CN109934281B - Unsupervised training method of two-class network - Google Patents
Unsupervised training method of two-class network Download PDFInfo
- Publication number
- CN109934281B CN109934281B CN201910175530.2A CN201910175530A CN109934281B CN 109934281 B CN109934281 B CN 109934281B CN 201910175530 A CN201910175530 A CN 201910175530A CN 109934281 B CN109934281 B CN 109934281B
- Authority
- CN
- China
- Prior art keywords
- training
- network
- label
- images
- labels
- 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
- 238000012549 training Methods 0.000 title claims abstract description 129
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000012545 processing Methods 0.000 claims abstract description 7
- 239000011159 matrix material Substances 0.000 claims description 33
- 239000013598 vector Substances 0.000 claims description 33
- 238000011176 pooling Methods 0.000 claims description 10
- 238000004458 analytical method Methods 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 4
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 abstract description 3
- 238000003909 pattern recognition Methods 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 13
- 230000004913 activation Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 210000002569 neuron Anatomy 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000003213 activating effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides an unsupervised training method of a two-class network, and belongs to the field of image processing, deep learning and pattern recognition. According to the invention, through a series of processing of sample data, the clustering network is trained in different stages, and the trained clustering network is used for classifying images, so that the problems of lack of supervision samples, overlarge data acquisition difficulty and overhigh cost required by data annotation are avoided, and the training of the clustering network and the realization of classification results are simpler and more efficient.
Description
Technical Field
The invention belongs to the field of image processing, deep learning and pattern recognition, and particularly relates to an unsupervised training method for a two-class network.
Background
With the development of deep learning, the classification network is widely applied to various fields such as face recognition, image retrieval, public monitoring, biological recognition, intelligent automobiles, medical assistance, remote sensing and the like. Due to the excellent characteristic characterization capability of the deep neural network, the deep classification network can obtain the classification performance close to human.
Most of the current classification networks are supervised learning networks, and the networks are trained by using images and corresponding labels (namely, in a supervised training mode), so that a large amount of labeled data is required. The data annotation process typically requires a significant amount of labor and time. Therefore, in some application fields with difficult data acquisition and high cost, the lack of supervision samples becomes a main factor for limiting the effect of the deep classification network.
Unlike supervised networks, clustering methods do not require supervised samples and can achieve classification through analysis of a set of samples. Therefore, the research on how to realize the unsupervised training of the classification network has important significance.
Disclosure of Invention
The invention aims to solve the problem that a supervised learning network in the prior art needs a large number of labeled samples, and provides an unsupervised training method for a two-class network.
An unsupervised training method of a two-class network comprises the following steps:
s1, collecting images to construct a data set S, wherein the data set comprises images to be clustered, the number of the images in the data set is M, and any image is I;
s2, constructing a clustering network, wherein the clustering network comprises a convolution layer, a pooling layer and a full-connection layer, and the output of the full-connection layer is used as the output of the clustering network;
s3, training a clustering network, generating corresponding label sets of images participating in training in several stages of training, and training the clustering network based on the images participating in training and the corresponding label sets to obtain final network model training parameters;
and S4, inputting the data set into the trained clustering network to obtain the classified output of the images in the data set.
Further, the step S2 includes:
constructing a clustering network, wherein the clustering network comprises a convolutional layer, a pooling layer and a fully-connected layer, and the output of the fully-connected layer is used as the output of the clustering network;
unifying the sizes of the images in the data set S into (H, W, C), randomly disorganizing and inputting the images into the clustering network, wherein H represents the image height, W represents the image width, and C represents the number of image channels; the input of the clustering network is a four-dimensional vector (M, H, W, C), the number of classified output classes is 2, the output of the clustering network is a two-class output matrix V represented by a two-dimensional tensor (M, 2), and a prediction class vector L represented by a one-dimensional vector (M,1) can be obtained from the matrix V, wherein an element L in LiRepresents a cluster class, represented as
Li＝argmax(V(i,1),V(i,2))
Where V (i,1), V (i,2) respectively represent the element values at the corresponding indices in the matrix V.
