CN112418331A - Clustering fusion-based semi-supervised learning pseudo label assignment method - Google Patents
Clustering fusion-based semi-supervised learning pseudo label assignment method Download PDFInfo
- Publication number
- CN112418331A CN112418331A CN202011348264.8A CN202011348264A CN112418331A CN 112418331 A CN112418331 A CN 112418331A CN 202011348264 A CN202011348264 A CN 202011348264A CN 112418331 A CN112418331 A CN 112418331A
- Authority
- CN
- China
- Prior art keywords
- label
- data
- pseudo
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 80
- 230000004927 fusion Effects 0.000 title claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 43
- 238000003064 k means clustering Methods 0.000 claims abstract description 12
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 11
- 238000013528 artificial neural network Methods 0.000 claims abstract description 9
- 238000004364 calculation method Methods 0.000 claims description 14
- 238000000605 extraction Methods 0.000 claims description 8
- 238000004422 calculation algorithm Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 238000012935 Averaging Methods 0.000 claims description 3
- OAICVXFJPJFONN-UHFFFAOYSA-N Phosphorus Chemical compound [P] OAICVXFJPJFONN-UHFFFAOYSA-N 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims 1
- 239000002994 raw material Substances 0.000 claims 1
- 230000009466 transformation Effects 0.000 claims 1
- 238000013135 deep learning Methods 0.000 abstract description 3
- 230000000694 effects Effects 0.000 description 5
- 238000012360 testing method Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 2
- 238000012163 sequencing technique Methods 0.000 description 2
- 102100040160 Rabankyrin-5 Human genes 0.000 description 1
- 101710086049 Rabankyrin-5 Proteins 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000006467 substitution reaction 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
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
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)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (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)
- Probability & Statistics with Applications (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a clustering fusion-based semi-supervised learning pseudo-label assignment method, which comprises the steps of carrying out neural network pre-training by using labeled data and unlabeled data aiming at convolutional neural network semi-supervised learning with unlabeled data sets, and extracting data characteristics by using a trained network; a nearest neighbor method is utilized to endow pseudo labels to N pieces of label-free data which are nearest to the labeled data; analyzing the whole data information by using k-means clustering, and endowing the data which is not endowed with the label with a clustered pseudo label; and continuously training the convolutional neural network by using the obtained label data and the pseudo label data to obtain an optimal network for label assignment. The method can be suitable for semi-supervised learning under deep learning in various fields; the information of the label-free data can be fully mined, and training data with richer contents are provided for the network; the principle is clear, the understanding is easy, and the code is easy to realize.
Description
Technical Field
The invention relates to the technical field of pseudo label assignment of semi-supervised learning, in particular to a semi-supervised learning pseudo label assignment method based on cluster fusion.
Background
With the increasing development of deep learning, fully supervised learning for training a neural network by using labeled data has achieved a good effect. However, in daily life, manually labeling data often consumes a great deal of labor and financial cost, while unlabeled data is often extremely easily available in a large amount, so semi-supervised and unsupervised learning has received the attention of researchers in recent years. The semi-supervised learning is between the supervised learning and the unsupervised learning, not only gives consideration to the accuracy of the supervised learning, but also gives consideration to the practicability of the unsupervised learning, and is a key problem in the field of pattern recognition and machine learning. The method mainly solves the problem of how to train by using labeled data and unlabeled data simultaneously when part of data in training data is unlabeled. The pseudo label method is a classic method of semi-supervised learning and unsupervised learning, and the main principle of the method is to endow a virtual label to unlabeled data, convert the unlabeled data into labeled data and then participate in training. The pseudo label method is classified into a pseudo label method of unsupervised learning and a pseudo label method of supervised learning.
There are two main methods of unsupervised learning with pseudo-label: one is a clustering-based method, which carries out integral clustering (such as k-means) on data characteristics and takes a label obtained after clustering as a pseudo label; one is to obtain the label by calculating the distance of the non-label feature from the reference feature based on the image features or similarity. Among them, the pseudo label method based on clustering is proved to be more effective and maintains the most advanced precision at present.
One of the more applied pseudo label methods for supervised learning is the nearest neighbor method, which is a special case of the k-neighbor method. The k-nearest neighbor method is based on the principle that a pseudo label is obtained by calculating Euclidean distances between a characteristic of labeled data and a characteristic of unlabeled data, and k unlabeled data closest to the labeled data are allocated to a label of the labeled data. The nearest neighbor method is to assign the label of the label data to one non-label data nearest to the label data.
