CN111488905B - Robust image recognition method based on high-dimensional PCANet - Google Patents

Robust image recognition method based on high-dimensional PCANet Download PDF

Info

Publication number
CN111488905B
CN111488905B CN202010147000.XA CN202010147000A CN111488905B CN 111488905 B CN111488905 B CN 111488905B CN 202010147000 A CN202010147000 A CN 202010147000A CN 111488905 B CN111488905 B CN 111488905B
Authority
CN
China
Prior art keywords
feature
dimensional
image
pcanet
atlas
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
CN202010147000.XA
Other languages
Chinese (zh)
Other versions
CN111488905A (en
Inventor
李小薪
徐晨雅
胡海根
周乾伟
郝鹏翼
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.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
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 Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN202010147000.XA priority Critical patent/CN111488905B/en
Publication of CN111488905A publication Critical patent/CN111488905A/en
Application granted granted Critical
Publication of CN111488905B publication Critical patent/CN111488905B/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
    • G06F18/24—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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • 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
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

A robust image recognition method based on high-dimensional PCANet comprises robust feature extraction and nearest neighbor classification based on chi-square distance, wherein the robust feature extraction process combines flat convolution and three-dimensional convolution of a feature map, the three-dimensional convolution fully considers correlation among channels, the flat convolution can fully decompose main directions of each channel of an input image, and the obtained pattern map has richer features compared with the original PCANet and can effectively improve the robustness of the PCANet; the classification process comprises the following steps: step 1, acquiring distance measurement from an image to be identified to each training image based on chi-square distance in a high-dimensional histogram feature space; and 2, obtaining a class mark corresponding to the training sample with the minimum distance measurement as the class mark of the image to be identified. The method and the device can effectively process the changes such as shielding, illumination change, resolution difference and the like in the image to be identified, and effectively improve the identification rate of the offset image.

Description

Robust image recognition method based on high-dimensional PCANet
Technical Field
The invention relates to the field of image processing and pattern recognition, in particular to robust image recognition with large difference between an image to be recognized and a training image, which is mainly used for processing and recognizing images in reality.
Background
In the existing field of computer vision and image recognition, deep neural networks (Deep Neural Network, DNN), represented by convolutional neural networks (Convolutional Neural Networks, CNN), have met with great success, and on some disclosed data sets, the classification capabilities of leading edge deep learning methods even exceed those of humans, for example: authentication accuracy on LFW face database, image classification accuracy on ImageNet, and handwriting digital recognition accuracy on MNIST, etc. However, in practice, the image to be identified tends to have a large difference in "distribution" or "structure" from the training image, which can cause DNN to suffer from a large-scale recognition error, which phenomenon is called "Covariate Shift" in the field of deep learning. The covariate offset causes the technical defects of lower accuracy and poor feasibility of the existing image recognition method.
Disclosure of Invention
In order to overcome the defects of low image recognition accuracy and poor feasibility caused by the existing covariate offset, the invention provides a robust image recognition method with High accuracy and good feasibility based on High-dimensional PCANet (HPCANet), which can effectively overcome the recognition problem caused by the covariate offset, and particularly can greatly improve the image recognition performance when the images to be recognized have offset with larger amplitudes such as shielding, illumination change, resolution difference and the like.