Further, the step S3 includes:
s31, setting labels of all images in the data set S to be (0, 0), and forming a first training sample set dataset with the original images0Training the clustering network, wherein the training loss function is a two-norm loss function
Wherein, G is an image corresponding label, and V (i) represents an output vector corresponding to the ith image corresponding to the matrix V;
training network weight by adopting a random gradient descent method through s1After step iteration, a first network weight omega is obtained1；
S32, randomly selecting one image in the data set S as a positive sample, setting the label of the positive sample to be (1,0), setting the corresponding category to be A, setting the labels of the rest images to be (0, 0), copying M parts of the positive sample, and forming a second training sample set dataset together with the data set1；
Set dataset of the second training samples1The data in (1) is randomly disturbed to be used as a training sample, and the first network weight omega is used as the training sample1Training the network for initial weights, s2After step iteration, a second network weight omega is obtained2；
S33, the network pair dataset obtained by training in the step S320The predicted values corresponding to the category A are subjected to size sorting to obtain Q images corresponding to the largest Q predicted values, the corresponding label values of the Q images are set to be (1,0), the labels of the other images are set to be (0, 0), and a third training sample set dataset is obtained2；
Set dataset of the third training samples2The data in (1) is randomly disturbed to be used as a training sample, and the second network weight omega is used2Training the network for initial weights, s3After step iteration, a third network weight omega is obtained3；
S34, the network pair dataset obtained by training in the step S332The predicted values corresponding to the category B are subjected to size sorting to obtain 2 XQ images corresponding to the largest 2 XQ predicted values, and Q images are randomly extractedSet its corresponding sample tag to (0,1) while preserving dataset2The image and the label corresponding to the middle category A are obtained, and a fourth training sample set dataset consisting of Q category A labels, Q category B labels and the corresponding image is obtained3；
Set dataset of the fourth training sample3With the first network weight ω as a training sample1Training the network for initial weights, s4After step iteration, a fourth network weight omega is obtained4；
S35, performing output analysis according to the network obtained by training in the step S34 to generate target labels, exchanging the generated target labels to obtain a label set, training the network by taking the label set as a category label, and processing the training by S5After step iteration, the final network weight omega is obtained5。
Further, the step S35 includes:
s351, performing output analysis according to the network obtained by training in the step S34 to generate a target label;
s352, exchanging the generated target labels to obtain a label set;
s353, taking the label set as a class label, training the network, and obtaining the training result S5After step iteration, the final network weight omega is obtained5。
Further, the step S351 includes:
searching in the matrix V to obtain the maximum value p in the matrix VmaxThe maximum value corresponding position is (m, n);
searching in a dimension V (:, n) where n is located, obtaining a sample corresponding to the largest T labels in the dimension, and setting the labels of the samples as n, wherein: representing all elements under the index, and T is a hyper-parameter;
removing the T data with the set labels from the data set S, and updating the data set without the labeled labels to be Snew(ii) a Removing vectors corresponding to the T data with the set labels from the matrix V to obtain a matrix Vnew(ii) a If n is 0, then SnewOf the middle sampleThe label is set to B; if n is 1, then SnewThe label of the middle sample is set as A;
obtaining a fifth training sample set (S, G) comprising the original image and the corresponding label0) Is denoted as dataset4Wherein G is0Is a one-dimensional tensor (M,1) representing the corresponding generated labelset of the dataset S.
Further, the step S352 includes:
dividing the matrix V into two sub-matrices V according to the corresponding prediction class vector L1And V2Wherein V is1Denotes a matrix composed of all vectors corresponding to (1,0), V2A matrix composed of all vectors corresponding to (0,1) is represented;
extraction of V1And V2Respectively obtaining the predicted values at the indexes corresponding to the respective categories in the vector v1And v2Respectively to vector v1And v2Sorting the medium elements to obtain the minimum M × r predicted values and corresponding indexes i in respective vectors1And i2Wherein r is a hyperparameter;
exchange index i1And i2And obtaining a label set according to the corresponding label.