The semi-supervised learning comprises both the label data and the non-label data, so that the characteristics of both the supervised learning and the unsupervised learning are determined, and the semi-supervised learning can be considered from the viewpoint of both the supervised learning and the unsupervised learning. Therefore, the application of the patent combines a k-means clustering method in unsupervised learning and a nearest neighbor method in unsupervised learning, and provides a semi-supervised learning pseudo label assignment method based on clustering fusion.
Disclosure of Invention
This section is for the purpose of summarizing some aspects of embodiments of the invention and to briefly introduce some preferred embodiments. In this section, as well as in the abstract and the title of the invention of this application, simplifications or omissions may be made to avoid obscuring the purpose of the section, the abstract and the title, and such simplifications or omissions are not intended to limit the scope of the invention.
The invention is provided in view of the problems of the existing semi-supervised learning.
Therefore, the technical problem solved by the invention is as follows: when some data in the training data have no label, the training cannot be performed, and a large amount of labor and financial cost is consumed by manually marking the data.
In order to solve the technical problems, the invention provides the following technical scheme: aiming at convolutional neural network semi-supervised learning with a label-free data set, performing neural network pre-training by using labeled data and label-free data, and extracting data characteristics by using the trained network; a nearest neighbor method is utilized to endow pseudo labels to N pieces of label-free data which are nearest to the labeled data; analyzing the whole data information by using k-means clustering, and endowing the rest unlabeled data with clustered pseudo labels; and continuously training the convolutional neural network by using the obtained label data and the pseudo label data to obtain an optimal network for label assignment.
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: the pre-training of the feature extraction convolutional neural network comprises training a resnet101 network by using an imagenet database, extracting the features of all samples in a training data set by using the pre-trained network, and setting the label data features as fl(xj;μj),xjFor jth tag data,. mu.jThe unlabeled data characteristic is f for its corresponding labelu(xi),xiIs the ith unlabeled data.
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: the method for assigning the pseudo labels to the non-label data by using the nearest neighbor method comprises the steps of respectively calculating Euclidean distances between the non-label data characteristics and the label data characteristics, selecting the label data which is closest to the non-label data and corresponds to each piece of non-label data, and only selecting part of most reliable non-label data to assign values when the nearest neighbor method is used for assigning the pseudo labels.
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: selecting the most reliable part of the non-tag data for assignment comprises the steps of sequencing the calculated Euclidean distance from small to large, selecting the first N non-tag data with the smallest distance to assign pseudo tags, wherein the pseudo tags are the tags of the corresponding tag data, and the calculation formula is as follows:
wherein: n is the number of the pseudo labels selected in the round, stThe first N data with the smallest distance are the non-label data with the pseudo label as the corresponding label data xjTag of (a)j。
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: the step of sequencing the calculated Euclidean distances from small to large comprises that a calculation formula of the minimum Euclidean distance between each non-label data characteristic and each label data characteristic is as follows:
wherein: f. ofu(xi) For the ith unlabeled data feature, fl(xj) For the jth label data characteristic, | ·| non-woven phosphor2Is Euclidean distance, L is a label data set, d (x)i) For each unlabeled data feature and the minimum euclidean distance between the respective labeled data features.
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: the pseudo label giving clustering to the data without labels comprises the steps of obtaining clustering pseudo labels for all data characteristics by using a K-means clustering algorithm, clustering the sample characteristics into K clusters by using the K-means clustering algorithm (K is a value set manually), setting the number of all clustering samples to be m, and setting the initial class set of cluster division to be mWherein i is 1-k, and k clustering centers are randomly selected to have the characteristic of mu1,μ2……μk(ii) a For each sample i, calculating the class to which the sample i belongs, and adding the calculation result into a set C of the class(i)In (1), the calculation formula is as follows:
wherein: x is the number ofiIs the data characteristic of the ith participating in the clustering, j is the serial number of the clustering center, c(i)The class number of the data feature i closest to the k classes is one of 1 to k, and for each class j, the class number isTo recalculate the cluster center of the class, the calculation formula is as follows:
wherein: i Cj| is the number of data features contained in each class, μjAveraging each type of feature; and continuously calculating the attribution class and the clustering center of the attribution class until the clustering center is not changed any more, namely reaching local convergence, and converting the pseudo label obtained by clustering to accord with the original label content of the data.
