CN110991554A - Improved PCA (principal component analysis) -based deep network image classification method - Google Patents

Improved PCA (principal component analysis) -based deep network image classification method Download PDF

Info

Publication number
CN110991554A
CN110991554A CN201911291420.9A CN201911291420A CN110991554A CN 110991554 A CN110991554 A CN 110991554A CN 201911291420 A CN201911291420 A CN 201911291420A CN 110991554 A CN110991554 A CN 110991554A
Authority
CN
China
Prior art keywords
image
features
pca
reduction processing
dimension reduction
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
CN201911291420.9A
Other languages
Chinese (zh)
Other versions
CN110991554B (en
Inventor
蒋强
陈凯
冯永新
隋涛
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.)
Shenyang Ligong University
Original Assignee
Shenyang Ligong 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 Shenyang Ligong University filed Critical Shenyang Ligong University
Priority to CN201911291420.9A priority Critical patent/CN110991554B/en
Publication of CN110991554A publication Critical patent/CN110991554A/en
Application granted granted Critical
Publication of CN110991554B publication Critical patent/CN110991554B/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
    • G06F18/24—Classification techniques
    • 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
    • G06F18/211—Selection of the most significant subset of features
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (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)

Abstract

The invention discloses a depth network image classification method based on improved PCA, which comprises the steps of firstly extracting features of an input image by using a depth convolution neural network, then screening the features by calculating image information entropy and setting an image information entropy threshold value, and reducing the dimension of the screened features by PCA, further simplifying the image features and improving the feature quality, thereby effectively solving the problem of too low identification speed caused by too high data dimension in the process of classifying a large data image set, and simultaneously effectively reducing the huge calculation amount of a covariance matrix calculated by the PCA in the process of reducing the dimension of the features by screening the features by the image information entropy, and improving the real-time property of image classification.

