CN108416389B - Image classification method based on noise reduction sparse automatic encoder and density space sampling - Google Patents

Image classification method based on noise reduction sparse automatic encoder and density space sampling Download PDF

Info

Publication number
CN108416389B
CN108416389B CN201810212714.7A CN201810212714A CN108416389B CN 108416389 B CN108416389 B CN 108416389B CN 201810212714 A CN201810212714 A CN 201810212714A CN 108416389 B CN108416389 B CN 108416389B
Authority
CN
China
Prior art keywords
image
training
noise reduction
automatic encoder
fisher vector
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810212714.7A
Other languages
Chinese (zh)
Other versions
CN108416389A (en
Inventor
张辉
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.)
Yancheng Teachers University
Original Assignee
Yancheng Teachers 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 Yancheng Teachers University filed Critical Yancheng Teachers University
Priority to CN201810212714.7A priority Critical patent/CN108416389B/en
Publication of CN108416389A publication Critical patent/CN108416389A/en
Application granted granted Critical
Publication of CN108416389B publication Critical patent/CN108416389B/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
    • G06F18/214—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
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns

Landscapes

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

Abstract

The invention discloses an image classification method based on a noise reduction sparse automatic encoder and density space sampling. The method comprises the following steps: constructing an image block training set; constructing a noise reduction sparse automatic encoder of a single hidden layer, inputting an image block training set, and training the noise reduction sparse automatic encoder; performing density space sampling on each image in the training image dataset and the test image dataset; extracting local characteristic set information from a space region obtained by density space sampling of each image by using a noise reduction sparse automatic encoder; encoding the feature set information by using two layers of stacked Fisher vectors to obtain a final Fisher Vector of each image; and training a classifier by using the Fisher vector to realize image classification. The invention can accurately acquire the image information, improves the classification accuracy of the image, and can be used for constructing a large-scale image classification and retrieval system.