The technical scheme adopted for solving the technical problems is as follows:
a robust image recognition method based on high-dimensional PCANet comprises the following steps:
step 1 selecting J images A= { A 1 ,…,A J As training set, the corresponding class label is
Figure SMS_1
Z＝{Z 1 ,…,Z K The number is the set of images to be identified, i.e. the test set, here +.>
Figure SMS_2
Respectively represent the C on the real number domain 0 An image with length and width of m x n of E {1,3} channels;
step 2, initializing parameters and input data: order the
Figure SMS_3
Here, a->
Figure SMS_4
For indicating the stage at which the network is located,
Figure SMS_5
indicating that the network is in training phase->
Figure SMS_6
Indicating that the network is in a testing stage; let l=0, where l is used to indicate the number of layers of the input image or feature map in the network,/>
Figure SMS_7
Wherein (1)>
Figure SMS_8
N＝J；
Step 3 consists of
Figure SMS_9
Construction of matrix->
Figure SMS_10
Figure SMS_11
Wherein,,
Figure SMS_12
Figure SMS_13
is->
Figure SMS_14
Mean value of->
Figure SMS_15
Figure SMS_16
Representing from->
Figure SMS_17
B e {1,2, …, mn } feature blocks of size k×k extracted from the c-th channel, vec (·) represents the operation of stretching the matrix into column vectors;
step 4 if
Figure SMS_18
If the network is in the test stage, jumping to the step 7, otherwise, executing the next step;
step 5 calculation
Figure SMS_19
Main direction->
Figure SMS_20
Wherein (1)>
Figure SMS_21
Is covariance matrix->
Figure SMS_22
The i' th eigenvector of (a), the corresponding eigenvalue is lambda i′ And->
Figure SMS_23
Step 6 from V (l) Acquisition of C l+1 Three-dimensional filter bank
Figure SMS_24
Step 7 computing the feature atlas X of the 1+1th convolution layer (l+1) ；
Step 8, let l=l+1, execute the above steps 3 to 7 until l=l, where L represents the maximum convolution layer number given in advance;
step 9, initializing parameters and input data: order the
Figure SMS_25
l＝0，/>
Figure SMS_26
Wherein Y is i (l) ＝A i ，N＝J；
Step 10 consists of
Figure SMS_29
Construction of matrix->
Figure SMS_31
Figure SMS_32
Wherein (1)>
Figure SMS_28
Figure SMS_30
Is->
Figure SMS_33
Is used for the average value of (a),
Figure SMS_34
Figure SMS_27
representing the slave Y i (l) B e {1,2, …, mn } feature blocks of size k x k extracted from the c-th channel;
step 11 if
Figure SMS_35
Then jump to step 14, otherwise, execute the next step;
step 12 calculation
Figure SMS_36
Main direction->
Figure SMS_37
Wherein (1)>
Figure SMS_38
Is covariance matrix->
Figure SMS_39
The i "th eigenvector of (a), the corresponding eigenvalue is lambda i″ And->
Figure SMS_40
Step 13 consists of
Figure SMS_41
Acquisition of C l+1 Three-dimensional filter bank
Figure SMS_42
Step 14 calculates a feature atlas Y for the 1+1th convolution layer (l+1) ；
Step 15, let l=l+1, execute the above steps 10 to 14 until l=l;
step 16 sets the feature atlas X (L) And Y (L) Combining to form a new feature atlas F:
Figure SMS_43
step 17, performing pattern diagram coding on the feature atlas F to obtain a pattern atlas P: p= { P i,β } i＝1,…,N；β＝1,…,B Wherein, the method comprises the steps of, wherein,
Figure SMS_44
beta e {1, …, B } pattern diagram representing the ith sample, F i,· Representing feature map subset F i In>
Figure SMS_45
T represents the number of channels involved in the encoding of a single pattern diagram, USF (·) represents a unit step function (Unit Step Function, USF), and the input value is binarized by comparison with 0, i.e.:
Figure SMS_46
step 18 extracts histogram features H from the pattern atlas P: h= [ H ] i ] i＝1,…,N Wherein H is i ＝[H i,1 ,…,H i,B ] T ，H i,β ＝Qhist(P i,β )，Qhist(P i,β ) Representing a pattern diagramP i,β Divided into Q blocks, a histogram is extracted from each block, each histogram using 2 T The number of packets, i.e. the code value of the statistical pattern diagram is 2 for each feature block T The frequency of occurrence in the individual packets;
step 19 if
Figure SMS_47
Make H Te =h, jump to step 21; otherwise, let H Tr =h, perform the next step;
step 20 order
Figure SMS_48
l＝0，/>
Figure SMS_49
Wherein n=k,>
Figure SMS_50
executing the steps 3 to 19;
step 21 calculates a metric matrix m= [ M ] i,j ] i＝1,…,J；j＝1,…,K Wherein, the method comprises the steps of, wherein,
Figure SMS_51
here the number of the elements is the number,
Figure SMS_52
wherein D represents
Figure SMS_53
And->
Figure SMS_54
Length of->
Figure SMS_55
Representation->
Figure SMS_56
The d element of (a)>
Figure SMS_57
Representation of
Figure SMS_58
The d element of (a);
step 22 calculates class id= [ Id ] of each sample in the test set Z i ] i＝1,…,K ：
Figure SMS_59
Wherein M is i Represents the ith column vector in the metric matrix M, minIndx (·) represents M i Index of the smallest element in the (c).
