: the invention adopts a two-stage strategy, so that the network spontaneously detects closed set noise samples and open set noise samples in the training process, finally reuses the closed set noise samples, deletes the open set noise samples from the training set, and can improve the recognition rate of noise. The model training results on the cleaned data set are also greatly superior to the training results on the uncleaned data set, which is proved in common test sets such as LFW, Age-DB and cfp-fp.
Drawings
FIG. 1 is a schematic diagram of the process steps 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 derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, as shown in fig. 1, the method steps of the present invention are performed by a two-stage strategy, so that a network spontaneously detects a closed set noise sample and an open set noise sample during a training process, and finally reuses the closed set noise sample and deletes the open set noise sample from a training set.
This example performed experiments on a WebFace dataset that included 10572 categories, 490623 samples. The data set denoted as D1, which is a data set that has been labeled by human or machine, and that has been subjected to face detection and alignment, can be used directly to train a neural network model. Although labeling of sample labels is already performed, a human face data set is huge, and a large number of noise samples exist in the data set, so that the performance of a trained model is reduced, and therefore the data set needs to be cleaned of noise. In this embodiment, we replace part of samples with closed set noise, replace part of noise with open set noise, extract features using the ResNet50 model, and perform experiments.
In the first stage, closed set noise samples in the training set and their corresponding real classes are detected. Training is carried out by using ResNet50, BoundarryF 1 Loss detects the closed set noise, and finally a closed set noise list file containing the closed set noise is output. The formula for BoundaryF1 Loss is as follows:
Figure 137816DEST_PATH_IMAGE024
Figure 509279DEST_PATH_IMAGE026
；
wherein the content of the first and second substances,
Figure 636635DEST_PATH_IMAGE003
indicating the number of samples in a batch,
Figure 723408DEST_PATH_IMAGE004
indicating the second within a batch
Figure 492781DEST_PATH_IMAGE004
The number of the samples is one,
Figure 981400DEST_PATH_IMAGE007
and
Figure 279657DEST_PATH_IMAGE006
all represent a category of a sample label, and
Figure 119306DEST_PATH_IMAGE007
and the first
Figure 426791DEST_PATH_IMAGE004
The labels of the individual samples correspond to each other,
Figure 507267DEST_PATH_IMAGE008
representing the total number of classes in the training set,
Figure 242005DEST_PATH_IMAGE009
in order to be a scaling factor, the scaling factor,
Figure 303371DEST_PATH_IMAGE010
represents all and
Figure 804759DEST_PATH_IMAGE004
the samples are of a different class and,
Figure 753123DEST_PATH_IMAGE011
in order to be a penalty term,
Figure 908030DEST_PATH_IMAGE012
is shown as
Figure 207424DEST_PATH_IMAGE004
The feature vector of each sample and
Figure 640067DEST_PATH_IMAGE007
the included angle between class center feature vectors of the classes;
Figure 833151DEST_PATH_IMAGE013
is shown as
Figure 175271DEST_PATH_IMAGE004
The feature vector of each sample and
Figure 476808DEST_PATH_IMAGE006
the included angle between class center feature vectors of the classes is represented as geodesic distance on the normalized hypersphere.
According to the list generated in the previous training, the closed set noise samples are moved to the corresponding category list in the original training set (the sample of each category in the original data set is in a list corresponding to each category, therefore, if the closed set noise is in the list 1 and belongs to the 1 st category, but we detect that he should belong to the nth category, we move the closed set noise samples to the nth list (the absolute path where the monitored closed set noise samples are stored in the list file and the category corresponding to the monitored closed set noise samples)), so that the reuse of the closed set noise is completed, and the data set D2 is generated. The closed set noise is moved to a catalogue of a category corresponding to the closed set noise, and is not deleted, so that the closed set noise can be better utilized to train a model, and the method for processing the closed set noise is called the reuse of the closed set noise.
For the first stage, through our experiments, a noise recognition rate of 84% of the closed set noise can be achieved.
And in the second stage, inputting the data set D2 into a ResNet50 network for continuous training, using a NoiseDropLoss loss function for training, saving the vector product of the feature layer and the weight center layer of each sample in the training process, and enabling the vector value corresponding to the class to which the sample belongs to be equal to
Figure 54420DEST_PATH_IMAGE014
. Using a fixed size queue to keep the updates
Figure 242956DEST_PATH_IMAGE014
And discarding the last
Figure 262035DEST_PATH_IMAGE014
. As training progresses, the queue q under hold will reflect the state of the training. If the queue is drawnIn a histogram, the histogram will exhibit two peaks. At this time, the data are fitted by a gaussian mixture model, and the number of submodels is set to 3, so that three parameters with gaussian distribution are obtained. The probability function of the gaussian mixture model is as follows:
Figure 660655DEST_PATH_IMAGE027
wherein the content of the first and second substances,
Figure 917324DEST_PATH_IMAGE016
represents the data of one observation, and the data of one observation,
Figure 209634DEST_PATH_IMAGE017
all of the parameters of the representative model are,
Figure 283769DEST_PATH_IMAGE018
represents the first
Figure 779472DEST_PATH_IMAGE018
The sub-models are,
Figure 357609DEST_PATH_IMAGE019
represents the number of the sub-models,
Figure 379791DEST_PATH_IMAGE020
represents that the observed data belongs to
Figure 234615DEST_PATH_IMAGE018
The probability of a sub-model,
Figure 732461DEST_PATH_IMAGE021
represents the first
Figure 721146DEST_PATH_IMAGE018
A gaussian distribution density function for the submodels.
And taking the mean value of the second Gaussian distribution as a judgment threshold, taking the sample lower than the threshold as a noise sample, and outputting the noise sample to the open set noise list to realize the cleaning of the noise sample.
After the threshold d is obtained by using the above-mentioned gaussian mixture model, a noisdroploss function is used, which is formed by combining the following two formulas:
Figure 473201DEST_PATH_IMAGE028
Figure 16703DEST_PATH_IMAGE029
wherein the content of the first and second substances,Lrepresents the loss function for a batch of samples,Tis a modulation function based ondAnd judging, and assigning different values to different samples. Statistics shows that the method can correctly identify more than 90% of open set noise and greatly improve the training effect of the data set.
And cleaning the noise sample according to list after training. The whole process can be carried out again, and whether the detected noise quantity and the shape of the histogram reach a cleaner level is judged to decide whether to stop. In the second stage, the noise identification rate above the open set 98 can be realized. After the data set is cleaned, the effect of the trained model is improved by 1-2 percentage points on lfw compared with the effect of the unwashed data set, and can be 2-8 percentage points on other data sets such as Age DB, CFP-FP, CALFW, CPLFW and SLLFW.
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 various changes in the embodiments and/or modifications of the invention can be made, and equivalents and modifications of some features of the invention can be made without departing from the spirit and scope of the invention.