Description

Image classification method based on noise reduction sparse automatic encoder and density space sampling
Technical Field
The invention belongs to the technical field of image classification, and particularly relates to an image classification method based on a noise reduction sparse automatic encoder and density space sampling.
Background
With the development of multimedia technology, image classification has become a hot problem for research in the field of computer vision. The image classification is to classify the images into different preset categories according to certain attributes of the images. How to effectively express the images is the key for improving the accuracy of image classification, wherein the problems of feature selection and feature extraction are key and difficult problems existing in the current image classification. Although the traditional methods for artificially designing features, such as Gabor filter, SIFT, LBP, HOG, and the like, have a certain effect in image classification, the methods need to be carefully designed and cannot be well applied to specific problems. In recent years, deep learning has been highly successful in a variety of application fields such as image recognition and speech recognition. A Convolutional Neural Network (CNN) in a deep learning model makes breakthrough progress in feature extraction, but the CNN needs a large amount of manually labeled label data, and a noise reduction sparse automatic encoder can learn the features describing image contents by using a large amount of label-free data. The traditional spatial pyramid is widely applied to image classification, but the grid proportion of the traditional spatial pyramid classification is fixed, the classification method is single, and the density spatial sampling method based on the sliding window can capture more image spatial information.
Disclosure of Invention
The invention provides an image classification method based on a noise reduction sparse automatic encoder and density space sampling, aims to solve the problems of image feature extraction and encoding, overcomes the defects of the existing image classification method, reduces the calculation cost and improves the classification precision.
In order to achieve the technical purpose, the technical scheme of the invention is as follows:
an image classification method based on a noise reduction sparse automatic encoder and density space sampling comprises the following steps:
(1) constructing an image block training set;
(2) constructing a noise reduction sparse automatic encoder of a single hidden layer, inputting the image block training set obtained in the step (1), and training the noise reduction sparse automatic encoder until an iteration stopping condition is met;
(3) for training image data set M1And a test image dataset M2Performing density space sampling on each image in the image data set, and setting a training image data set M1Number of images of m1Testing the image dataset M2Number of images of m2；
(4) Taking a trained noise reduction sparse automatic encoder as a feature extractor, and extracting local feature set information from a space region obtained by density space sampling of each image;
(5) encoding the feature set information extracted in the step (4) by adopting two layers of stacked Fisher vectors to obtain a final Fisher Vector which is used as a feature Vector of image classification;
(6) using a training image dataset M1Training a classifier by using the Fisher vector and the label corresponding to the Fisher vector to test the image data set M2And inputting the Fisher vector into a trained classifier to classify the image.
Further, the specific process of step (1) is as follows:
(1a) obtaining unlabeled image sets
Figure BDA0001597668740000021
Wherein the content of the first and second substances,
Figure BDA0001597668740000023
representing the ith image, wherein N is the number of unlabeled images;
(1b) from each image
Figure BDA0001597668740000024
Randomly extracting M image blocks of r x r size to obtain MN image blocks in total to form an image block training set
Figure BDA0001597668740000022
n＝r*r*3。
Further, the specific process of step (2) is as follows:
(2a) carrying out noise adding treatment on the image block training set X: randomly pairing image block data X in X according to a ratio qjAdding noise to obtain xj'～q(xj'|xj) Wherein x isj' is to scale q the vector xjSetting a part of values as 0, wherein j belongs to {1, MN };