The invention has the beneficial effects that: the invention provides an unsupervised training method of a two-class network, which trains a clustering network at different stages through a series of processing of sample data, avoids the problems of lack of supervised samples, overlarge data acquisition difficulty and overhigh cost required by data labeling, and ensures that the training of the clustering network and the realization of classification results are simpler and more efficient.
Drawings
Fig. 1 is a flow chart provided by an embodiment of the present invention.
Fig. 2 is a flowchart of step S3 in fig. 1.
Detailed Description
Before describing the present invention, the following definitions of terms are made:
definition 1, convolution layer
The convolutional layer is realized by four-dimensional tensors (N, K, K, C)1) Convolution kernel of representationPerforming convolution operation with the input to extract different features of the input to obtain a four-dimensional tensor (N, H)0，W0，C2) The output of the representation. N, K, C therein1Respectively representing the number of input feature maps, the size of a convolution kernel and the number of feature map channels. Its input can be the input layer in definition 1 or the feature map in definition 3. Wherein H0，W0Height and width, C, representing the output characteristic image of the convolutional layer2Representing the number of convolutions.
Definition 2, pooling layer
The pooling layer is an operation that enables downsampling of the feature map in definition 4. Its input is the four-dimensional tensor (N, H) in definition 3f，Wf，Cf) The output of the characteristic diagram is Hf，WfAnd (5) reducing the characteristic diagram.
Define 3, activate function
The activation function is a nonlinear function after convolution. We call the four-dimensional tensor (N, H) output after activating the functionf，Wf，Cf) Is a characteristic diagram. The activation functions in the network may be chosen to be different functions, as defined by f (x) max (x,0), f (x) tanh (x),
define 4, Softmax layer
The Softmax layer is a Softmax function used at the output layer of the classification network and is defined as
Definition 5, fully connected layer
The fully-connected layer is a network structure in which each neuron in the neural network is connected with each neuron in the previous layer.
Definition 6, learning Rate
The learning rate is a coefficient before the parameter updating amount when the model training is carried out by adopting a back propagation algorithm, and is used for controlling the parameter updating amount each time.
Definition 7, image normalization
Image normalization refers to adjusting the gray value range of each channel in the image to be within a specific range, and is defined as follows:
wherein,
σ denotes standard deviation, NpRepresenting the number of pixels of the image I.
The embodiments of the present invention will be further described with reference to the accompanying drawings.
Referring to fig. 1, the present invention provides an unsupervised training method for a two-class network, and the implementation process is implemented by a tensoflow frame, and specifically implemented by the following steps:
and S1, acquiring images to construct a data set S, wherein the data set acquired in the invention comprises images to be clustered, the number of the images in the data set is M, and any image is I.
In this embodiment, the data set S is a MNIST data set, and the present invention selects 400 images (200 images per class) of 2 classes in the MNIST data set, that is, M is 400. The training settings N, H, W, C are 400, 28, 1, respectively. The MNIST data set is from the national institute of standards and technology, and the training set consists of numbers handwritten from 250 different people, 50% of which are high school students and 50% of which are from the staff of the census bureau of population. The test set is also handwritten digital data of the same scale. In the invention, only the image data in the MNIST data set is selected, and the corresponding class label file is not used.
S2, constructing a clustering network, wherein the clustering network comprises a convolution layer, a pooling layer and a full-connection layer, and the output of the full-connection layer is used as the output of the clustering network.
In the embodiment, a clustering network is constructed, wherein the clustering network comprises 3 convolutional layers, 2 pooling layers and 2 full-connection layers, and the output of the full-connection layers is used as the output of the clustering network; the activation function is a relu function, defined as relu (x) max (x,0), the pooling step in the pooling layer is 2, and the number of neurons in the fully-connected layer is 1024 and 2, respectively. The network removes a softmax layer which is commonly used in a classification network, and directly uses the output of a full connection layer as the output of the clustering network. The invention carries out standardized preprocessing operation on the image and then uses the image as network input.