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: the step of converting the pseudo labels obtained through clustering comprises the steps of calculating Euclidean distances between characteristics of the label-free data and characteristics of all label data for the label-free data of which the pseudo labels cannot be obtained, and selecting labels of the label data corresponding to the minimum distance as the pseudo labels of the label-free data, so that the unification of all data labels is realized.
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: after all data are endowed with labels, network training is carried out once, the obtained label data and pseudo label data are input, the loss function used in the network training is classified loss and triple loss, the purpose is to enable the loss function to be continuously close to a minimum value, after the network training is carried out once, the pseudo label assignment of the next round is continued until the loss value obtained by the loss function is reduced and changed extremely little in a certain training process, which is called network convergence, and then the optimal network is obtained.
The invention is a preferable scheme of the clustering fusion-based semi-supervised learning pseudo-label assignment method, wherein: the step of extracting the optimal data features and giving the pseudo labels comprises that as the neural network is trained, the feature extraction performance of the network is better than that of the prior network, so that the number of the pseudo labels selected by a nearest neighbor method can be gradually increased in the next round of pseudo label selection process, more reliable non-label data close to the label data are selected for training, the number selected by the unsupervised k-means method is reduced, and the training is stopped until the network converges.
The invention has the beneficial effects that: the method can be suitable for semi-supervised learning under deep learning in various fields; the information of the label-free data can be fully mined, and training data with richer contents are provided for the network; the principle is clear, the understanding is easy, and the code is easy to realize.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without inventive exercise. Wherein:
fig. 1 is a schematic flowchart of a semi-supervised learning pseudo tag assignment method based on cluster fusion according to a first embodiment of the present invention;
fig. 2 is a schematic diagram illustrating a principle of a semi-supervised learning pseudo tag assignment method based on cluster fusion according to a first embodiment of the present invention.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, specific embodiments accompanied with figures are described in detail below, and it is apparent that the described embodiments are a part of the embodiments of the present invention, not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making creative efforts based on the embodiments of the present invention, shall fall within the protection scope of the present invention.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, but the present invention may be practiced in other ways than those specifically described and will be readily apparent to those of ordinary skill in the art without departing from the spirit of the present invention, and therefore the present invention is not limited to the specific embodiments disclosed below.
Furthermore, reference herein to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one implementation of the invention. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments.
The present invention will be described in detail with reference to the drawings, wherein the cross-sectional views illustrating the structure of the device are not enlarged partially in general scale for convenience of illustration, and the drawings are only exemplary and should not be construed as limiting the scope of the present invention. In addition, the three-dimensional dimensions of length, width and depth should be included in the actual fabrication.
Meanwhile, in the description of the present invention, it should be noted that the terms "upper, lower, inner and outer" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, and are only for convenience of describing the present invention and simplifying the description, but do not indicate or imply that the referred device or element must have a specific orientation, be constructed in a specific orientation and operate, and thus, cannot be construed as limiting the present invention. Furthermore, the terms first, second, or third are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.
The terms "mounted, connected and connected" in the present invention are to be understood broadly, unless otherwise explicitly specified or limited, for example: can be fixedly connected, detachably connected or integrally connected; they may be mechanically, electrically, or directly connected, or indirectly connected through intervening media, or may be interconnected between two elements. The specific meanings of the above terms in the present invention can be understood in specific cases to those skilled in the art.
Example 1
Referring to fig. 1 to 2, a first embodiment of the present invention provides a clustering fusion-based semi-supervised learning pseudo tag assignment method, including:
s1: and constructing a feature extraction convolutional neural network, performing neural network pre-training by using labeled data and unlabeled data, and extracting data features by using the trained network. In which it is to be noted that,
the pre-training of the feature extraction convolutional network comprises training a resnet101 network with an imagenet database, extracting features of all samples in a training data set with the pre-trained network, and setting label data features as fl(xj;μj),xjFor jth tag data,. mu.jThe unlabeled data characteristic is f for its corresponding labelu(xi),xiIs the ith unlabeled data.