(2b) random initialization model parameter W1、W2、b1、b2Setting the number of neurons in an input layer of the noise reduction sparse automatic encoder as n and the number of neurons in a hidden layer as s, wherein,
Figure BDA0001597668740000031
the weight of the code is represented by,
Figure BDA0001597668740000032
the weight of the decoding is represented by the weight,
Figure BDA0001597668740000033
a term representing the coding bias is represented,
Figure BDA0001597668740000034
representing a decoding bias term;
let alphaj＝g(W1xj'+b1) Wherein, in the step (A),
Figure BDA0001597668740000035
Figure BDA0001597668740000036
(2c) defining an overall cost function Loss of the noise reduction sparse automatic encoder:
Figure BDA0001597668740000037
where λ is the weight attenuation coefficient, β is the sparse penalty weight, ρ is the target sparse value,
Figure BDA0001597668740000038
is the average response of all training samples on the t-th hidden neuron;
(2d) calculating the gradient of the overall cost function Loss by using an error back propagation algorithm; solving the minimization problem of the overall cost function Loss through an improved quasi-Newton method L-BFGS (Levenberg-Beckmann-Straussler) so as to obtain a trained model parameter W1、b1、W2、b2。
Further, the specific process of step (3) is as follows:
(3a) separately acquiring training image data sets M1And a test image dataset M2Each image of (a);
(3b) performing density space sampling on each image acquired in the step (3a) from left to right and from top to bottom in turn in an iterative manner by using a sliding window with variable dimensions, wherein each image sampling obtains R spatial regions; defining the size of an initial sliding window as [ w, h ], the step length of each sliding of the sliding window as t, and defining the space region acquired by each sliding of the sliding window as a quadruple Area (m, n, w, h); where m and n represent the coordinate positions of the top left corner of the sliding window in the image, and w and h represent the width and height of the initial sliding window.
Further, the specific process of step (4) is as follows:
(4a) randomly extracting from each spatial region of each image obtained in step (3b)
Figure BDA0001597668740000039
Graph of rA block;
(4b) taking the trained noise reduction sparse automatic encoder as a feature extractor to obtain the I-th space region in the image i
Figure BDA0001597668740000046
Local feature set of r image blocks
Figure BDA0001597668740000041
Each image acquires R local feature sets, corresponding to R spatial regions.
Further, the specific steps of step (5) are as follows:
(5a) training image data set M by adopting Gaussian mixture model containing K Gaussian components1Modeling all local feature sets, and solving to obtain a parameter theta of the model;
(5b) on the basis of the step (5a), coding a local feature set of the ith spatial region of the image i by using a Fisher vector of the first layer to obtain a Fisher vector bl(ii) a Obtaining a Fisher vector set of the image i after Fisher vector coding of the first layer
Figure BDA0001597668740000042
(5c) For training image data set M1And a test image dataset M2Performing PCA (principal component analysis) dimensionality reduction on Fisher vector sets of all the images, wherein the Fisher vector set of the image i after dimensionality reduction is
Figure BDA0001597668740000043
(5d) Training image data set M by adopting Gaussian mixture model containing K' Gaussian components1Modeling the Fisher vector set after dimension reduction of all the images, and solving to obtain a parameter theta' of the model;
(5e) on the basis of the step (5d), a Fisher vector set B of the image i after dimensionality reduction is carried out by utilizing a Fisher vector of the second layeri' encoding to obtain Fisher vector fiObtaining a training image data set M1And a test image dataset M2Each figure isLike the final Fisher vector.
Further, in step (5a), a Gaussian mixture model of K Gaussian units is recorded as
Figure BDA0001597668740000044
Wherein, atAs a single local feature, pi(at) Represents the (i) th gaussian unit of the image,
Figure BDA0001597668740000045
d represents a local feature atDimension of (d), pi,μi,∑iRespectively representing the weight, the mean value and the covariance matrix of the ith Gaussian unit; by training the image dataset M1To estimate all parameters theta ═ pi of the gaussian mixture modeli,μi,∑i,i＝1,...K}。
