CN111401434A - Image classification method based on unsupervised feature learning - Google Patents

Image classification method based on unsupervised feature learning Download PDF

Info

Publication number
CN111401434A
CN111401434A CN202010173425.8A CN202010173425A CN111401434A CN 111401434 A CN111401434 A CN 111401434A CN 202010173425 A CN202010173425 A CN 202010173425A CN 111401434 A CN111401434 A CN 111401434A
Authority
CN
China
Prior art keywords
image
matrix
original
feature
feature map
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.)
Granted
Application number
CN202010173425.8A
Other languages
Chinese (zh)
Other versions
CN111401434B (en
Inventor
聂飞平
陆继韬
王榕
李学龙
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.)
Northwestern Polytechnical University
Original Assignee
Northwestern Polytechnical 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 Northwestern Polytechnical University filed Critical Northwestern Polytechnical University
Priority to CN202010173425.8A priority Critical patent/CN111401434B/en
Publication of CN111401434A publication Critical patent/CN111401434A/en
Application granted granted Critical
Publication of CN111401434B publication Critical patent/CN111401434B/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
    • 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
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

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

Abstract

The invention provides an image classification method based on unsupervised feature learning. Firstly, constructing a block matrix, utilizing a PCA algorithm to learn a filter bank, reducing the dimension of a feature map, taking output as input, repeating the process to construct a deep network, and obtaining a feature map set after two layers of dimension reduction; then, binarizing the feature map subjected to dimension reduction, calculating a hash value and counting a histogram in blocks to obtain feature embedding of the original image; secondly, training a classifier by using original image labels and feature embedding; and finally, computing features of the unidentified images to be classified, embedding the features and processing the features by using a trained classifier to obtain a final classification result.