S2: and (3) assigning pseudo labels to a batch of non-label data closest to the label data by using a nearest neighbor method. In which it is to be noted that,
the method for assigning the pseudo labels to the non-label data by utilizing the nearest neighbor method comprises the following steps that as the label data and the non-label data exist in semi-supervised learning at the same time to participate in training, the labeled data can provide richer information than the non-label data, so that the nearest neighbor method is considered for assignment of the pseudo labels under semi-supervision; respectively calculating Euclidean distances between the label-free data features and the label data features, and selecting the label data corresponding to each label-free data and closest to the label data, wherein the neural network is only pre-trained, so that only part of the most reliable label-free data is selected for assignment when a nearest neighbor method is used for giving a pseudo label;
further, selecting the most reliable part of the unlabeled data for assignment includes sorting the calculated euclidean distances from small to large, and selecting the first N unlabeled data with the smallest distance to assign a pseudo label, where the pseudo label is a label of the corresponding label data, and the calculation formula is as follows:
wherein: f. ofu(xi) For the ith unlabeled data feature, fl(xj) For the jth label data characteristic, | ·| non-woven phosphor2Is Euclidean distance, L is a label data set, d (x)i) For each label-free data feature and the minimum Euclidean distance between the label data features, N is the number of pseudo labels selected in the current round, stThe first N data with the smallest distance are the non-label data with the pseudo label as the corresponding label data xjTag of (a)j。
S3: and analyzing the whole data information by using k-means clustering, and assigning a pseudo label of the clustering to the data which is not assigned with the label. In which it is to be noted that,
pseudo-labels that assign clusters to unlabeled data include,
for the overall data characteristics, a K-means clustering algorithm is used for obtaining clustering pseudo labels, the K-means clustering algorithm clusters the sample characteristics into K clusters (K is a value set manually), the number of all clustering samples is set to be m, and the initial classification set of cluster division isWherein i is 1-k, and k clustering centers are randomly selected to have the characteristic of mu1,μ2……μk(ii) a For each sample i, calculating the class to which the sample i belongs, and adding the calculation result into a set C of the class(i)In (1), the calculation formula is as follows:
wherein: x is the number ofiIs the data characteristic of the ith participating in the clustering, j is the serial number of the clustering center, c(i)The class number of the data feature i closest to the k classes is one of 1 to k, and for each class j, the cluster center of the class needs to be recalculated, and the calculation formula is as follows:
wherein: i Cj| is the number of data features contained in each class, μjAveraging each type of feature; and continuously calculating the attribution class and the clustering center of the attribution class until the clustering center is not changed any more, namely achieving local convergence, and converting the pseudo label obtained by clustering to accord with the original label content of the data.
Further, converting the pseudo labels obtained by clustering comprises that the pseudo labels obtained by clustering are not consistent with original labels of data, so that further conversion is needed, for each label-free data of which the pseudo labels cannot be obtained, label data with the same clustering label is searched in a label data set according to the clustering label, and the label of the label data is assigned to the label-free data; the clustering labels generate certain noise due to factors such as a k-means algorithm, a network and the like, label data containing the same clustering labels may not be found, for the individual label-free data, Euclidean distances between the characteristics of the individual label-free data and the characteristics of all label data are calculated, and the label of the label data corresponding to the minimum distance is selected as a pseudo label of the label-free data, so that the unification of all data labels is realized.
S4: and continuously training the convolutional neural network by using the obtained label data and the pseudo label data to obtain an optimal network for label assignment. In which it is to be noted that,
when all the label-free data obtain corresponding pseudo labels, training the network by using classification loss and triple loss of all the data, and continuously extracting the characteristics of all the data by using the trained network; because the neural network is trained, the characteristic extraction performance of the network is better than that of the prior network, so that the number of pseudo labels selected by a nearest neighbor method can be gradually increased in the next round of pseudo label selection process, more reliable non-label data close to label data are selected for training, the number selected by the unsupervised k-means method is reduced, and the training is stopped until the network converges.