Further, in step (6), the classifier employs a support vector machine.
Adopt the beneficial effect that above-mentioned technical scheme brought:
(1) the invention adopts the noise reduction sparse automatic encoder as the feature extractor, the model is an unsupervised feature learning method, and the invention can use the label-free samples for training and can solve the problem that a large amount of manually marked samples are needed at present. In addition, the noise-reduced sparse auto-encoder is able to learn a more robust feature representation than a normal sparse auto-encoder.
(2) The invention adopts a density space sampling method to replace the traditional space pyramid structure. The traditional space pyramid has fixed image segmentation, and the density space sampling method can more flexibly segment the image, so that more image space information can be captured.
(3) The invention adopts the stacked Fisher vector to carry out feature coding, compared with a standard single-layer structure, the stacked structure can extract the semantic information of the image in a hierarchical structure, and can obtain better precision and performance in image classification.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Detailed Description
The technical scheme of the invention is explained in detail in the following with the accompanying drawings.
The software and hardware environment of the test experiment of this example is as follows:
hardware type:
computer type: a desktop machine;
CPU：Intel(R)Core(TM)i5-5200U CPU@2.20GHz
memory: 8.00GB
The system type is as follows: 64-bit operating system
And (3) developing a language: matlab
The technical scheme of the invention is explained in detail in the following with the accompanying drawings. The example is exemplified by an STL-10 database containing 10 types of RGB images, each 96 x 96 in size. The number of training samples for supervised training is 5000, 5000 training samples are divided into ten folds, the number of training samples for supervised training is 1000 each time, and the number of test samples is 8000.
The invention provides an image classification method based on a noise reduction sparse automatic encoder and density space sampling, which is matched with the image classification method shown in figure 1 and comprises the following specific steps.
Step 1, constructing an image block training set:
1a, obtaining an unlabeled image set in STL-10
Figure BDA0001597668740000061
Wherein the content of the first and second substances,
Figure BDA0001597668740000068
representing the ith image, the number of unlabeled images is 100000;
1b from each image
Figure BDA0001597668740000069
Randomly extracting 10 image blocks with 8-by-8 sizes to form an image block training set
Figure BDA0001597668740000062
Multiplication by 3 is because there are R, G, B three channels.
Step 2, constructing a single hidden layer noise reduction sparse automatic encoder, inputting the image block training set obtained in the step 1b, and training the noise reduction sparse automatic encoder until the iteration stopping condition is met:
2a, adding noise to the image block training set X: randomly pairing image block vector data X in X according to a certain proportion qjAdding noise to obtain xj'～q(xj'|xj). Wherein x isj' is to scale q the vector xjSome of the values are set to 0. Where q is set to 0.5, j ∈ {1,1000000 };
2b, randomly initializing model parameters W1、W2、b1、b2The number of neurons in the input layer of the noise reduction sparse automatic encoder is 192, and the number of neurons in the hidden layer is 500. Wherein the content of the first and second substances,
Figure BDA0001597668740000063
in order to encode the weights, the weights are,
Figure BDA0001597668740000064
in order to decode the weights, the weights are decoded,
Figure BDA0001597668740000065
in order to encode the bias term(s),
Figure BDA0001597668740000066
to decode the bias term. Suppose alphaj＝g(W1xj'+b1) Wherein
Figure BDA0001597668740000067
2c, overall cost function Loss defining DSAE:
Figure BDA0001597668740000071
where λ is the weight attenuation coefficient, β is the sparse penalty weight, ρ is the target sparse value,
Figure BDA0001597668740000072
is the average response of all training samples on the tth hidden neuron, preferably λ 0.03, β 3, ρ 0.1;