Description

Improved PCA (principal component analysis) -based deep network image classification method
Technical Field
The invention belongs to the technical field of image processing and pattern recognition, and particularly relates to a depth network image classification method based on improved PCA.
Background
In the current society, as the cloud era comes, big data attracts more and more attention, and images as the main expression form of data information become important means for people to acquire information due to the characteristics of rich content, visual reflection and the like, and the quantity of the images is rapidly increasing at an astonishing speed. However, the problem of disorder of image information becomes increasingly prominent with the increase of image data. How to automatically identify, retrieve and classify massive image data by using an artificial intelligence technology has become a research focus in the field of computer vision identification at present.
The traditional image classification method, such as Scale Invariant Feature Transform (SIFT), Histogram of Oriented Gradients (HOG), etc., has shallow structure hierarchy and small calculation amount, and can complete model training and analysis without taking a large number of images as a basis. However, the traditional model cannot acquire semantic features and depth features of higher layers from the original image, and image features are not easy to extract under the condition of large data. With the rise of deep networks, many excellent image classification methods based on deep learning emerge, for example: AlexNet, VGGNet, google lenet, ResNet, etc. The deep learning identification method can obtain deeper image features, the image feature expression is richer, the image feature extraction is more accurate, excellent classification results are obtained, and the classification effect of partial deep networks even exceeds the precision of human beings.
The image classification is widely applied in many fields, and the application in the aspects of biological feature recognition, intelligent transportation, medical auxiliary diagnosis and the like brings great convenience to our lives, but the deep learning method still has the problems of large amount of images as a basis, large calculation amount, long model training time, high requirement on hardware environment, long classification process and the like, and along with the solution of the problems, the deep learning plays a greater role in the image classification field.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a depth network image classification method based on improved PCA, which comprises the following steps:
step 1: inputting m images in a CIFAR-100 image data set into a deep convolutional neural network model, carrying out gray scale and filtering pretreatment, and eliminating noise interference to obtain the original characteristics of each image;
step 2: constructing a feature extraction module in a deep convolutional neural network model, and extracting the image features of each image by using the constructed feature extraction module;
and step 3: carrying out improved PCA dimension reduction processing on the image features of each image, wherein the improved PCA dimension reduction processing means that firstly, the image features of each image are preliminarily screened through image information entropy, and then, the PCA dimension reduction processing is carried out on the preliminarily screened image features, and the method is specifically represented as follows:
step 3.1: calculating the image information entropy H of the extracted image features by using a formula (1), and primarily screening the image features according to an information entropy threshold;
Figure BDA0002319189840000021
where H denotes the entropy of the image information of the image feature, psRepresenting the probability value corresponding to the s-th gray value in each image, wherein n represents the total number of the input gray values of each image;
step 3.2: carrying out PCA (principal component analysis) dimension reduction processing on the image characteristics of each image obtained by primary screening;
and 4, step 4: inputting the image characteristics Y subjected to the dimension reduction processing into a Softmax classifier to finish classification processing, and outputting the result of image classification as a real classification result;
and 5: selecting a loss function of the deep convolutional neural network model during training as a cross entropy loss function, and calculating a difference value H (p, q) between a prediction classification result and a real result according to the cross entropy loss function, wherein the cross entropy loss function is expressed as:
Figure BDA0002319189840000022
wherein p (m) represents the predicted classification result, and q (m) represents the true classification result;
step 6: if the difference value H (p, q) between the predicted classification result and the real result is larger than the expected difference value H' (p, q), reversely propagating the cross entropy loss function through a deep convolutional network back propagation algorithm, and continuously adjusting the network weight wuvUntil the difference H (p, q) between the calculated predicted classification result and the true result is less than or equal to the desired difference H' (p, q), or is reachedAnd until a preset training time M, wherein u represents the u-th neuron of the previous layer, and v represents the v-th neuron of the next layer.
The feature extraction module in the step 2 is specifically expressed as follows: designing the number of layers of a deep convolutional neural network model to be 13 layers according to a CIFAR-100 image dataset, wherein the 13 layers comprise L2～L14Layer of which L1Represents an input layer, L2～L14The number of the layer convolution kernels is respectively 64, 64, 128, 128, 256, 256, 256, 512, 512, 512, 512, 512, 512 and L in sequence2～L14The convolution kernel size of the layers is 3 x 3, the pooling mode is chosen as Maxpool, and the activation function is chosen as ReLU.
The step 3.2 is specifically expressed as:
3.2.1) centralizing the image characteristic matrix of each image obtained by primary screening;
3.2.2) calculating the covariance of different dimensions of the image features after the centralization treatment, and forming a covariance matrix with the dimension m;
3.2.3) calculating the eigenvalue of each covariance matrix and the eigenvector corresponding to the eigenvalue;
3.2.4) firstly determining the value of the contribution rate f according to the actual situation, and then determining K characteristic values needing to be reserved by using a formula (3);
Figure BDA0002319189840000031
in the formula, λjRepresents the j-th eigenvalue to be preserved in each covariance matrix, K represents the total number of eigenvalues to be preserved in each covariance matrix, and lambdaiThe ith eigenvalue of each covariance matrix is represented, and m represents the total number of eigenvalues in each covariance matrix;
3.2.5) forming a transformation basis P by the preserved eigenvectors corresponding to the K eigenvalues, and finishing the dimension reduction processing by using a formula (4) according to the transformation basis P;
Y＝PTX (4)