Example 2
In order to better verify and explain the technical effects adopted in the method, the second embodiment of the invention selects the application of pedestrian re-identification for testing, and compares the test results by means of scientific demonstration to verify the real effect of the method;
the pedestrian re-identification application is used as an experimental object for carrying out experimental test, namely an image of a monitored pedestrian is given, images of pedestrians at different angles under the crossing equipment are retrieved, and identification of the pedestrian is realized, namely, a machine can recognize the same person in different scenes; the method is used for carrying out comparison test with the traditional pseudo label assignment method, wherein the traditional pseudo label assignment method only assigns values to a batch of non-label data closest to the label data by using a nearest neighbor method, and the rest non-label data are not assigned with pseudo labels; on the basis of assigning the nearest label-free data to the label data by using a nearest neighbor method, the method of the invention assigns pseudo labels to the remaining unselected samples by using an integral K-means clustering mode, so that the data can be trained by using classification loss and the like, the information contained in the data is fully utilized, and the network training effect is improved; under the task of pedestrian re-identification marked by a single sample, the results on the data set Market1501 by using the traditional method and the method of the invention are shown in the following table 1, wherein MAP refers to mean average precision, rank-1 refers to the probability that the image with the highest identification probability in the search results is the correct result, and rank-5 and rank-10 refer to the probability that the correct result exists in the images with the highest identification probability of the first 5 and the first 10 respectively.
Table 1: the method compares the resulting data.
As can be seen from Table 1, the results obtained by using the method of the present invention are higher than those obtained by using the conventional method regardless of the average accuracy or the probability of identifying the correct result, and therefore, the effect of network training can be improved by adding the K-means pseudo label labeling method.
It should be noted that the above-mentioned embodiments are only for illustrating the technical solutions of the present invention and not for limiting, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, which should be covered by the claims of the present invention.
Claims (9)
1. A semi-supervised learning pseudo label assignment method based on cluster fusion is characterized in that: comprises the steps of (a) preparing a mixture of a plurality of raw materials,
aiming at convolutional neural network semi-supervised learning with a label-free data set, performing neural network pre-training by using labeled data and label-free data, and extracting data characteristics by using the trained network;
a nearest neighbor method is utilized to endow pseudo labels to N pieces of label-free data which are nearest to the labeled data;
analyzing the whole data information by using k-means clustering, and endowing the residual unlabeled data with clustered pseudo labels;
and continuously training the convolutional neural network by using the obtained label data and the pseudo label data to obtain an optimal network to realize label assignment and feature extraction.
2. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 1, wherein: the pre-training of the feature extraction network comprises,
pre-training a convolutional neural network by using an imagenet database, then extracting the characteristics of all samples in a training data set by using the pre-trained network, and setting the characteristics of label data as fl(xj;μj),xjFor jth tag data,. mu.jThe unlabeled data characteristic is f for its corresponding labelu(xi),xiIs the ith unlabeled data.
3. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 1 or 2, wherein: the using the nearest neighbor method to assign the pseudo label to the non-label data includes,
respectively calculating Euclidean distances between the non-label data characteristics and the label data characteristics, selecting the label data corresponding to each non-label data and having the closest distance, and only selecting part of the most reliable non-label data to assign when the nearest neighbor method is used for assigning the pseudo label.
4. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 3, wherein: the selecting the most reliable part of the non-labeled data to assign a value comprises,
sorting the calculated Euclidean distances from small to large, selecting the first N label-free data with the smallest distance to assign a pseudo label, wherein the pseudo label is a label of corresponding label data, and the calculation formula is as follows:
wherein: n is the number of the pseudo labels selected in the round, stThe first N data with the smallest distance are the non-label data with the pseudo label as the corresponding label data xjTag of (a)j。
5. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 4, wherein: the ordering of the calculated euclidean distances from small to large includes,
the calculation formula of the minimum Euclidean distance between each label-free data characteristic and each label data characteristic is as follows:
wherein: f. ofu(xi) For the ith unlabeled data feature, fl(xj) For the jth label data characteristic, | ·| non-woven phosphor2Is Euclidean distance, L is a label data set, d (x)i) For each unlabeled data feature and the minimum euclidean distance between the respective labeled data features.
6. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 4 or 5, wherein: the pseudo-labels that assign k-means clusters to the remaining unlabeled data include,
using a K-means clustering algorithm to cluster the sample characteristics into K clusters (K is a value set manually), setting the number of all the clustered samples as m, and setting the initial cluster set of cluster division asWherein i is 1-k, and k clustering centers are randomly selected to have the characteristic of mu1,μ2……μk(ii) a For each sample i, calculating the class to which the sample i belongs, and adding the calculation result into a set C of the class(i)In (1), the calculation formula is as follows:
wherein: x is the number ofiIs the data characteristic of the ith participating in the clustering, j is the serial number of the clustering center, c(i)The class number of the data feature i closest to the k classes is one of 1 to k, and for each class j, the cluster center of the class needs to be recalculated, and the calculation formula is as follows:
wherein: i Cj| is the number of data features contained in each class,μjaveraging each type of feature; and continuously calculating the attribution class and the clustering center of the attribution class until the clustering center is not changed any more, namely reaching local convergence, and converting the pseudo label obtained by clustering to accord with the original label content of the data.
7. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 6, wherein: the transformation of the pseudo label obtained by k-means clustering comprises,
and calculating Euclidean distances between the characteristics of the label-free data and the characteristics of all label data for the label-free data of which the pseudo labels cannot be obtained, and selecting the label of the label data corresponding to the minimum distance as the pseudo label of the label-free data, so that the unification of all data labels is realized.
8. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 1 or 7, wherein: the continuous training of the convolutional neural network includes,
after all data are endowed with labels, network training is carried out once, the obtained label data and pseudo label data are input, the loss function used in the network training is classified loss and triple loss, the purpose is to enable the loss function to be continuously close to a minimum value, after the network training is carried out once, the pseudo label assignment of the next round is continued until the loss value obtained by the loss function is reduced and changed extremely little in a certain training process, which is called network convergence, and then the optimal network is obtained.
9. The cluster fusion-based semi-supervised learning pseudo-label assignment method according to claim 8, wherein: the extracting the best data features assigns pseudo-tags including,
because the neural network is trained, the characteristic extraction performance of the network is better than that of the prior network, so that the number of pseudo labels selected by a nearest neighbor method can be gradually increased in the next round of pseudo label selection process, more reliable non-label data close to label data are selected for training, the number selected by the unsupervised k-means method is reduced, and the training is stopped until the network converges.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011348264.8A CN112418331A (en) | 2020-11-26 | 2020-11-26 | Clustering fusion-based semi-supervised learning pseudo label assignment method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011348264.8A CN112418331A (en) | 2020-11-26 | 2020-11-26 | Clustering fusion-based semi-supervised learning pseudo label assignment method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN112418331A true CN112418331A (en) | 2021-02-26 |
Family
ID=74843313
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011348264.8A Pending CN112418331A (en) | 2020-11-26 | 2020-11-26 | Clustering fusion-based semi-supervised learning pseudo label assignment method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112418331A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113609927A (en) * | 2021-07-19 | 2021-11-05 | 上海电力大学 | Pedestrian re-recognition network training method based on branch learning and layered pseudo labels |
CN113836300A (en) * | 2021-09-24 | 2021-12-24 | 中国电信股份有限公司 | Log analysis method, system, device and storage medium |
CN113989596A (en) * | 2021-12-23 | 2022-01-28 | 深圳佑驾创新科技有限公司 | Training method of image classification model and computer readable storage medium |
CN117235448A (en) * | 2023-11-14 | 2023-12-15 | 北京阿丘科技有限公司 | Data cleaning method, terminal equipment and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109697469A (en) * | 2018-12-26 | 2019-04-30 | 西北工业大学 | A kind of self study small sample Classifying Method in Remote Sensing Image based on consistency constraint |
CN110942025A (en) * | 2019-11-26 | 2020-03-31 | 河海大学 | Unsupervised cross-domain pedestrian re-identification method based on clustering |
CN111027421A (en) * | 2019-11-26 | 2020-04-17 | 西安宏规电子科技有限公司 | Graph-based direct-push type semi-supervised pedestrian re-identification method |
CN111192158A (en) * | 2019-12-11 | 2020-05-22 | 国网甘肃省电力公司电力科学研究院 | Transformer substation daily load curve similarity matching method based on deep learning |
CN111563424A (en) * | 2020-04-20 | 2020-08-21 | 清华大学 | Pedestrian re-identification method and device based on semi-supervised learning |
-
2020
- 2020-11-26 CN CN202011348264.8A patent/CN112418331A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109697469A (en) * | 2018-12-26 | 2019-04-30 | 西北工业大学 | A kind of self study small sample Classifying Method in Remote Sensing Image based on consistency constraint |
CN110942025A (en) * | 2019-11-26 | 2020-03-31 | 河海大学 | Unsupervised cross-domain pedestrian re-identification method based on clustering |