and 2d, calculating the gradient of the overall cost function Loss by using an error back propagation algorithm. Solving the minimization problem of the cost function Loss by an improved quasi-Newton method L-BFGS (bidirectional Forwarding-binary scaling) method so as to obtain a trained model parameter W1、b1、W2、b2。
Step 3, training image data set M1And a test image dataset M2Wherein the training image dataset M is subjected to density space sampling1Number of images 1000, test image dataset M2Number of images 8000:
3a, respectively acquiring training image data sets M1And a test image dataset M2Each image of (a);
3b, density spatial sampling is performed on each image in step 3a in an iterative manner from left to right and from top to bottom in turn using a sliding window of variable dimensions. Defining the size of an initial sliding window as [ w, h ], the step length of each sliding of the sliding window is 10, and the space region acquired by each sliding of the sliding window is defined as a quadruple Area (m, n, w, h). Where m and n represent the coordinate positions of the top left corner of the sliding window in the image, and w and h represent the width and height of the initial sliding window. m and n are initially set to 0, w and h are initially set to 46, and each image can be sampled to obtain 90 spatial regions.
Step 4, taking the trained noise reduction sparse automatic encoder as a feature extractor, and extracting local feature set information from a space region obtained by density space sampling of each image:
4a, randomly extracting 100 8 × 8 image blocks from each spatial region of each image obtained in the step 3 c;
4b, taking the trained noise reduction sparse automatic encoder as a feature extractor to obtain local feature sets of 100 8 x 8 image blocks in the ith space region in the image i
Figure BDA0001597668740000073
Each image may capture 90 feature sets.
And 5, encoding the feature set information in the step 4b by using two layers of stacked Fisher vectors to obtain a final Fisher Vector as a feature Vector of image classification:
5a, training image data set M by adopting a Gaussian mixture model containing 128 Gaussian components1Modeling all local feature sets, and solving to obtain a parameter theta of the model; a Gaussian mixture model of 128 Gaussian units is recorded as
Figure BDA0001597668740000081
Wherein, atAs a single local feature, pi(at) Represents the (i) th gaussian unit of the image,
Figure BDA0001597668740000082
d is a local feature atD192, pii,μi,∑iRespectively representing the weight, mean and covariance matrix of the ith gaussian unit. By training the image dataset M1To estimate all parameters theta ═ pi of the gaussian mixture modeli,μi,∑i,i＝1,...128}；
5b, on the basis of the step 5a, coding a local feature set A of the ith spatial region of the image i by using a Fisher vector of the first layer to obtain a Fisher vector bl. After Fisher vector encoding of the first layer, a Fisher vector set of the image i can be obtained
Figure BDA0001597668740000083
5c, for training image data set M1And a test image dataset M2Performing PCA (principal component analysis) dimensionality reduction on Fisher vector sets of all the images, wherein the Fisher vector set of the image i after dimensionality reduction is
Figure BDA0001597668740000084
5d, training the image data set M by adopting a Gaussian mixture model containing 32 Gaussian components1Modeling the Fisher vector set after dimension reduction of all the images, and solving to obtain a parameter theta' of the model;
5e, on the basis of the step 5d, utilizing the Fisher vector of the second layer to perform Fisher vector set B on the dimensionality reduced image ii' encoding to obtain Fisher vector fiFrom which a training image data set M can be derived1And a test image dataset M2Final Fisher vector f for each image.
Step 6, adopting the training image data set M in the step 5e1Training a support vector machine by using the Fisher vectors and the labels corresponding to the Fisher vectors to test the image data set M2The Fisher vector is input into a trained classifier to classify the image.
The embodiments are only for illustrating the technical idea of the present invention, and the technical idea of the present invention is not limited thereto, and any modifications made on the basis of the technical scheme according to the technical idea of the present invention fall within the scope of the present invention.