Description

Image classification method based on unsupervised feature learning
Technical Field
The invention belongs to the technical field of machine learning and computer vision, and particularly relates to an image classification method based on unsupervised feature learning.
Background
With the rapid increase in computing power and the explosive growth of data, deep learning methods, represented by convolutional neural networks (ConvNet or CNN), have enjoyed tremendous success in a variety of computer vision tasks. It is well known that feature extraction of data has an important influence on the performance of machine learning. One of the main reasons for the success of CNN is that it extracts multi-level semantic information of an image through a cascaded convolution filter, which is superior to the characteristics of the conventional manual design. CNNs are usually supervised, and the free data sets published on the network are not necessarily suitable for the current task, so users usually need to pay high cost to collect data suitable for the task to be completed and hire manual tagging of the data for training of the network. In order to overcome the dependence on data labels and reduce the cost of acquiring data, many unsupervised learning methods have been proposed, i.e. to automatically learn features from a large number of unlabelled samples of pictures and videos. The unsupervised method is used for replacing the current supervised method applied in a large scale, has extremely high economic value and social value, and is a hot field in the current machine learning research.
The image classification problem is one of the most basic, most important and most challenging tasks in computer vision, and is the basis of other 'high-level' tasks, such as target detection, semantic segmentation, pedestrian recognition and the like, and the improvement of the performance of image classification can indirectly improve the performance of other dozens of computer vision tasks. The task of image classification is quite challenging, mainly because of the potentially large intra-class differences in the same class of images.
For example, the document "Jain A K, Farrokhia F. Understand textual definition using Gabor filters [ J ]. Pattern recognition,1991,24(12): 1167. sup. H1186." features for texture classification were designed, "Ahonen T, Hadid A, Pietikon M. facedescription with local patterns: Application to surface recognition [ J ]. IEEE Transactions on Pattern Analysis & Machine integration, 2006(12): 2037. sup. 2041. design features for facial classification," document "L. visual classification [ D. G.8978. distribution & Machine integration," design features for facial classification, "2005. 12. sup. H2004. 2041. special features for human classification," design features for human classification, "2005. D.8978. distribution, and so on,. 1. although these special features for simple design of objects, such as" design of objects, special features for simple classification, special manual classification, see No. 12. No. 1. for manual classification of objects, No. 12. special features for human classification, No. 12. special for design of objects.
Learning features from data of interest is considered a remedy for manually designed features, typically deep neural networks. The core idea of the deep neural network is to learn multiple levels of feature representations, and it is expected that the features at the higher level can represent more abstract semantics in the data, and the abstract representations have stronger robustness to intra-class differences. The document Bruna J, Mallat S.Invaring correlation networks [ J ]. IEEE transactions on pattern analysis and machine interpretation, 2013,35(8): 1872-. Despite the use of a fixed filter bank, it still exhibits superior performance in the tasks of handwritten digit recognition, texture recognition. However, the structure has insufficient generalization performance in tasks with obvious changes in illumination conditions, such as face recognition. The document "Chan T H, Jia K, Gao S, et al PCANet: A simple depleting baseline for image classification? [J] IEEE transactions on imaging 2015,24(12): 5017-. The network does not use labels in the stage of feature learning, and achieves excellent effects in tasks such as handwritten number recognition and face recognition. However, the dimensionality of the feature extracted by the algorithm increases with the network depth index, and it is difficult to extract semantic information with richer levels by improving the depth.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides an image classification method based on unsupervised feature learning. Firstly, constructing a block matrix, utilizing a PCA algorithm to learn a filter bank, reducing the dimension of a feature map, taking output as input, repeating the process to construct a deep network, and obtaining a feature map set after two layers of dimension reduction; then, binarizing the feature map subjected to dimension reduction, calculating a hash value and counting a histogram in blocks to obtain feature embedding of the original image; secondly, training a classifier by using original image labels and feature embedding; and finally, computing features of the unidentified images to be classified, embedding the features and processing the features by using a trained classifier to obtain a final classification result. The method can solve the problem of exponential increase of the feature dimension of the existing PCANet algorithm, and can effectively extract the features of the image for classification tasks through a simple linear PCA operator.
An image classification method based on unsupervised feature learning is characterized by comprising the following steps:
step 1: for the original ith input image, i is 1,2, …, N is the number of input images, and the feature map set after dimensionality reduction is extracted according to the following process
Figure BDA0002409578710000031
Wherein,
Figure BDA0002409578710000032
a j-th dimension-reduced feature map representing the original i-th input image, j being 1,2, …, D1，D1For the set target dimension:
step 1.1, extracting image blocks which take each pixel as a center in an image by using a sliding window with the step length of 1, then stretching each image block into one-dimensional vectors, respectively carrying out de-equalization processing on each vector, and taking each de-equalized vector as a column of a matrix to obtain a matrix X, wherein the size of an input image is p × p, the size of the image block is k × k, and the value range of k is more than 1 and less than k and less than p;
step 1.2: an orthogonal matrix V formed by a group of standard orthogonal bases is solved by adopting a principal component analysis algorithm, and the following requirements are met:
Figure BDA0002409578710000033
wherein, L1The value range of the set number of the filters is not less than 1 and not more than L1＜k2；
Transforming each column of the matrix V into a matrix with the size of k × k from a vector, taking each matrix obtained by transformation as a filter, performing convolution operation on each filter and the original input image respectively, and enabling the size of a feature map obtained after convolution to be the same as that of the original image through zero filling operation to obtain L1A feature map with a web size of p × p;