Further, in the step 7, the feature atlas X of the (1) th convolution layer is calculated according to the following steps (l+1) :7.1 Will) be
Figure SMS_62
Projected to W (l+1) ：/>
Figure SMS_64
7.2 Will->
Figure SMS_66
The elements in (a) are reorganized into feature atlas +.>
Figure SMS_61
Wherein (1)>
Figure SMS_63
And is also provided with
Figure SMS_69
Figure SMS_70
Here, a->
Figure SMS_60
Representation->
Figure SMS_65
Column vectors of rows a to b of column c, a% b representing a remainder of b,/-a%>
Figure SMS_67
Representation rounding down the real number a, mat m×n (v) represents that an arbitrary column vector is +.>
Figure SMS_68
Rearranged into an mxn matrix.
Still further, in the step 14, the feature atlas Y of the (1+1) th convolution layer is calculated as follows (l+1) :14.1 Will) be
Figure SMS_71
Projected to W (l+1) ：/>
Figure SMS_75
14.2 Will->
Figure SMS_77
The elements in (a) are reorganized into feature atlas +.>
Figure SMS_73
Wherein (1)>
Figure SMS_74
Figure SMS_76
And->
Figure SMS_78
Here the number of the elements is the number,
Figure SMS_72
the representation connects the matrices in the set in the channel direction.
The technical conception of the invention is as follows: when the images to be identified and the training set images have large offsets such as shielding, illumination change, resolution difference and the like, the identification performance of the existing neural network model is often greatly reduced, and PCANet can better solve the problems. However, PCANet suffers from two drawbacks: (1) The PCANet adopts flat convolution, and correlation among channels of the feature map is not fully considered; (2) PCANet compresses the generated feature map 8 times when encoding the pattern map, so that the acquired pattern map lacks rich discriminative features. In order to solve the above problems, the invention combines the stereo convolution and the flat convolution, wherein the stereo convolution can fully consider the correlation between channels, and the flat convolution can fully decompose the main direction of each channel of the input image, so the obtained pattern diagram has richer features compared with the original PCANet, and the robustness of the PCANet can be effectively improved.
The beneficial effects of the invention are mainly shown in the following steps: the method can more effectively process the changes such as shielding, illumination change, resolution difference and the like in the image to be identified, thereby effectively improving the identification rate of the offset image.