Claims (4)

1. The image classification method based on the noise reduction sparse automatic encoder and the density space sampling is characterized by comprising the following steps of:
(1) constructing an image block training set; the specific process of the step is as follows:
(1a) obtaining unlabeled image sets
Figure FDA0002390211910000011
Wherein the content of the first and second substances,
Figure FDA0002390211910000012
representing the ith image, wherein N is the number of unlabeled images;
(1b) from each image
Figure FDA0002390211910000013
Randomly extracting M image blocks of r x r size to obtain MN image blocksForming a training set of image blocks
Figure FDA0002390211910000014
(2) Constructing a noise reduction sparse automatic encoder of a single hidden layer, inputting the image block training set obtained in the step (1), and training the noise reduction sparse automatic encoder until an iteration stopping condition is met;
(3) for training image data set M1And a test image dataset M2Performing density space sampling on each image in the image data set, and setting a training image data set M1Number of images of m1Testing the image dataset M2Number of images of m2(ii) a The specific process of the step is as follows:
(3a) separately acquiring training image data sets M1And a test image dataset M2Each image of (a);
(3b) performing density space sampling on each image acquired in the step (3a) from left to right and from top to bottom in turn in an iterative manner by using a sliding window with variable dimensions, wherein each image sampling obtains R spatial regions; defining the size of an initial sliding window as [ w, h ], the step length of each sliding of the sliding window as t, and defining the space region acquired by each sliding of the sliding window as a quadruple Area (m, n, w, h); wherein m and n represent the coordinate position of the upper left corner of the sliding window in the image, and w and h represent the width and height of the initial sliding window;
(4) taking a trained noise reduction sparse automatic encoder as a feature extractor, and extracting local feature set information from a space region obtained by density space sampling of each image; the specific process of the step is as follows:
(4a) randomly extracting from each spatial region of each image obtained in step (3b)
Figure FDA0002390211910000015
Image blocks of r;
(4b) taking the trained noise reduction sparse automatic encoder as a feature extractor to obtain the I-th space region in the image i
Figure FDA0002390211910000021
Local feature set of r image blocks
Figure FDA0002390211910000022
Obtaining R local feature sets of each image, wherein the R local feature sets correspond to R spatial regions;
(5) encoding the feature set information extracted in the step (4) by adopting two layers of stacked Fisher vectors to obtain a final Fisher Vector which is used as a feature Vector of image classification; the specific process of the step is as follows:
(5a) training image data set M by adopting Gaussian mixture model containing K Gaussian components1Modeling all local feature sets, and solving to obtain a parameter theta of the model;
(5b) on the basis of the step (5a), coding a local feature set of the ith spatial region of the image i by using a Fisher vector of the first layer to obtain a Fisher vector bl(ii) a Obtaining a Fisher vector set of the image i after Fisher vector coding of the first layer
Figure FDA0002390211910000023
(5c) For training image data set M1And a test image dataset M2Performing PCA (principal component analysis) dimensionality reduction on Fisher vector sets of all the images, wherein the Fisher vector set of the image i after dimensionality reduction is
Figure FDA0002390211910000024
(5d) Training image data set M by adopting Gaussian mixture model containing K' Gaussian components1Modeling the Fisher vector set after dimension reduction of all the images, and solving to obtain a parameter theta' of the model;
(5e) on the basis of the step (5d), a Fisher vector set B of the image i after dimensionality reduction is carried out by utilizing a Fisher vector of the second layeri' encoding to obtain Fisher vector fiObtaining a training image data set M1And a test image dataset M2Final Fisher vector of each image;
(6) using a training image dataset M1Training a classifier by using the Fisher vector and the label corresponding to the Fisher vector to test the image data set M2And inputting the Fisher vector into a trained classifier to classify the image.
2. The image classification method based on the noise reduction sparse automatic encoder and the density space sampling according to claim 1, wherein the specific process of the step (2) is as follows:
(2a) carrying out noise adding treatment on the image block training set X: randomly pairing image block data X in X according to a ratio qjAdding noise to obtain xj'～q(xj'|xj) Wherein x isj' is to scale q the vector xjSetting a part of values as 0, wherein j belongs to {1, MN };
(2b) random initialization model parameter W1、W2、b1、b2Setting the number of neurons in an input layer of the noise reduction sparse automatic encoder as n and the number of neurons in a hidden layer as s, wherein,
Figure FDA0002390211910000031
the weight of the code is represented by,
Figure FDA0002390211910000032
the weight of the decoding is represented by the weight,
Figure FDA0002390211910000033
a term representing the coding bias is represented,
Figure FDA0002390211910000034
representing a decoding bias term;
let alphaj＝g(W1xj'+b1) Wherein, in the step (A),
Figure FDA0002390211910000035
g(z)＝1/(1+exp(-z))，
Figure FDA0002390211910000036
Figure FDA0002390211910000037
(2c) defining an overall cost function Loss of the noise reduction sparse automatic encoder:
Figure FDA0002390211910000038
where λ is the weight attenuation coefficient, β is the sparse penalty weight, ρ is the target sparse value,
Figure FDA0002390211910000039
is the average response of all training samples on the t-th hidden neuron;
(2d) calculating the gradient of the overall cost function Loss by using an error back propagation algorithm; solving the minimization problem of the overall cost function Loss through an improved quasi-Newton method L-BFGS (Levenberg-Beckmann-Straussler) so as to obtain a trained model parameter W1、b1、W2、b2。
3. The method for classifying an image based on a noise reduction sparse automatic encoder and density spatial sampling as claimed in claim 1, wherein in step (5a), a Gaussian mixture model of K Gaussian units is recorded as
Figure FDA00023902119100000310
Wherein, atAs a single local feature, pi(at) Represents the (i) th gaussian unit of the image,
Figure FDA00023902119100000311
d represents a local feature atDimension of (d), pi,μi,∑iRespectively representing the weight, the mean value and the covariance matrix of the ith Gaussian unit; by training the image dataset M1To estimate all parameters of the gaussian mixture modelθ＝{πi,μi,∑i,i＝1,...K}。
4. The method for classifying an image based on a noise reduction sparse automatic encoder and density spatial sampling according to claim 1, wherein in the step (6), the classifier adopts a support vector machine.
CN201810212714.7A 2018-03-15 2018-03-15 Image classification method based on noise reduction sparse automatic encoder and density space sampling Active CN108416389B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810212714.7A CN108416389B (en) 2018-03-15 2018-03-15 Image classification method based on noise reduction sparse automatic encoder and density space sampling