step 1.3: stretching each characteristic diagram output in the step 1.2 into a one-dimensional vector, and taking the one-dimensional vector as a row of the matrix to obtain an original characteristic diagram matrix
Figure BDA0002409578710000034
Then, a projection matrix U formed by a group of standard orthogonal bases is solved by adopting a principal component analysis algorithm1And satisfies the following conditions:
Figure BDA0002409578710000035
wherein D is1The value range is not less than 1 and not more than D for the set target dimension1≤L1；
According to
Figure BDA0002409578710000036
Calculating to obtain a matrix P after dimension reduction, converting each row vector of P into a matrix with the size of P × P, wherein each matrix is a feature map after dimension reduction to obtain D1The feature map after dimension reduction is formed into a feature map set after dimension reduction of the original input image;
step 2: for each feature map set obtained in step 1 after dimension reduction
Figure BDA0002409578710000037
With each image therein
Figure BDA0002409578710000038
For the input image, j is 1,2, …, D1And let the procedure described in step 1.2The number of the filters is L2The value range is not less than 1 and not more than L2＜k2And the target dimension in step 1.3 is D2The value range is not less than 1D2≤L2Calculating according to the step 1 to obtain a feature map set after the dimension reduction for the second time
Figure BDA0002409578710000039
And step 3: for each feature map set obtained in step 2
Figure BDA00024095787100000310
All images are subjected to binarization processing to enable all pixel values in the images to be 0 or 1, and then all pixel point values at the same position in different images are extracted to form a length D2The binary strings are extracted from the pixel points at all positions, each binary string is converted into a decimal number, and each decimal number is used as a new pixel value of the pixel position extracted from the corresponding binary string to obtain a new image which is marked as
Figure BDA0002409578710000041
Thus, for the original ith input image, a new image set is obtained
Figure BDA0002409578710000042
Figure BDA0002409578710000043
And 4, step 4: obtaining a new image set of the original ith input image for step 3iI is 1,2, …, N, each image is subject to block extraction by a sliding window with step length s and size b × b, wherein, 1 is not less than b not less than p, 1 is not less than s not less than b, then each block is statistically provided with
Figure BDA0002409578710000044
Connecting all the histograms of the bins into a vector; general collectionConnecting all the images in the image collection according to the vectors obtained by the process, and embedding the connected vectors as the features of the original ith input image;
for each original input image a new image setiProcessing the images i to 1,2, …, N according to the above process to obtain corresponding feature embeddings, and obtaining N feature embeddings of all the original N input images;
and 5: embedding the N characteristics obtained in the step (4) into input data serving as a classifier, taking labels of the original N input images as input labels of the classifier, and training the classifier to obtain a trained classifier; the classifier comprises a nearest neighbor classifier and a support vector machine;
step 6: and (3) taking M unlabeled images to be classified as input images, calculating according to the steps 1-4 to obtain M feature embeddings, and inputting the M feature embeddings into the trained classifier obtained in the step 5 to obtain a classification result.
The invention has the beneficial effects that: because gradient back propagation is not used, PCA is used for unsupervised learning of the multi-stage filter bank, compared with a convolutional neural network, the calculation amount is greatly reduced, and the method is suitable for rapid calculation under a non-special hardware environment; because the feature map after dimensionality reduction is used as the next-layer input image, compared with the PCANet, the problem that the parameter number increases exponentially along with the increase of the layer number is solved, and the construction of a deeper network to extract semantic information of a higher layer is possible; because the feature extraction process does not depend on the labeled data, a large amount of unlabeled data on the network can be added into the training set under the condition of not increasing the additional manual labeling cost, and the feature extraction performance is further improved; since each step can be realized in a parallelization mode to shorten the running time, the method can be used for large-scale data set training on a distributed cluster.
Drawings
FIG. 1 is a basic flow chart of an image classification method based on unsupervised feature learning according to the present invention;
FIG. 2 is a schematic diagram of the process of computing a filter bank using a principal component analysis algorithm according to the present invention.
Detailed Description
The present invention will be further described with reference to the following drawings and examples, which include, but are not limited to, the following examples.
As shown in fig. 1, the present invention provides an image classification method based on unsupervised feature learning, which is implemented as follows:
1. and extracting a feature map set after the dimension reduction of the first layer.
(1) A matrix of tiles is constructed.
Hypothesis training set
Figure BDA0002409578710000051
Consists of N images of p × p size, for each image, a tile of size k × k centered around each pixel in the image is extracted in the form of a sliding window, with a sliding step of 1.
Figure BDA0002409578710000052
Wherein x isi,jA column vector representing the jth tile of the ith image,
Figure BDA0002409578710000053
a column vector representing the jth tile of the ith image after the mean removal, 1 being a vector having all dimensions 1, i being 1,2, …, N, j being 1,2, …, m, m being (p-k +1)2The number of patches extracted for each image.
For the original ith image, i is 1,2, …, N, and each of the averaged patch vectors is used as a column of the matrix to obtain a matrix
Figure BDA0002409578710000054
(2) The filter bank is solved and the eigen map is calculated.
Solving a group of standard orthogonal bases by using a principal component analysis algorithm to minimize the reconstruction error, namely:
Figure BDA0002409578710000055
wherein, L1The value range of the set number of the filters is not less than 1 and not more than L1＜k2. The solution to the optimization problem is the covariance matrix XiXi TMaximum L1The eigenvectors capture the main differences between the training sample blocks, and are transformed L1And k × k, namely obtaining a filter bank of the first layer.
L using the first layer1And (3) convolving each filter with the original input image to obtain an original output characteristic diagram of a first layer:
Figure BDA0002409578710000057