Drawings
Fig. 1 is a feature map extraction process of the high-dimensional PCANet according to the present invention, wherein,
Figure SMS_79
a flat convolution operation is shown, and the details of the step 7 are shown in the invention content; the U represents merging the feature map subsets; />
Figure SMS_80
The block histogram feature extraction of the pattern diagram is shown, and the step 18 of the invention is detailed;
FIG. 2 is a classification process of the high-dimensional PCANet of the present invention, see step 21 and step 22 of the summary, wherein NN represents the nearest neighbor classifier, id represents the final class of the image to be identified;
FIG. 3 is a training set sample and a test set sample from an AR face database, where (a) a sample of test set I, (b) a sample of test set II, (c) a sample of test set III, and (d) a sample of training set;
FIG. 4 (a) is a process of stretching a matrix into column vectors by Vec (·) operator, and FIG. 4 (b) is mat m×n A process in which the (-) operator resets the column vector to a matrix;
FIG. 5 is a process diagram of extracting feature blocks from a feature map in a flat convolution, where (a) is the original feature map, (b) is boundary zero padding, (c) is feature block selection, and (d) is the selected multi-channel feature block;
FIG. 6 is a process diagram of extracting feature blocks from a feature map in a stereo convolution, where (a) is the original feature map, (b) is boundary zero padding, (c) is feature block selection, and (d) is the selected multi-channel feature block;
fig. 7 (a) is a one-dimensional illustration of a flat filter, and fig. 7 (b) is a one-dimensional illustration of a stereo filter;
FIG. 8 is a two-dimensional illustration of a flat filter/stereo filter, wherein (a) represents the flat convolution kernel of convolution layer 1, (b) represents the flat convolution kernel of convolution layer 2, and (c) represents the stereo convolution kernel of convolution layer 2;
FIG. 9 is a model of a feature map generated from an image to be identified through a 2-layer flat convolution and a stereo convolution, where (a) represents the image to be identified (with illumination changes and occlusions), (b) represents the model of 64 feature maps generated through a 2-layer flat convolution, and (c) represents the model of 64 feature maps generated through a 2-layer stereo convolution;
fig. 10 is 16 pattern diagrams generated by the high-dimensional PCANet method, wherein the 8 pattern diagrams of the first row are from the characteristic diagrams generated by the flat convolution, and the 8 pattern diagrams of the second row are from the characteristic diagrams generated by the stereo convolution.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Referring to fig. 1 to 10, a robust image recognition method based on High-dimensional PCANet (HPCANet), the method comprising the steps of:
step 1 selecting J images A= { A 1 ,…,A J As training set, the corresponding class label is
Figure SMS_81
Z＝{Z 1 ,…,Z K And the image to be identified is a set, namely a test set. Here, a->
Figure SMS_82
Respectively represent the C on the real number domain 0 An image with length and width of m x n of E {1,3} channels; FIG. 3 illustrates three sample subsets of training set samples and images to be identified from an AR face database;
step 2, initializing parameters and input data: order the
Figure SMS_83
Here, a->
Figure SMS_84
For indicating the stage at which the network is located,
Figure SMS_85
indicating that the network is in training phase->
Figure SMS_86
Indicating that the network is in a testing stage; let l=0, where l is used to indicate the number of layers of the input image or feature map in the network,/>
Figure SMS_87
Wherein (1)>
Figure SMS_88
N＝J；
Step 3 consists of
Figure SMS_89
Construction of matrix->
Figure SMS_90
Figure SMS_91
Wherein,,
Figure SMS_92
Figure SMS_93
is->
Figure SMS_94
Mean value of->
Figure SMS_95
Figure SMS_96
Representing from->
Figure SMS_97
B e {1,2, …, mn } feature blocks of size k×k extracted from the c-th lane, vec (·) represents the operation of stretching the matrix into column vectors. FIG. 4 (a) specifically depicts the process of Vec (-) stretching the matrix into column vectors, and FIG. 5 details the process of extracting feature blocks from feature maps in a flat convolution;
step 4 if
Figure SMS_98
If the network is in the test stage, jumping to the step 7, otherwise, executing the next step;
step 5 calculation
Figure SMS_99
Main direction->
Figure SMS_100
Wherein (1)>
Figure SMS_101
Is covariance matrix->
Figure SMS_102
The i' th eigenvector of (a), the corresponding eigenvalue is lambda i′ And->
Figure SMS_103