Unifying the sizes of the images in the data set S into (H, W, C), randomly disorganizing and inputting the images into a clustering network, wherein H represents the image height, W represents the image width, and C represents the number of image channels; the input of the clustering network is a four-dimensional vector (M, H, W, C), the number of classified output categories is 2, the output of the clustering network is a two-category output matrix V represented by a two-dimensional tensor (M, 2), and a prediction category vector L represented by a one-dimensional vector (M,1) can be obtained from the matrix V, wherein an element L in the LiRepresents a cluster class, represented as
Li＝argmax(V(i,1),V(i,2))
Where V (i,1), V (i,2) respectively represent the element values at the corresponding indices in the matrix V.
And S3, training a clustering network, generating corresponding label sets of the images participating in training in several stages of training, and training the clustering network based on the images participating in training and the corresponding label sets to obtain final network model training parameters.
Referring to fig. 2, in the present embodiment, step S3 is implemented by the following sub-steps:
s31, first-stage training
The labels of all images in the data set S are set to (0, 0), and together with the original images, a first training sample set dataset is formed0Training the clustering network, wherein the training loss function is a two-norm loss function
Wherein, G is an image corresponding label, and V (i) represents an output vector corresponding to the ith image corresponding to the matrix V;
training by using random gradient descent methodTraining the network weight, through s1After step iteration, a first network weight omega is obtained1；
In the subsequent steps, random gradient descent method and two-norm loss function are adopted for training.
S32 second-stage training
Randomly selecting an image in a data set S as a positive sample, setting the label of the positive sample to be (1,0), and setting the corresponding category to be A; and setting the labels of the rest images to be (0, 0) and constructing a new training set. The method comprises a sample copying step to solve the problem of uneven samples, and comprises the following specific steps: copying M parts of randomly selected 1 positive samples, and forming a new training set, namely a second training sample set dataset, together with the original data set1；
Set dataset of second training samples1The data in (1) is randomly disturbed to be used as a training sample, and the first network weight omega is used1Training the network for initial weights, s2After step iteration, a second network weight omega is obtained2。
S33, training in the third stage: obtaining more positive samples
The pairs of networks dataset trained in step S320The predicted values corresponding to the category A are subjected to size sorting to obtain Q images corresponding to the largest Q predicted values, the corresponding label values of the Q images are set to be (1,0), the labels of the other images are set to be (0, 0), and a third training sample set dataset is obtained2；
Set dataset of third training samples2The data in (1) is randomly disturbed to be used as a training sample, and the second network weight omega is used2Training the network for initial weights, s3After step iteration, a third network weight omega is obtained3。
S34, fourth-stage training: randomly extracting Q negative samples
The pairs of networks dataset trained in step S332The predicted values corresponding to the category B are subjected to size sorting to obtain 2 × Q images corresponding to the largest 2 × Q predicted values, and the sample labels corresponding to the Q images are set to be (0,1) in random extraction.While conserving dataset2The image and the label corresponding to the middle category A are obtained to obtain a new training set consisting of Q labels of the category A, Q labels of the category B and the corresponding image, namely a fourth training sample set dataset3；
Set dataset of fourth training samples3The data in (1) is used as a training sample, and the first network weight omega is used1Training the network for initial weights, setting a smaller learning rate, s4After step iteration, a fourth network weight omega is obtained4。
S35, fifth stage training
Performing output analysis according to the network obtained by training in the step S34 to generate target labels, exchanging the generated target labels to obtain a label set, training the network by taking the label set as a category label, and processing the training by S5After step iteration, the final network weight omega is obtained5。
In this embodiment, step S35 is implemented by the following sub-steps:
s351, performing output analysis according to the network obtained by training in the step S34 to generate a target label, wherein the process is as follows:
searching in the matrix V to obtain the maximum value p in the matrix VmaxThe maximum value corresponding position is (m, n);
searching in the dimension where n is located, namely V (:, n), obtaining a sample corresponding to the maximum T labels in the dimension, and setting the labels thereof as n, wherein, ": "means all elements under the index are included, T is a hyper-parameter;
removing the T data with the set labels from the data set S, and updating the data set without the labeled labels to be Snew(ii) a Removing the vector corresponding to the T data with the set label from the matrix V to obtain the matrix Vnew. If n is 0 (corresponding to category a), SnewThe label of the middle sample can be set as B; if n is 1 (corresponding to class B), SnewThe label of the middle sample may be set to a.