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810212714.7A CN108416389B (en) 2018-03-15 2018-03-15 Image classification method based on noise reduction sparse automatic encoder and density space sampling

Publications (2)

Publication Number Publication Date
CN108416389A CN108416389A (en) 2018-08-17
CN108416389B true CN108416389B (en) 2020-10-30

Family

ID=63131586

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810212714.7A Active CN108416389B (en) 2018-03-15 2018-03-15 Image classification method based on noise reduction sparse automatic encoder and density space sampling

Country Status (1)

Country Link
CN (1) CN108416389B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109389171B (en) * 2018-10-11 2021-06-25 云南大学 Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology
CN115496761B (en) * 2022-11-17 2023-03-03 湖南自兴智慧医疗科技有限公司 AE-based method and system for phase-splitting screening of low power lens and storage medium

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8553994B2 (en) * 2008-02-05 2013-10-08 Futurewei Technologies, Inc. Compressive sampling for multimedia coding
CN105404858A (en) * 2015-11-03 2016-03-16 电子科技大学 Vehicle type recognition method based on deep Fisher network
CN105868796B (en) * 2016-04-26 2019-03-01 中国石油大学(华东) The design method of linear discriminant rarefaction representation classifier based on nuclear space
CN107133640A (en) * 2017-04-24 2017-09-05 河海大学 Image classification method based on topography's block description and Fei Sheer vectors
CN107609648B (en) * 2017-07-21 2021-02-12 哈尔滨工程大学 Genetic algorithm combined with stacking noise reduction sparse automatic encoder

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Action Recognition with Stacked Fisher Vectors;Xiaojiang Peng等;《European Conference on Computer Vision-ECCV2014》;20140912;第581-595页 *
Human Action Recognition from Local Part Model;Feng Shi等;《2011 IEEE International Workshop on Haptic Audio Visual Environments and Games》;20111017;第1-4页 *
融入形状信息的图像分类研究;嵇朋朋;《中国优秀硕士学位论文全文数据库信息科技辑》;20150715;第I138-1387页 *

Also Published As

Publication number Publication date
CN108416389A (en) 2018-08-17

Similar Documents

Publication Publication Date Title
CN107122809B (en) Neural network feature learning method based on image self-coding
CN108804530B (en) Subtitling areas of an image
CN109345508B (en) Bone age evaluation method based on two-stage neural network
CN111160533B (en) Neural network acceleration method based on cross-resolution knowledge distillation
CN108108751B (en) Scene recognition method based on convolution multi-feature and deep random forest
Kong Facial expression recognition method based on deep convolutional neural network combined with improved LBP features
CN110188827B (en) Scene recognition method based on convolutional neural network and recursive automatic encoder model
CN111475622A (en) Text classification method, device, terminal and storage medium
Sekma et al. Human action recognition based on multi-layer fisher vector encoding method
Yang et al. STA-TSN: Spatial-temporal attention temporal segment network for action recognition in video
CN111079514A (en) Face recognition method based on CLBP and convolutional neural network
Kastaniotis et al. HEp-2 cell classification with vector of hierarchically aggregated residuals
Ouyed et al. Feature weighting for multinomial kernel logistic regression and application to action recognition
Jin et al. Multiple graph regularized sparse coding and multiple hypergraph regularized sparse coding for image representation
Song et al. MPPCANet: A feedforward learning strategy for few-shot image classification
Fan et al. A hierarchical Dirichlet process mixture of generalized Dirichlet distributions for feature selection
CN108416389B (en) Image classification method based on noise reduction sparse automatic encoder and density space sampling
Tabernik et al. Towards deep compositional networks
CN112163114A (en) Image retrieval method based on feature fusion
Yin et al. Object recognition in art drawings: Transfer of a neural network
CN108388918B (en) Data feature selection method with structure retention characteristics
CN113378021A (en) Information entropy principal component analysis dimension reduction method based on semi-supervision
CN117115817A (en) Cell morphology identification method and device based on multi-mode fusion
Fan et al. [Retracted] Accurate Recognition and Simulation of 3D Visual Image of Aerobics Movement
CN114882288B (en) Multi-view image classification method based on hierarchical image enhancement stacking self-encoder

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