wherein, represents a two-dimensional convolution operation, IiRepresenting the original i-th input image, W1 lThe ith filter representing the first layer,
Figure BDA0002409578710000058
representing the feature map generated by convolving the ith original input image with the ith filter. Before convolution in the original image IiAnd (5) zero filling is carried out on the periphery, so that the size of the feature image output after convolution is kept the same as that of the original image before convolution.
(3) And reducing the dimension of the feature map.
If the above steps (1) - (2) are directly repeated to construct a deep network, that is, all the feature maps obtained in step (2) are input into step (1) and iterated several times, the number of output feature maps will increase exponentially. In order to avoid the exponential increase of the number of feature maps with the depth of the network, the invention adds a dimension reduction module before inputting the next layer. Stretching all the characteristic diagrams output in the step (2) into one-dimensional vectors to obtain an original characteristic diagram matrix
Figure BDA0002409578710000061
Then adopts the principal componentProjection matrix U formed by solving a group of standard orthogonal bases through a sub-analysis algorithm1And satisfies the following conditions:
Figure BDA0002409578710000062
wherein D is1The value range is not less than 1 and not more than D for the set target dimension1≤L1. Solution U of the optimization problem1Is a covariance matrix
Figure BDA00024095787100000613
Maximum D1And the feature vectors corresponding to the feature values form an orthogonal matrix.
According to
Figure BDA0002409578710000064
Calculating to obtain a matrix P after dimensionality reduction, wherein the size of the matrix P is
Figure BDA0002409578710000065
Reducing each row vector of P into a matrix with the size of P × P to obtain a characteristic diagram after dimension reduction, and obtaining D by the matrix P in a co-correspondence way1Forming a feature map set after dimensionality reduction by using a feature map with the amplitude of p × p after dimensionality reduction;
to this end, for the original I-th input image IiObtaining a feature map set after dimension reduction
Figure BDA0002409578710000066
Figure BDA0002409578710000067
2. And (5) iteratively constructing a deep network.
Each feature map set after dimension reduction obtained in the step 1 is collected
Figure BDA00024095787100000614
As input to the next layer, i.e. in each set
Figure BDA00024095787100000615
Each image of
Figure BDA0002409578710000068
For the input image, j is 1,2, …, D1The calculation is carried out according to the above process starting from step 1, and the number of the filters in step (2) is L2The value range is not less than 1 and not more than L2＜k2The target dimension in the step (3) is D2The value range is not less than 1D2≤L2. To this end, for each original image IiI 1,2, …, N, to yield D1A set of images
Figure BDA0002409578710000069
Wherein each image set
Figure BDA00024095787100000610
And further comprises D2The images being, i.e.
Figure BDA00024095787100000611
3. And (4) carrying out hash calculation.
For step 2, obtaining each feature map set
Figure BDA00024095787100000612
Each set is encoded as a new image, processed as follows:
(1) and (6) binarization.
And (4) carrying out binarization processing on all images in the set, wherein if the negative number is 0 and the non-negative number is 1, the pixel values in the images are all changed into 0 or 1.
(2) A binary hash.
Extracting D in the set2All D of the same position of the image2A binary value constituting a length D2The binary string of (2) is extracted from the pixel points at all positions in this way. Each binary string is treated as a binary representation constituting a decimal number, converted into a decimal number, and the output D2The amplitude and real number characteristic diagram is converted intoAn integer image, namely:
Figure BDA0002409578710000071
wherein,
Figure BDA0002409578710000078
representation collection
Figure BDA0002409578710000079
H (-) represents the binarization processing of the image in the step (1) above,
Figure BDA00024095787100000710
representation collection
Figure BDA0002409578710000075
The d-th image of (1). Each pixel in the new image obtained is a closed interval
Figure BDA00024095787100000711
Is an integer of (1).
Thus, for the original ith input image, a new image set is obtained
Figure BDA0002409578710000076
Figure BDA00024095787100000712
4. And (5) partitioning a statistical histogram.
Obtaining a new image set of the original ith input image for step 3iDividing each image into blocks by adopting a sliding window of B × B, wherein B is more than or equal to 1 and less than or equal to p, the value range of the sliding step length s is more than or equal to 1 and less than or equal to B, setting the number of the obtained blocks as B, counting the number of the blocks containing the blocks
Figure BDA00024095787100000713
Histogram of each bin, and histogram of all B binsAre concatenated into a vector. Connecting the vectors obtained by the process in all the images in the set, so far, for each setiObtain a vector
Figure BDA0002409578710000077
Embedded as a feature of the original ith input image.
For each original input image a new image setiIf all the i-1, 2, …, N are processed according to the above procedure to obtain the corresponding feature embeddings, N feature embeddings of all the original N input images can be obtained.
5. The classifier is trained using feature embedding.
Embedding the N characteristics obtained in the step (4) into input data serving as a classifier, taking labels of the original N input images as input labels of the classifier, and training the classifier to obtain a trained classifier; the classifier comprises a nearest neighbor classifier and a support vector machine;
6. and (5) classifying the images.
Using M unlabelled images to be classified as input images, and keeping L filter number set when embedding training image features1、L2And the resulting projection matrix U1、U2And (4) calculating to obtain M feature embeddings according to the steps 1-4, and inputting the M feature embeddings into the trained classifier obtained in the step 5 to obtain a classification result.
In this embodiment, a simulation experiment is performed on an Ubuntu 16.04 operating system with a central processing unit of intel (r) xeon (r) CPU E5-2680 [email protected] and a memory 512G by using Python and pytorreh software, and the experiment is performed by using data sets MNIST and CIFAR-10, respectively, and the information of the data sets is shown in table 1. The accuracy of the feature extraction performed by the steps 1-4 of the present invention and the classification performed by using the linear SVM as the classifier is shown in Table 2, and the time taken for the feature extraction is shown in Table 3. Therefore, the method can extract the features of large-scale data in a short time, and well complete classification tasks by using the extracted features.
TABLE 1
Data set Number of samples Number of categories
MNIST 60000 10
CIFAR-10 50000 10
TABLE 2
Data set Accuracy of
MNIST 99.39％
CIFAR-10 77.09％
TABLE 3
Data set Time (seconds)
MNIST 587.88
CIFAR-10 2564.04