Step 6 from V (l) Acquisition of C l+1 Three-dimensional filter bank
Figure SMS_104
Fig. 7 (a) and 8 (b) show one-dimensional and two-dimensional representations of a flat filter, respectively;
step 7 the feature atlas X of the (1) th convolution layer is calculated as follows (l+1) :7.1 Will) be
Figure SMS_106
Projected to
Figure SMS_110
7.2 Will->
Figure SMS_112
The elements in (a) are reorganized into feature atlas +.>
Figure SMS_107
Wherein (1)>
Figure SMS_108
And is also provided with
Figure SMS_111
Figure SMS_114
Here, a->
Figure SMS_105
Representation->
Figure SMS_109
Column vectors of rows a to b of column c, a% b representing a remainder of b,/-a%>
Figure SMS_113
Representation rounding down the real number a, mat m×n (v) represents that an arbitrary column vector is +.>
Figure SMS_115
Rearranging into an m×n matrix;
step 8, let l=l+1, execute the above steps 3 to 7 until l=l, where L represents the maximum convolution layer number given in advance;
step 9, initializing parameters and input data: order the
Figure SMS_116
l＝0，/>
Figure SMS_117
Wherein Y is i (l) ＝A i ，N＝J；
Step 10 consists of
Figure SMS_119
Construction of matrix->
Figure SMS_121
Figure SMS_123
Wherein (1)>
Figure SMS_120
Figure SMS_122
Is->
Figure SMS_124
Is used for the average value of (a),
Figure SMS_125
Figure SMS_118
representing the slave Y i (l) B e {1,2, …, mn } feature blocks of size k x k extracted from the c-th channel; FIG. 6 illustrates in detail the process of extracting feature blocks from a feature map in a stereo convolution;
step 11 if
Figure SMS_126
Then jump to step 14, otherwise, execute the next step;
step 12 calculation
Figure SMS_127
Main direction->
Figure SMS_128
Wherein (1)>
Figure SMS_129
As covariance momentMatrix->
Figure SMS_130
The i "th eigenvector of (a), the corresponding eigenvalue is lambda i″ And->
Figure SMS_131
Step 13 consists of
Figure SMS_132
Acquisition of C l+1 Three-dimensional filter bank
Figure SMS_133
Fig. 7 (b) and 8 (c) show one-dimensional and two-dimensional representations of a stereo filter, respectively;
step 14 the feature atlas Y of the 1+1th convolution layer is calculated as follows (l+1) :14.1 Will) be
Figure SMS_136
Projected to
Figure SMS_138
14.2 Will->
Figure SMS_139
Reorganizing elements in a feature atlas
Figure SMS_135
Wherein (1)>
Figure SMS_137
Figure SMS_140
And is also provided with
Figure SMS_141
Here, a->
Figure SMS_134
Representing connecting the matrices in the set in the channel direction;
step 15, let l=l+1, execute the above steps 3 to 14 until l=l;
step 16 sets the feature atlas X (L) And Y (L) Combining to form a new feature atlas F:
Figure SMS_142
fig. 9 shows the model values of 64×2=128 feature maps generated by 2-layer flat convolution and stereo convolution of the image to be identified (fig. 9 (a));
step 17, performing pattern diagram coding on the feature atlas F to obtain a pattern atlas P: p= { P i,β } i＝1,...,N；β＝1,…,B Wherein, the method comprises the steps of, wherein,
Figure SMS_143
beta e {1, …, B } pattern diagram representing the ith sample, F i,· Representing feature map subset F i In>
Figure SMS_144
T represents the number of channels involved in the encoding of a single pattern diagram, USF (·) represents a unit step function (Unit Step Function, USF), and the input value is binarized by comparison with 0, i.e.:
Figure SMS_145
FIG. 10 illustrates a pattern diagram generated by the high-dimensional PCANet method;
step 18 extracts histogram features H from the pattern atlas P: h= [ H ] i ] i＝1,…,N Wherein H is i ＝[H i,1 ,…,H i,B ] T ，H i,β ＝Qhist(P i,β )，Qhist(P i,β ) Representing the pattern diagram P i,β Divided into Q blocks, a histogram is extracted from each block, each histogram using 2 T The number of packets, i.e. the code value of the statistical pattern diagram is 2 for each feature block T The frequency of occurrence in the individual packets;
step 19 if
Figure SMS_146
Make H Te =h, jump to step 21; otherwise, let H Tr =h, perform the next step;
step 20 order
Figure SMS_147
l＝0，/>
Figure SMS_148
Wherein n=k,>
Figure SMS_149
executing the steps 3 to 19;
step 21 calculates a metric matrix m= [ M ] i,j ] i＝1,…,J；j＝1,…,K Wherein, the method comprises the steps of, wherein,
Figure SMS_150
here the number of the elements is the number,
Figure SMS_151
wherein D represents
Figure SMS_152
And->
Figure SMS_153
Length of->
Figure SMS_154
Representation->
Figure SMS_155
The d element of (a)>
Figure SMS_156
Representation of
Figure SMS_157
The d element of (a);
step 22 calculates class id= [ Id ] of each sample in the test set Z i ] i＝1,…,K ：
Figure SMS_158
Wherein M is i Represents the ith column vector in the metric matrix M, minIndx (·) represents M i Index of the smallest element in the (c).