CN111027421A (en) * | 2019-11-26 | 2020-04-17 | 西安宏规电子科技有限公司 | Graph-based direct-push type semi-supervised pedestrian re-identification method |
CN111192158A (en) * | 2019-12-11 | 2020-05-22 | 国网甘肃省电力公司电力科学研究院 | Transformer substation daily load curve similarity matching method based on deep learning |
CN111563424A (en) * | 2020-04-20 | 2020-08-21 | 清华大学 | Pedestrian re-identification method and device based on semi-supervised learning |
Non-Patent Citations (2)
Title |
---|
ANKITA SHUKLA ET AL.: "Semi-Supervised Clustering with Neural Networks", 《2020 IEEE SIXTH INTERNATIONAL CONFERENCE ON MULTIMEDIA BIG DATA》, pages 152 - 161 * |
YU WU ET AL.: "Progressive Learning for Person Re-Identification With One Example", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》, pages 2872 - 2881 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113609927A (en) * | 2021-07-19 | 2021-11-05 | 上海电力大学 | Pedestrian re-recognition network training method based on branch learning and layered pseudo labels |
CN113609927B (en) * | 2021-07-19 | 2023-09-29 | 上海电力大学 | Pedestrian re-recognition network training method based on branch learning and layered pseudo tag |
CN113836300A (en) * | 2021-09-24 | 2021-12-24 | 中国电信股份有限公司 | Log analysis method, system, device and storage medium |
CN113989596A (en) * | 2021-12-23 | 2022-01-28 | 深圳佑驾创新科技有限公司 | Training method of image classification model and computer readable storage medium |
CN113989596B (en) * | 2021-12-23 | 2022-03-22 | 深圳佑驾创新科技有限公司 | Training method of image classification model and computer readable storage medium |
CN117235448A (en) * | 2023-11-14 | 2023-12-15 | 北京阿丘科技有限公司 | Data cleaning method, terminal equipment and storage medium |
CN117235448B (en) * | 2023-11-14 | 2024-02-06 | 北京阿丘科技有限公司 | Data cleaning method, terminal equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112418331A (en) | Clustering fusion-based semi-supervised learning pseudo label assignment method | |
CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
CN110298042A (en) | Based on Bilstm-crf and knowledge mapping video display entity recognition method | |
CN110532379B (en) | Electronic information recommendation method based on LSTM (least Square TM) user comment sentiment analysis | |
CN105808752B (en) | A kind of automatic image marking method based on CCA and 2PKNN | |
CN109886072B (en) | Face attribute classification system based on bidirectional Ladder structure | |
CN111241992B (en) | Face recognition model construction method, recognition method, device, equipment and storage medium | |
CN113642547B (en) | Unsupervised domain adaptive character re-identification method and system based on density clustering | |
CN113706547B (en) | Unsupervised domain adaptive semantic segmentation method based on category dissimilarity guidance | |
CN111931562A (en) | Unsupervised feature selection method and system based on soft label regression | |
CN113806494A (en) | Named entity recognition method based on pre-training language model | |
TW201604821A (en) | Collaborative face annotation method and collaborative face annotation system | |
CN114898136B (en) | Small sample image classification method based on characteristic self-adaption | |
CN115187910A (en) | Video classification model training method and device, electronic equipment and storage medium | |
Perina et al. | A hybrid generative/discriminative classification framework based on free-energy terms | |
CN114550091A (en) | Unsupervised pedestrian re-identification method and unsupervised pedestrian re-identification device based on local features | |
CN112668633B (en) | Adaptive graph migration learning method based on fine granularity field | |
CN113743239A (en) | Pedestrian re-identification method and device and electronic equipment | |
CN116775798A (en) | Cross-modal hash method based on feature fusion between graph network and modalities | |
CN116029394A (en) | Self-adaptive text emotion recognition model training method, electronic equipment and storage medium | |
CN113609927B (en) | Pedestrian re-recognition network training method based on branch learning and layered pseudo tag | |
CN114418111A (en) | Label prediction model training and sample screening method, device and storage medium | |
CN114299336A (en) | Photographic image aesthetic style classification method based on self-supervision learning and deep forest | |
CN113378955A (en) | Intrusion detection method based on active learning | |
CN113495964A (en) | Method, device and equipment for screening triples and 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 |