Claims (1)

1. An image classification method based on unsupervised feature learning is characterized by comprising the following steps:
step 1: for the original ith input image, i is 1,2, …, N is the number of input images, and the feature map set after dimensionality reduction is extracted according to the following process
Figure FDA0002409578700000011
Wherein,
Figure FDA0002409578700000012
a j-th dimension-reduced feature map representing the original i-th input image, j being 1,2, …, D1，D1For the set target dimension:
step 1.1, extracting image blocks which take each pixel as a center in an image by using a sliding window with the step length of 1, then stretching each image block into one-dimensional vectors, respectively carrying out de-equalization processing on each vector, and taking each de-equalized vector as a column of a matrix to obtain a matrix X, wherein the size of an input image is p × p, the size of the image block is k × k, and the value range of k is more than 1 and less than k and less than p;
step 1.2: an orthogonal matrix V formed by a group of standard orthogonal bases is solved by adopting a principal component analysis algorithm, and the following requirements are met:
Figure FDA0002409578700000013
wherein, L1The value range of the set number of the filters is not less than 1 and not more than L1＜k2；
Transforming each column of the matrix V into a matrix with the size of k × k from a vector, taking each matrix obtained by transformation as a filter, performing convolution operation on each filter and the original input image respectively, and enabling the size of a feature map obtained after convolution to be the same as that of the original image through zero filling operation to obtain L1A feature map with a web size of p × p;
step 1.3: stretching each characteristic diagram output in the step 1.2 into a one-dimensional vector, and taking the one-dimensional vector as a row of the matrix to obtain an original characteristic diagram matrix
Figure FDA0002409578700000014
Then, a projection matrix U formed by a group of standard orthogonal bases is solved by adopting a principal component analysis algorithm1And satisfies the following conditions:
Figure FDA0002409578700000015
wherein D is1The value range is not less than 1 and not more than D for the set target dimension1≤L1；
According to
Figure FDA0002409578700000016
Calculating to obtain a matrix P after dimension reduction, converting each row vector of P into a matrix with the size of P × P, wherein each matrix is a feature map after dimension reduction to obtain D1The feature map after dimension reduction is formed into a feature map set after dimension reduction of the original input image;
step 2: for each feature map set obtained in step 1 after dimension reduction
Figure FDA0002409578700000017
With each image therein
Figure FDA0002409578700000018
For the input image, j is 1,2, …, D1And the number of the filters in the step 1.2 is L2The value range is not less than 1 and not more than L2＜k2And the target dimension in step 1.3 is D2The value range is not less than 1D2≤L2Calculating according to the step 1 to obtain a feature map set after the dimension reduction for the second time
Figure FDA0002409578700000019
And step 3: for each feature map set obtained in step 2
Figure FDA00024095787000000110
All images are subjected to binarization processing to enable all pixel values in the images to be 0 or 1, and then all pixel point values at the same position in different images are extracted to form a length D2The binary strings are extracted from the pixel points at all positions, each binary string is converted into a decimal number, and each decimal number is used as a new pixel value of the pixel position extracted from the corresponding binary string to obtain a new image which is marked as
Figure FDA0002409578700000021
Thus, for the original ith input image, a new image set is obtained
Figure FDA0002409578700000022
Figure FDA0002409578700000023
And 4, step 4: obtaining a new image set of the original ith input image for step 3iI is 1,2, …, N, each image is subject to block extraction by a sliding window with step length s and size b × b, wherein, 1 is not less than b not less than p, 1 is not less than s not less than b, then each block is statistically provided with
Figure FDA0002409578700000024
Connecting all the histograms of the bins into a vector; connecting vectors obtained by all images in the set according to the process, and embedding the connected vectors as the features of the original ith input image;
for each original input image a new image setiProcessing the images i to 1,2, …, N according to the above process to obtain corresponding feature embeddings, and obtaining N feature embeddings of all the original N input images;
and 5: embedding the N characteristics obtained in the step (4) into input data serving as a classifier, taking labels of the original N input images as input labels of the classifier, and training the classifier to obtain a trained classifier; the classifier comprises a nearest neighbor classifier and a support vector machine;
step 6: and (3) taking M unlabeled images to be classified as input images, calculating according to the steps 1-4 to obtain M feature embeddings, and inputting the M feature embeddings into the trained classifier obtained in the step 5 to obtain a classification result.
CN202010173425.8A 2020-03-12 2020-03-12 Image classification method based on unsupervised feature learning Active CN111401434B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010173425.8A CN111401434B (en) 2020-03-12 2020-03-12 Image classification method based on unsupervised feature learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010173425.8A CN111401434B (en) 2020-03-12 2020-03-12 Image classification method based on unsupervised feature learning