Table 1 compares the recognition rates of three versions of HPCANet (HPCANet-1, HPCANet-2, HPCANet-3) with the existing method (VGG-Face, LCNN, PCANet) for the training set and test set given in FIG. 3. Here, two-layer convolution is adopted by all three versions of HPCANet, the number of flat convolution kernels adopted is 8 (convolution layer 1) +8 (convolution layer 2), the number of three-dimensional convolution kernels adopted by HPCANet-1 is 8 and 24 respectively, the number of three-dimensional convolution kernels adopted by HPCANet-2 is 8 and 32 respectively, and the number of three-dimensional convolution kernels adopted by HPCANet-3 is 8 and 40 respectively.
Figure SMS_159
TABLE 1
As can be seen from Table 1, HPCANet-1 through HPCANet-3 all exhibit better performance than PCANet, especially when the resolution of the image to be identified is low, this advantage is more pronounced; in addition, it can be seen that from HPCANet-1 to HPCANet-3, the recognition performance of HPCANet gradually increases with the increase of feature dimensions.

Claims (3)

1. A robust image recognition method based on high-dimensional PCANet, the method comprising the steps of:
step 1 selecting J images A= { A 1 ,…,A J As training set, the corresponding class label is
Figure QLYQS_1
Z＝{Z 1 ,…,Z K The number is the set of images to be identified, i.e. the test set, here +.>
Figure QLYQS_2
Respectively represent the C on the real number domain 0 ∈{An image of 1,3 channels having a length-width of mxn;
step 2, initializing parameters and input data: order the
Figure QLYQS_3
Here, a->
Figure QLYQS_4
For indicating the stage in which the network is located, +.>
Figure QLYQS_5
Indicating that the network is in training phase->
Figure QLYQS_6
Indicating that the network is in a testing stage; let l=0, where l is used to indicate the number of layers of the input image or feature map in the network,/>
Figure QLYQS_7
Wherein (1)>
Figure QLYQS_8
Step 3 consists of
Figure QLYQS_9
Construction of matrix->
Figure QLYQS_10
Figure QLYQS_11
Wherein,,
Figure QLYQS_12
Figure QLYQS_13
is->
Figure QLYQS_14
Mean value of->
Figure QLYQS_15
Figure QLYQS_16
Representing from->
Figure QLYQS_17
B e {1,2, …, mn } feature blocks of size k×k extracted from the c-th channel, vec (·) represents the operation of stretching the matrix into column vectors;
step 4 if
Figure QLYQS_18
If the network is in the test stage, jumping to the step 7, otherwise, executing the next step;
step 5 calculation
Figure QLYQS_19
Main direction->
Figure QLYQS_20
Wherein (1)>
Figure QLYQS_21
Is covariance matrix->
Figure QLYQS_22
The i' th eigenvector of (a), the corresponding eigenvalue is lambda i′ And->
Figure QLYQS_23
Step 6 from V (l) Acquisition of C l+1 Three-dimensional filter bank
Figure QLYQS_24
C l+1 ≤k 2 ；
Step 7 computing the feature atlas X of the 1+1th convolution layer (l+1) ；
Step 8, let l=l+1, execute the above steps 3 to 7 until l=l, where L represents the maximum convolution layer number given in advance;
step 9, initializing parameters and input data: order the
Figure QLYQS_25
l＝0，/>
Figure QLYQS_26
Wherein Y is i (l) ＝A i ，N＝J；
Step 10 consists of