in the formula, Y represents the image feature after the dimension reduction processing，PTA transpose matrix representing the exchange base P and X representing the image feature after the centering process.
The invention has the beneficial effects that:
1) the invention introduces PCA dimension reduction on the basis of the deep convolutional neural network, and reduces the characteristic dimension.
2) The invention further introduces the image information entropy, thereby effectively reducing the huge calculation amount and hardware occupation of the calculation covariance matrix in the PCA dimension reduction process.
3) According to the invention, Softmax is used for replacing a full connection layer of a traditional deep network, so that the classification speed is effectively improved.
Drawings
FIG. 1 is a schematic structural diagram of a deep convolutional neural network model based on improved PCA in the present invention.
FIG. 2 is a comparison graph of image classification time before and after the improved PCA dimensionality reduction process of the present invention.
FIG. 3 is a comparison graph of classification accuracy of 4 experimental images before and after the improved PCA dimensionality reduction process in the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific examples.
As shown in fig. 1, according to a structural schematic diagram of a built depth convolution neural network model based on improved PCA, a depth network image classification method based on improved PCA first extracts features of an input image using a depth convolution neural network, then filters the features by calculating an image information entropy and setting an image information entropy threshold, dimensionality reduction is performed on the filtered features through PCA (principal component analysis), image features are further simplified, feature quality is improved, and finally the filtered features are sent to a Softmax classifier to realize classification, which specifically comprises the following steps:
step 1: 5000 images in a CIFAR-100 image data set are input into a deep convolution neural network model, gray level and filtering preprocessing is carried out, noise interference is eliminated, the original characteristics of each image are obtained, and different classification labels are attached to each image and used as prediction classification results of the original characteristics;
step 2: constructing a feature extraction module in a deep convolutional neural network model, and extracting the image features of each image by using the constructed feature extraction module;
as shown in fig. 1, the feature extraction module is specifically expressed as: designing the number of layers of a deep convolutional neural network model to be 13 layers according to a CIFAR-100 image dataset, wherein the 13 layers comprise L2～L14Layer of which L1Represents an input layer, L2～L14The number of the layer convolution kernels is respectively 64, 64, 128, 128, 256, 256, 256, 512, 512, 512, 512, 512, 512 and L in sequence2～L14The convolution kernel size of the layers is 3 x 3, the pooling mode is chosen as Maxpool, and the activation function is chosen as ReLU.
And step 3: carrying out improved PCA dimension reduction processing on the image features of each image, wherein the improved PCA dimension reduction processing means that firstly, the image features of each image are preliminarily screened through image information entropy, and then, the PCA dimension reduction processing is carried out on the preliminarily screened image features, and the method is specifically represented as follows:
step 3.1: calculating the image information entropy H of the extracted image features by using a formula (1), and primarily screening the image features according to an information entropy threshold, wherein the information entropy threshold is set to be 0.15 in the embodiment;
Figure BDA0002319189840000041
where H denotes the entropy of the image information of the image feature, psRepresenting the probability value corresponding to the s-th gray value in each image, wherein n represents the total number of the input gray values of each image;
step 3.2: carrying out PCA (principal component analysis) dimension reduction processing on the image features of each image obtained by primary screening, wherein the specific expression is as follows:
3.2.1) centralizing the image characteristic matrix of each image obtained by primary screening;
3.2.2) calculating the covariance of different dimensions of the image features after the centralization treatment, and forming a covariance matrix with the dimension m;
3.2.3) calculating the eigenvalue of each covariance matrix and the eigenvector corresponding to the eigenvalue;
3.2.4) determining the contribution rate f to be 90% in the embodiment, and then determining K characteristic values needing to be reserved by using a formula (3);
Figure BDA0002319189840000042
in the formula, λjRepresents the j-th eigenvalue to be preserved in each covariance matrix, K represents the total number of eigenvalues to be preserved in each covariance matrix, and lambdaiThe ith eigenvalue of each covariance matrix is represented, and m represents the total number of eigenvalues in each covariance matrix;
3.2.5) forming a transformation basis P by the preserved eigenvectors corresponding to the K eigenvalues, and finishing the dimension reduction processing by using a formula (4) according to the transformation basis P;
Y＝PTX (4)
wherein Y represents an image feature after dimension reduction processing, PTA transpose matrix representing the exchange base P and X representing the image feature after the centering process.
And 4, step 4: inputting the image characteristics Y subjected to the dimension reduction processing into a Softmax classifier to finish classification processing, and outputting the result of image classification as a real classification result;
and 5: selecting a loss function of the deep convolutional neural network model during training as a cross entropy loss function, and calculating a difference value H (p, q) between a prediction classification result and a real result according to the cross entropy loss function, wherein the cross entropy loss function is expressed as:
Figure BDA0002319189840000051
wherein p (m) represents the predicted classification result, and q (m) represents the true classification result;
step 6: if the difference value H (p, q) between the predicted classification result and the real result is larger than the expected difference value H' (p, q), reversely propagating the cross entropy loss function through a deep convolutional network back propagation algorithm, and continuously adjusting the network weight wuvUntil the difference H (p, q) between the calculated predicted classification result and the real result is less than or equal to the expected difference H' (p, q), or a preset training number M is reached, wherein u represents the u-th neuron of the previous layer and v represents the v-th neuron of the next layer.
As shown in fig. 2 and fig. 3, the technical scheme can effectively reduce the feature dimension and the calculation amount in the feature extraction stage, improve the image classification effect, and obviously improve the image classification time and the accuracy after the improved PCA dimension reduction processing.