Publications (2)

Publication Number Publication Date
CN111401434A true CN111401434A (en) 2020-07-10
CN111401434B CN111401434B (en) 2024-03-08

Family

ID=71432455

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010173425.8A Active CN111401434B (en) 2020-03-12 2020-03-12 Image classification method based on unsupervised feature learning

Country Status (1)

Country Link
CN (1) CN111401434B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111898710A (en) * 2020-07-15 2020-11-06 中国人民解放***箭军工程大学 Method and system for selecting characteristics of graph
CN112200763A (en) * 2020-08-24 2021-01-08 江苏科技大学 Liver function grading method based on liver CT image
CN112766367A (en) * 2021-01-18 2021-05-07 北京工商大学 Image classification method and device based on PCANet

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090297046A1 (en) * 2008-05-29 2009-12-03 Microsoft Corporation Linear Laplacian Discrimination for Feature Extraction
US20140079297A1 (en) * 2012-09-17 2014-03-20 Saied Tadayon Application of Z-Webs and Z-factors to Analytics, Search Engine, Learning, Recognition, Natural Language, and Other Utilities
CN104881682A (en) * 2015-05-26 2015-09-02 东南大学 Image classification method based on locality preserving mapping and principal component analysis
CN105426919A (en) * 2015-11-23 2016-03-23 河海大学 Significant guidance and unsupervised feature learning based image classification method
CN106845525A (en) * 2016-12-28 2017-06-13 上海电机学院 A kind of depth confidence network image bracket protocol based on bottom fusion feature
CN108460391A (en) * 2018-03-09 2018-08-28 西安电子科技大学 Based on the unsupervised feature extracting method of high spectrum image for generating confrontation network
CN108921029A (en) * 2018-06-04 2018-11-30 浙江大学 A kind of SAR automatic target recognition method merging residual error convolutional neural networks and PCA dimensionality reduction
CN110473140A (en) * 2019-07-18 2019-11-19 清华大学 A kind of image dimension reduction method of the extreme learning machine based on figure insertion