The process in step S351 described above is a single-class label generation process.
In the second category problem, after the generation process of the single category label is completed, the category label in the remaining data can be set as another type label, so that the above process is only needed to be executed once.
After the above process is completed, a fifth training sample set (S, G) containing the original image and the corresponding label can be obtained0) Is denoted as dataset4. Wherein G is0Is a one-dimensional tensor (M,1) representing the set of generated labels for the data set S.
S352, the label set G obtained in the step S3510And exchanging the labels with the proportion of r to generate a new label set, wherein r is a hyperparameter, the labels are set to be different in size in different training step lengths, and the number of samples corresponding to the exchanged labels is M multiplied by r. In the training iteration process, the ratio r is gradually reduced to 0. The process of exchanging labels is as follows:
for class two problems, the matrix V is divided into two sub-matrices V according to the corresponding prediction class vector L1And V2Wherein V is1Denotes a matrix composed of all vectors corresponding to (1,0), V2A matrix composed of all vectors corresponding to (0,1) is represented;
extraction of V1And V2Respectively obtaining the predicted values at the indexes corresponding to the respective categories in the vector v1And v2Respectively to vector v1And v2Sorting the medium elements to obtain the minimum M × r predicted values and corresponding indexes i in respective vectors1And i2；
Exchange index i1And i2And obtaining a label set according to the corresponding label.
S353, training the network by taking the label set as a class label required by the classification network training, and obtaining the label set5After step iteration, the final network weight omega is obtained5。
In this embodiment, steps S352 and S353 need to be executed for multiple times, where the number of times of execution is equal to the number of different values r in the super-parameter r setting.
In this embodiment, when the clustering network is trained in step 3, the gradientdescnoptimizing optimizer is adopted, and the learning rates at different stages are set to be 0.05, 0.1, 0.01, 0.05,0.01, training step size(s) in different stages1To s5) Are respectively set to 30, 20, 200, 400. Step 3 comprises five stages of training, wherein Q is set to 5 in the third stage of training and the fourth stage of training. In the fifth stage of training, the parameter T is set to be 200, and the parameter r is set to be a piecewise constant of {0.3,0.2,0.1 and 0}, and the training is changed every 100 steps.
And S4, inputting the data set into the trained clustering network to obtain the classified output of the images in the data set.
In this embodiment, the data set S in step S1 is input into the network model constructed in step S2 and trained in step S3, and classification output is obtained.
In this embodiment, when the clustering network is tested, the images in the data set S are classified. In the test, the inputs are set to N ═ 1, H ═ 28, W ═ 28, and C ═ 1. And the input image is preprocessed by image standardization operation and then is accessed into a clustering network, and a classification result is obtained by output of the clustering network.
It will be appreciated by those of ordinary skill in the art that the examples provided herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited examples and embodiments. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.