Claims (3)

1. A depth network image classification method based on improved PCA is characterized by comprising the following steps:
step 1: inputting m images in a CIFAR-100 image data set into a deep convolutional neural network model, carrying out gray scale and filtering pretreatment, and eliminating noise interference to obtain the original characteristics of each image;
step 2: constructing a feature extraction module in a deep convolutional neural network model, and extracting the image features of each image by using the constructed feature extraction module;
and step 3: carrying out improved PCA dimension reduction processing on the image features of each image, wherein the improved PCA dimension reduction processing means that firstly, the image features of each image are preliminarily screened through image information entropy, and then, the PCA dimension reduction processing is carried out on the preliminarily screened image features, and the method is specifically represented as follows:
step 3.1: calculating the image information entropy H of the extracted image features by using a formula (1), and primarily screening the image features according to an information entropy threshold;
Figure FDA0002319189830000011
where H denotes the entropy of the image information of the image feature, psRepresenting the probability value corresponding to the s-th gray value in each image, wherein n represents the total number of the input gray values of each image;
step 3.2: carrying out PCA (principal component analysis) dimension reduction processing on the image characteristics of each image obtained by primary screening;
and 4, step 4: inputting the image characteristics Y subjected to the dimension reduction processing into a Softmax classifier to finish classification processing, and outputting the result of image classification as a real classification result;
and 5: selecting a loss function of the deep convolutional neural network model during training as a cross entropy loss function, and calculating a difference value H (p, q) between a prediction classification result and a real result according to the cross entropy loss function, wherein the cross entropy loss function is expressed as:
Figure FDA0002319189830000012
wherein p (m) represents the predicted classification result, and q (m) represents the true classification result;
step 6: if the difference value H (p, q) between the predicted classification result and the real result is larger than the expected difference value H' (p, q), reversely propagating the cross entropy loss function through a deep convolutional network back propagation algorithm, and continuously adjusting the network weight wuvUntil the difference H (p, q) between the calculated predicted classification result and the real result is less than or equal to the expected difference H' (p, q), or a preset training number M is reached, wherein u represents the u-th neuron of the previous layer and v represents the v-th neuron of the next layer.
2. The method for classifying depth network images based on improved PCA as claimed in claim 1, wherein the feature extraction module in the step 2 is specifically expressed as: designing the number of layers of a deep convolutional neural network model to be 13 layers according to a CIFAR-100 image dataset, wherein the 13 layers comprise L2～L14Layer of which L1Represents an input layer, L2～L14The number of the layer convolution kernels is respectively 64, 64, 128, 128, 256, 256, 256, 512, 512, 512, 512, 512, 512 and L in sequence2～L14The convolution kernel size of the layers is 3 x 3, the pooling mode is chosen as Maxpool, and the activation function is chosen as ReLU.
3. The method for classifying depth network images based on improved PCA as claimed in claim 1, wherein said step 3.2 is specifically expressed as:
3.2.1) centralizing the image characteristic matrix of each image obtained by primary screening;
3.2.2) calculating the covariance of different dimensions of the image features after the centralization treatment, and forming a covariance matrix with the dimension m;
3.2.3) calculating the eigenvalue of each covariance matrix and the eigenvector corresponding to the eigenvalue;
3.2.4) firstly determining the value of the contribution rate f according to the actual situation, and then determining K characteristic values needing to be reserved by using a formula (3);
Figure FDA0002319189830000021
in the formula, λjRepresents the j-th eigenvalue to be preserved in each covariance matrix, K represents the total number of eigenvalues to be preserved in each covariance matrix, and lambdaiThe ith eigenvalue of each covariance matrix is represented, and m represents the total number of eigenvalues in each covariance matrix;
3.2.5) forming a transformation basis P by the preserved eigenvectors corresponding to the K eigenvalues, and finishing the dimension reduction processing by using a formula (4) according to the transformation basis P;
Y＝PTX (4)
wherein Y represents an image feature after dimension reduction processing, PTA transpose matrix representing the exchange base P and X representing the image feature after the centering process.
CN201911291420.9A 2019-12-16 2019-12-16 Improved PCA (principal component analysis) -based deep network image classification method Active CN110991554B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911291420.9A CN110991554B (en) 2019-12-16 2019-12-16 Improved PCA (principal component analysis) -based deep network image classification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911291420.9A CN110991554B (en) 2019-12-16 2019-12-16 Improved PCA (principal component analysis) -based deep network image classification method