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090297046A1 (en) * 2008-05-29 2009-12-03 Microsoft Corporation Linear Laplacian Discrimination for Feature Extraction
US20140079297A1 (en) * 2012-09-17 2014-03-20 Saied Tadayon Application of Z-Webs and Z-factors to Analytics, Search Engine, Learning, Recognition, Natural Language, and Other Utilities
CN104881682A (en) * 2015-05-26 2015-09-02 东南大学 Image classification method based on locality preserving mapping and principal component analysis
CN105426919A (en) * 2015-11-23 2016-03-23 河海大学 Significant guidance and unsupervised feature learning based image classification method
CN106845525A (en) * 2016-12-28 2017-06-13 上海电机学院 A kind of depth confidence network image bracket protocol based on bottom fusion feature
CN108460391A (en) * 2018-03-09 2018-08-28 西安电子科技大学 Based on the unsupervised feature extracting method of high spectrum image for generating confrontation network
CN108921029A (en) * 2018-06-04 2018-11-30 浙江大学 A kind of SAR automatic target recognition method merging residual error convolutional neural networks and PCA dimensionality reduction
CN110473140A (en) * 2019-07-18 2019-11-19 清华大学 A kind of image dimension reduction method of the extreme learning machine based on figure insertion

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
冯姝: "无监督局部特征学习的鲁棒性人脸识别", 计算机应用, 10 February 2017 (2017-02-10) *
王改华等: "采用无监督学习算法与卷积的图像分类模型", 华侨大学学报(自然科学版), 20 January 2018 (2018-01-20) *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111898710A (en) * 2020-07-15 2020-11-06 中国人民解放***箭军工程大学 Method and system for selecting characteristics of graph
CN111898710B (en) * 2020-07-15 2023-09-29 中国人民解放***箭军工程大学 Feature selection method and system of graph
CN112200763A (en) * 2020-08-24 2021-01-08 江苏科技大学 Liver function grading method based on liver CT image
CN112766367A (en) * 2021-01-18 2021-05-07 北京工商大学 Image classification method and device based on PCANet

Also Published As

Publication number Publication date
CN111401434B (en) 2024-03-08

Similar Documents

Publication Publication Date Title
CN111898736B (en) Efficient pedestrian re-identification method based on attribute perception
Kadam et al. Detection and localization of multiple image splicing using MobileNet V1
CN106203356B (en) A kind of face identification method based on convolutional network feature extraction
CN111401434B (en) Image classification method based on unsupervised feature learning
CN113674334B (en) Texture recognition method based on depth self-attention network and local feature coding
CN105046272B (en) A kind of image classification method based on succinct non-supervisory formula convolutional network
Sheng et al. Learning-based road crack detection using gradient boost decision tree
CN108932518B (en) Shoe print image feature extraction and retrieval method based on visual bag-of-words model
CN110211127B (en) Image partition method based on bicoherence network
CN114492619B (en) Point cloud data set construction method and device based on statistics and concave-convex performance
CN111079514A (en) Face recognition method based on CLBP and convolutional neural network
CN113076927A (en) Finger vein identification method and system based on multi-source domain migration
Wang et al. A new Gabor based approach for wood recognition
CN115527072A (en) Chip surface defect detection method based on sparse space perception and meta-learning
Huynh et al. Plant identification using new architecture convolutional neural networks combine with replacing the red of color channel image by vein morphology leaf
CN113192076B (en) MRI brain tumor image segmentation method combining classification prediction and multi-scale feature extraction
Zhang et al. Multi-path siamese convolution network for offline handwritten signature verification
CN110991554A (en) Improved PCA (principal component analysis) -based deep network image classification method
White et al. Digital fingerprinting of microstructures
CN112070116B (en) Automatic artistic drawing classification system and method based on support vector machine
CN111488797B (en) Pedestrian re-identification method
CN105844299B (en) A kind of image classification method based on bag of words
Özyurt et al. A new method for classification of images using convolutional neural network based on Dwt-Svd perceptual hash function
CN109840914B (en) Texture segmentation method based on user interaction
CN117036904A (en) Attention-guided semi-supervised corn hyperspectral image data expansion method

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