Claims (4)
1. An unsupervised training method of a two-class network is characterized by comprising the following steps:
s1, collecting images to construct a data set S, wherein the data set comprises images to be clustered, the number of the images in the data set is M, and any image is I;
s2, constructing a clustering network, wherein the clustering network comprises a convolution layer, a pooling layer and a full-connection layer, and the output of the full-connection layer is used as the output of the clustering network; the method comprises the following steps:
constructing a clustering network, wherein the clustering network comprises a convolutional layer, a pooling layer and a fully-connected layer, and the output of the fully-connected layer is used as the output of the clustering network;
unifying the sizes of the images in the data set S into (H, W, C), randomly disorganizing and inputting the images into the clustering network, wherein H represents the image height, W represents the image width, and C represents the number of image channels; the input of the clustering network is a four-dimensional vector (M, H, W, C), the number of classified output classes is 2, the output of the clustering network is a two-class output matrix V represented by a two-dimensional tensor (M, 2), and a prediction class vector L represented by a one-dimensional vector (M,1) can be obtained from the matrix V, wherein an element L in LiRepresents a cluster class, represented as
Li＝argmax(V(i，1)，V(i，2))
Wherein, V (i,1) and V (i,2) respectively represent the element values at the corresponding indexes in the matrix V;
s3, training a clustering network, generating corresponding label sets of images participating in training in several stages of training, and training the clustering network based on the images participating in training and the corresponding label sets to obtain final network model training parameters; the method comprises the following steps:
s31, setting labels of all images in the data set S to be (0, 0), and forming a first training sample set dataset with the original images0Training the clustering network, wherein the training loss function is a two-norm loss function
Wherein, G is an image corresponding label, and V (i) represents an output vector corresponding to the ith image corresponding to the matrix V;
training network weight by adopting a random gradient descent method through s1After step iteration, a first network weight omega is obtained1；
S32, randomly selecting one image in the data set S as a positive sample, setting the label of the positive sample to be (1,0), setting the corresponding category to be A, setting the labels of the rest images to be (0, 0), and copying M copies of the positive sampleForming a second set of training samples dataset together with said data set1；
Set dataset of the second training samples1The data in (1) is randomly disturbed to be used as a training sample, and the first network weight omega is used as the training sample1Training the network for initial weights, s2After step iteration, a second network weight omega is obtained2；
S33, the network pair dataset obtained by training in the step S320The predicted values corresponding to the category A are subjected to size sorting to obtain Q images corresponding to the largest Q predicted values, the corresponding label values of the Q images are set to be (1,0), the labels of the other images are set to be (0, 0), and a third training sample set dataset is obtained2；
Set dataset of the third training samples2The data in (1) is randomly disturbed to be used as a training sample, and the second network weight omega is used2Training the network for initial weights, s3After step iteration, a third network weight omega is obtained3；
S34, the network pair dataset obtained by training in the step S332The predicted values corresponding to the category B are subjected to size sorting to obtain 2 xQ images corresponding to the largest 2 xQ predicted values, the sample labels corresponding to the Q images are set to be (0,1) in the Q images, and simultaneously, the dataset is reserved2The image and the label corresponding to the middle category A are obtained, and a fourth training sample set dataset consisting of Q category A labels, Q category B labels and the corresponding image is obtained3；
Set dataset of the fourth training sample3With the first network weight ω as a training sample1Training the network for initial weights, s4After step iteration, a fourth network weight omega is obtained4；
S35, performing output analysis according to the network obtained by training in the step S34 to generate target labels, exchanging the generated target labels to obtain a label set, training the network by taking the label set as a category label, and processing the training by S5After step iteration, the final network weight omega is obtained5；
And S4, inputting the data set into the trained clustering network to obtain the classified output of the images in the data set.
2. The method for unsupervised training of a classification network of claim 1, wherein the step S35 comprises:
s351, performing output analysis according to the network obtained by training in the step S34 to generate a target label;
s352, exchanging the generated target labels to obtain a label set;
s353, taking the label set as a class label, training the network, and obtaining the training result S5After step iteration, the final network weight omega is obtained5。
3. The method for unsupervised training of a classification network of claim 2, wherein the step S351 comprises:
searching in the matrix V to obtain the maximum value p in the matrix VmaxThe maximum value corresponding position is (m, n);
searching in a dimension V (:, n) where n is located, obtaining samples corresponding to the largest T labels in the dimension, and setting the labels of the samples as n, wherein: representing all elements under the index, and T is a hyper-parameter;
removing the T data with the set labels from the data set S, and updating the data set without the labeled labels to be Snew(ii) a Removing vectors corresponding to the T data with the set labels from the matrix V to obtain a matrix Vnew(ii) a If n is 0, then SnewThe label of the middle sample is set as B; if n is 1, then SnewThe label of the middle sample is set as A;
obtaining a fifth training sample set (S, G) comprising the original image and the corresponding label0) Is denoted as dataset4Wherein G is0Is a one-dimensional tensor (M,1) representing the corresponding generated labelset of the dataset S.