Claims (3)

1. A method for cleaning double-stage noise of a large-scale face data set is characterized by comprising the following steps:
s1, constructing an initial face data set D1;
s2, detecting the closed set noise samples and the corresponding real categories in the data set D1, and outputting a closed set noise list file containing closed set noise;
s3, completing the reuse of closed set noise to generate a data set D2;
s4, training is continued by taking the data set D2 as input, and an open set noise list file containing open set noise is output;
s5, deleting the open set noise in the data set D2 according to the open set noise list file in the step S4, and finally generating a cleaned clean data set;
the closed set noise samples and the corresponding true classes in the detection data set D1 in the step S2 are specifically: inputting a data set D1 into a ResNet50 network, training, and detecting closed set noise by using a BoundryF 1 Loss function;
the reuse of the closed set noise in step S3 is completed, and a data set D2 is generated, specifically: according to the closed set noise list file output in the step S2, moving the closed set noise samples to the corresponding category directories in the data set D1, completing the reuse of the closed set noise and generating a data set D2;
the step S4 of outputting the open set noise list file containing the open set noise specifically includes: inputting a data set D2 into a ResNet50 network for continuous training, using a NoiseDropLoss loss function for training, storing the vector product of a feature layer and a weight center layer of each sample in the training process, enabling the vector value corresponding to the class of the sample to be cos theta, continuously storing the latest cos theta and discarding the last cos theta by using a fixed-size queue, storing the final queue along with the training, fitting the group of data by a Gaussian mixture model, setting the number of submodels to be 3, obtaining three parameters of Gaussian distribution, taking the mean value of the second Gaussian distribution as a judgment threshold, regarding the samples lower than the threshold as open-set noise samples, and outputting the samples as open-set noise list files;
the BoundarryF 1 Loss function formula is as follows:
Figure FDA0003329808140000021
where if max{cos(θl+m)for all l≠yi}-cos(θyi)>0∶yi＝l；
where N represents the number of samples in a batch, i represents the ith sample in a batch, yiAnd j each represent a category of a certain sample label, and yiCorresponding to the label of the ith sample, n represents the total number of classes of the training set, s is a scaling factor, l represents all classes different from the ith sample, m is a penalty item, and thetayiThe feature vector representing the ith sample and the yiThe included angle between class center feature vectors of the classes; thetajRepresenting an included angle between the characteristic vector of the ith sample and the class center characteristic vector of the jth class, and reflecting the included angle as a geodesic distance on the normalized hypersphere;
the noisledroploss function is as follows:
Figure FDA0003329808140000022
Figure FDA0003329808140000023
wherein L represents the loss function of a batch of samples, T is a modulation function, and the decision is made based on d, assigning different values to different samples.
2. The method of two-stage noise cleaning of large-scale face data sets of claim 1, wherein: the initial face data set D1 of the step S1 is a data set to which labels have been labeled, and face detection and alignment have been performed.
3. The method of two-stage noise cleaning of large-scale face data sets of claim 1, wherein: the probability function of the Gaussian mixture model is as follows:
Figure FDA0003329808140000031
wherein x represents one-time observation data, theta represents all parameters of the model, K represents the kth sub-model, K represents the number of the sub-models, and alphakRepresents the probability that the observed data belongs to the kth sub-model, phi (x | theta)k) Representing the gaussian distribution density function of the kth sub-model.
CN202111061863.6A 2021-09-10 2021-09-10 Double-stage noise cleaning method for large-scale face data set Active CN113505120B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111061863.6A CN113505120B (en) 2021-09-10 2021-09-10 Double-stage noise cleaning method for large-scale face data set