Publications (2)

Publication Number Publication Date
CN110991554A true CN110991554A (en) 2020-04-10
CN110991554B CN110991554B (en) 2023-04-18

Family

ID=70093848

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911291420.9A Active CN110991554B (en) 2019-12-16 2019-12-16 Improved PCA (principal component analysis) -based deep network image classification method

Country Status (1)

Country Link
CN (1) CN110991554B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111583217A (en) * 2020-04-30 2020-08-25 深圳开立生物医疗科技股份有限公司 Tumor ablation curative effect prediction method, device, equipment and computer medium
CN111738318A (en) * 2020-06-11 2020-10-02 大连理工大学 Super-large image classification method based on graph neural network
CN117520824A (en) * 2024-01-03 2024-02-06 浙江省白马湖实验室有限公司 Information entropy-based distributed optical fiber data characteristic reconstruction method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017215284A1 (en) * 2016-06-14 2017-12-21 山东大学 Gastrointestinal tumor microscopic hyper-spectral image processing method based on convolutional neural network
CN109242028A (en) * 2018-09-19 2019-01-18 西安电子科技大学 SAR image classification method based on 2D-PCA and convolutional neural networks

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017215284A1 (en) * 2016-06-14 2017-12-21 山东大学 Gastrointestinal tumor microscopic hyper-spectral image processing method based on convolutional neural network
CN109242028A (en) * 2018-09-19 2019-01-18 西安电子科技大学 SAR image classification method based on 2D-PCA and convolutional neural networks

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
宋丹;石勇;邓宸伟: "一种结合PCA与信息熵的SIFT特征向量自适应降维算法" *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111583217A (en) * 2020-04-30 2020-08-25 深圳开立生物医疗科技股份有限公司 Tumor ablation curative effect prediction method, device, equipment and computer medium
CN111738318A (en) * 2020-06-11 2020-10-02 大连理工大学 Super-large image classification method based on graph neural network
CN117520824A (en) * 2024-01-03 2024-02-06 浙江省白马湖实验室有限公司 Information entropy-based distributed optical fiber data characteristic reconstruction method

Also Published As

Publication number Publication date
CN110991554B (en) 2023-04-18

Similar Documents

Publication Publication Date Title
CN108615010B (en) Facial expression recognition method based on parallel convolution neural network feature map fusion
CN110532900B (en) Facial expression recognition method based on U-Net and LS-CNN
CN110991554B (en) Improved PCA (principal component analysis) -based deep network image classification method
CN110097060B (en) Open set identification method for trunk image
CN107918772B (en) Target tracking method based on compressed sensing theory and gcForest
CN101556600B (en) Method for retrieving images in DCT domain
CN107169117B (en) Hand-drawn human motion retrieval method based on automatic encoder and DTW
Tereikovskyi et al. The method of semantic image segmentation using neural networks
CN106650804B (en) A kind of face sample cleaning method and system based on deep learning feature
CN109710804B (en) Teaching video image knowledge point dimension reduction analysis method
CN110321805B (en) Dynamic expression recognition method based on time sequence relation reasoning
Yue et al. Face recognition based on histogram equalization and convolution neural network
CN109344720B (en) Emotional state detection method based on self-adaptive feature selection
CN108596044B (en) Pedestrian detection method based on deep convolutional neural network
CN111126155B (en) Pedestrian re-identification method for generating countermeasure network based on semantic constraint
CN116110118A (en) Pedestrian re-recognition and gait recognition method based on space-time feature complementary fusion
Liu et al. Image retrieval using CNN and low-level feature fusion for crime scene investigation image database
Julianto et al. A performance evaluation of convolutional neural network architecture for classification of rice leaf disease
CN110490028A (en) Recognition of face network training method, equipment and storage medium based on deep learning
CN114882278A (en) Tire pattern classification method and device based on attention mechanism and transfer learning
CN108710836B (en) Lip detection and reading method based on cascade feature extraction
Zhang et al. Application of convolution neural network in Iris recognition technology
CN116884067B (en) Micro-expression recognition method based on improved implicit semantic data enhancement
CN105844299B (en) A kind of image classification method based on bag of words
CN111401434A (en) Image classification method based on unsupervised feature learning

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