4. The method for unsupervised training of a classification network of claim 3, wherein the step S352 comprises:
dividing the matrix V into two sub-matrices V according to the corresponding prediction class vector L1And V2Wherein V is1Denotes a matrix composed of all vectors corresponding to (1,0), V2A matrix composed of all vectors corresponding to (0,1) is represented;
extraction of V1And V2Respectively obtaining the predicted values at the indexes corresponding to the respective categories in the vector v1And v2Respectively to vector v1And v2Sorting the medium elements to obtain the minimum M × r predicted values and corresponding indexes i in respective vectors1And i2Wherein r is a hyperparameter;
exchange index i1And i2And obtaining a label set according to the corresponding label.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910175530.2A CN109934281B (en) | 2019-03-08 | 2019-03-08 | Unsupervised training method of two-class network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910175530.2A CN109934281B (en) | 2019-03-08 | 2019-03-08 | Unsupervised training method of two-class network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109934281A CN109934281A (en) | 2019-06-25 |
| CN109934281B true CN109934281B (en) | 2021-01-26 |
Family
ID=66986748
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910175530.2A Active CN109934281B (en) | 2019-03-08 | 2019-03-08 | Unsupervised training method of two-class network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109934281B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112232439B (en) * | 2020-11-06 | 2024-04-05 | 四川云从天府人工智能科技有限公司 | Pseudo tag updating method and system in unsupervised ReID |
| CN113298145A (en) * | 2021-05-24 | 2021-08-24 | 中国邮政储蓄银行股份有限公司 | Label filling method and device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101615150A (en) * | 2009-07-24 | 2009-12-30 | 中兴通讯股份有限公司 | Data processing method and device based on embedded database |
| CN108009559A (en) * | 2016-11-02 | 2018-05-08 | 哈尔滨工业大学 | A kind of Hyperspectral data classification method based on empty spectrum united information |
| CN108197666A (en) * | 2018-01-30 | 2018-06-22 | 咪咕文化科技有限公司 | Image classification model processing method and device and storage medium |
| CN108710948A (en) * | 2018-04-25 | 2018-10-26 | 佛山科学技术学院 | A kind of transfer learning method based on cluster equilibrium and weight matrix optimization |
| CN109272040A (en) * | 2018-09-20 | 2019-01-25 | 中国科学院电子学研究所苏州研究院 | A kind of radar operation mode generation method |
Family Cites Families (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104850864A (en) * | 2015-06-01 | 2015-08-19 | 深圳英智源智能系统有限公司 | Unsupervised image recognition method based on convolutional neural network |
| CN105894046B (en) * | 2016-06-16 | 2019-07-02 | 北京市商汤科技开发有限公司 | Method and system, the computer equipment of convolutional neural networks training and image procossing |
| CN107545271B (en) * | 2016-06-29 | 2021-04-09 | 阿里巴巴集团控股有限公司 | Image recognition method, device and system |
| CN106897390B (en) * | 2017-01-24 | 2019-10-15 | 北京大学 | Target precise search method based on depth measure study |
| CN108765465B (en) * | 2018-05-31 | 2020-07-10 | 西安电子科技大学 | Unsupervised SAR image change detection method |
| CN108764366A (en) * | 2018-06-07 | 2018-11-06 | 南京信息职业技术学院 | Feature selection and clustering sampling integration two-classification method for unbalanced data |
| CN109145937A (en) * | 2018-06-25 | 2019-01-04 | 北京达佳互联信息技术有限公司 | A kind of method and device of model training |
| CN109035169B (en) * | 2018-07-19 | 2020-06-12 | 西安交通大学 | Unsupervised/semi-supervised CT image reconstruction depth network training method |
| CN109299270B (en) * | 2018-10-30 | 2021-09-28 | 云南电网有限责任公司信息中心 | Text data unsupervised clustering method based on convolutional neural network |