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111061863.6A CN113505120B (en) 2021-09-10 2021-09-10 Double-stage noise cleaning method for large-scale face data set

Publications (2)

Publication Number Publication Date
CN113505120A CN113505120A (en) 2021-10-15
CN113505120B true CN113505120B (en) 2021-12-21

Family

ID=78016726

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111061863.6A Active CN113505120B (en) 2021-09-10 2021-09-10 Double-stage noise cleaning method for large-scale face data set

Country Status (1)

Country Link
CN (1) CN113505120B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114756541B (en) * 2022-05-25 2022-12-06 济南银华信息技术有限公司 Big data feature cleaning decision method and system for artificial intelligence training
CN115145904B (en) * 2022-07-06 2023-04-07 北京正远达科技有限公司 Big data cleaning method and big data acquisition system for AI cloud computing training

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101251855A (en) * 2008-03-27 2008-08-27 腾讯科技（深圳）有限公司 Equipment, system and method for cleaning internet web page

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108399608B (en) * 2018-03-01 2021-10-15 桂林电子科技大学 High-dimensional image denoising method based on tensor dictionary and total variation
CN110245550B (en) * 2019-04-22 2021-05-11 北京云识图信息技术有限公司 Human face noise data set CNN training method based on total cosine distribution
CN110674120B (en) * 2019-08-09 2024-01-19 国电新能源技术研究院有限公司 Wind farm data cleaning method and device
US20210089964A1 (en) * 2019-09-20 2021-03-25 Google Llc Robust training in the presence of label noise
CN110598037B (en) * 2019-09-23 2022-01-04 腾讯科技（深圳）有限公司 Image searching method, device and storage medium
US11599792B2 (en) * 2019-09-24 2023-03-07 Salesforce.Com, Inc. System and method for learning with noisy labels as semi-supervised learning
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
CN113361346B (en) * 2021-05-25 2022-12-23 天津大学 Scale parameter self-adaptive face recognition method for replacing adjustment parameters
CN113361201B (en) * 2021-06-10 2023-08-25 南京大学 Crowd-sourced acquired tag data cleaning method based on noise tag learning

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101251855A (en) * 2008-03-27 2008-08-27 腾讯科技（深圳）有限公司 Equipment, system and method for cleaning internet web page

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
孟晓超等.基于主动学习的标签噪声清洗方法.《陕西师范大学学报(自然科学版)》.2020,(第02期), *
潘巍等.一种有效的多数据源RFID冗余数据清洗技术.《西北工业大学学报》.2011,(第03期), *

Also Published As

Publication number Publication date
CN113505120A (en) 2021-10-15

Similar Documents

Publication Publication Date Title
CN110796186A (en) Dry and wet garbage identification and classification method based on improved YOLOv3 network
CN109190442B (en) Rapid face detection method based on deep cascade convolution neural network
CN111414862B (en) Expression recognition method based on neural network fusion key point angle change
CN109492026B (en) Telecommunication fraud classification detection method based on improved active learning technology
CN113505120B (en) Double-stage noise cleaning method for large-scale face data set
CN110298391A (en) A kind of iterative increment dialogue intention classification recognition methods based on small sample
CN108229550B (en) Cloud picture classification method based on multi-granularity cascade forest network
CN105469080B (en) A kind of facial expression recognizing method
CN109214444B (en) Game anti-addiction determination system and method based on twin neural network and GMM
CN105930792A (en) Human action classification method based on video local feature dictionary
CN110414587A (en) Depth convolutional neural networks training method and system based on progressive learning
CN114092742A (en) Small sample image classification device and method based on multiple angles
CN115205521A (en) Kitchen waste detection method based on neural network
CN112800232A (en) Big data based case automatic classification and optimization method and training set correction method
CN109740672B (en) Multi-stream feature distance fusion system and fusion method
CN112766134B (en) Expression recognition method for strengthening distinction between classes
CN110827809B (en) Language identification and classification method based on condition generation type confrontation network
CN112801113A (en) Data denoising method based on multi-scale reliable clustering
JP2003256839A (en) Method for selecting characteristics of pattern, method for classifying pattern, method for judging pattern, and its program and its device
CN116524521A (en) English character recognition method and system based on deep learning
CN116403252A (en) Face recognition classification method based on multi-target feature selection of bidirectional dynamic grouping
CN112800959B (en) Difficult sample mining method for data fitting estimation in face recognition
CN115100694A (en) Fingerprint quick retrieval method based on self-supervision neural network
CN113989567A (en) Garbage picture classification method and device
CN113569835A (en) Water meter numerical value reading method based on target detection and segmentation identification

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