| CN109426813B (en) * | 2018-11-02 | 2022-06-24 | 中电科新型智慧城市研究院有限公司 | Remote sensing image user-defined interest point extraction method based on fuzzy clustering and neural network model |
| CN109342862A (en) * | 2018-12-14 | 2019-02-15 | 国网山东省电力公司电力科学研究院 | Based on Non-surveillance clustering with and svm classifier Diagnosis Method of Transformer Faults |
-
2019
- 2019-03-08 CN CN201910175530.2A patent/CN109934281B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101615150A (en) * | 2009-07-24 | 2009-12-30 | 中兴通讯股份有限公司 | Data processing method and device based on embedded database |
| CN108009559A (en) * | 2016-11-02 | 2018-05-08 | 哈尔滨工业大学 | A kind of Hyperspectral data classification method based on empty spectrum united information |
| CN108197666A (en) * | 2018-01-30 | 2018-06-22 | 咪咕文化科技有限公司 | Image classification model processing method and device and storage medium |
| CN108710948A (en) * | 2018-04-25 | 2018-10-26 | 佛山科学技术学院 | A kind of transfer learning method based on cluster equilibrium and weight matrix optimization |
| CN109272040A (en) * | 2018-09-20 | 2019-01-25 | 中国科学院电子学研究所苏州研究院 | A kind of radar operation mode generation method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109934281A (en) | 2019-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Mascarenhas et al. | A comparison between VGG16, VGG19 and ResNet50 architecture frameworks for Image Classification | |
| Shao et al. | Feature learning for image classification via multiobjective genetic programming | |
| CN110399821B (en) | Customer satisfaction acquisition method based on facial expression recognition | |
| Zhang et al. | Chromosome classification with convolutional neural network based deep learning | |
| CN104517122A (en) | Image target recognition method based on optimized convolution architecture | |
| CN111090764B (en) | Image classification method and device based on multitask learning and graph convolution neural network | |
| CN103177265B (en) | High-definition image classification method based on kernel function Yu sparse coding | |
| CN110837570B (en) | Method for unbiased classification of image data | |
| CN112818764A (en) | Low-resolution image facial expression recognition method based on feature reconstruction model | |
| CN108090472A (en) | Pedestrian based on multichannel uniformity feature recognition methods and its system again | |
| CN112766283A (en) | Two-phase flow pattern identification method based on multi-scale convolution network | |
| CN109934281B (en) | Unsupervised training method of two-class network | |
| CN112786160A (en) | Multi-image input multi-label gastroscope image classification method based on graph neural network | |
| CN109508640A (en) | Crowd emotion analysis method and device and storage medium | |
| Pratama et al. | Deep convolutional neural network for hand sign language recognition using model E | |
| CN114299304A (en) | Image processing method and related equipment | |
| CN111898614A (en) | Neural network system, image signal and data processing method | |
| CN117611838A (en) | Multi-label image classification method based on self-adaptive hypergraph convolutional network | |
| Goyal et al. | Morphological classification of galaxies using Conv-nets | |
| CN110569889A (en) | Convolutional neural network image classification method based on L2 normalization | |
| CN113723456B (en) | Automatic astronomical image classification method and system based on unsupervised machine learning | |
| CN115457366A (en) | Chinese herbal medicine multi-label recognition model based on graph convolution neural network | |
| CN115063374A (en) | Model training method, face image quality scoring method, electronic device and storage medium | |
| Chu et al. | A genetic programming approach to integrate multilayer cnn features for image classification | |
| CN114548291A (en) | Ecological biological identification method based on MR-CNN algorithm |
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 |