Figure QLYQS_29
Construction of matrix->
Figure QLYQS_30
Figure QLYQS_32
Wherein (1)>
Figure QLYQS_28
Figure QLYQS_31
Is->
Figure QLYQS_33
Is used for the average value of (a),
Figure QLYQS_34
Figure QLYQS_27
representing the slave Y i (l) B e {1,2, …, mn } feature blocks of size k x k extracted from the c-th channel;
step 11 if
Figure QLYQS_35
Jump to step 14, otherwise, execute the next step；
Step 12 calculation
Figure QLYQS_36
Main direction->
Figure QLYQS_37
Wherein (1)>
Figure QLYQS_38
Is covariance matrix->
Figure QLYQS_39
The i "th eigenvector of (a), the corresponding eigenvalue is lambda i ", and->
Figure QLYQS_40
Step 13 consists of
Figure QLYQS_41
Acquisition of C l+1 Three-dimensional filter bank>
Figure QLYQS_42
C l+1 ≤k 2 C l ；
Step 14 calculates a feature atlas Y for the 1+1th convolution layer (l+1) ；
Step 15, let l=l+1, execute the above steps 10 to 14 until l=l;
step 16 sets the feature atlas X (L) And Y (L) Combining to form a new feature atlas F:
Figure QLYQS_43
step 17, performing pattern diagram coding on the feature atlas F to obtain a pattern atlas P: p= { P i,β } i＝1,…,N；β＝1,…,B Wherein, the method comprises the steps of, wherein,
Figure QLYQS_44
beta e 1 representing the ith sample…, B pattern diagrams, F i,· Representing feature map subset F i In>T represents the number of channels involved in the encoding of a single pattern, USF (·) represents a unit step function, and the input value is binarized by comparison with 0, i.e.:
Figure QLYQS_46
step 18 extracts histogram features H from the pattern atlas P: h= [ H ] i ] i＝1,…,N Wherein H is i ＝[H i,1 ,…,H i,B ] T ，H i,β ＝Qhist(P i,β )，Qhist(P i,β ) Representing the pattern diagram P i,β Divided into Q blocks, a histogram is extracted from each block, each histogram using 2 T The number of packets, i.e. the code value of the statistical pattern diagram is 2 for each feature block T The frequency of occurrence in the individual packets;
step 19 if
Figure QLYQS_47
Make H Te =h, jump to step 21; otherwise, let H Tr =h, perform the next step;
step 20 order
Figure QLYQS_48
l＝0，/>
Figure QLYQS_49
Wherein n=k,>
Figure QLYQS_50
executing the steps 3 to 19;
step 21 calculates a metric matrix m= [ M ] i,j ] i＝1,…,J；j＝1,…,K Wherein, the method comprises the steps of, wherein,
Figure QLYQS_51
here the number of the elements is the number,
Figure QLYQS_52
wherein D represents
Figure QLYQS_53
And->
Figure QLYQS_54
Length of->
Figure QLYQS_55
Representation->
Figure QLYQS_56
The d element of (a)>
Figure QLYQS_57
Representation->
Figure QLYQS_58
The d element of (a);
step 22 calculates class id= [ Id ] of each sample in the test set Z i ] i＝1,…,K ：
Figure QLYQS_59
Wherein M is i Represents the ith column vector in the metric matrix M, minIndx (·) represents M i Index of the smallest element in the (c).
2. The robust image recognition method based on high-dimensional PCANet as recited in claim 1, wherein in said step 7, a feature atlas X of the 1 st+1 th convolution layer is calculated as follows (l+1) :7.1 Will) be
Figure QLYQS_62
Projected to W (l+1 )：
Figure QLYQS_64
7.2 Will->
Figure QLYQS_67
Reorganizing elements in a feature atlas X (l+1) ：
Figure QLYQS_61
Wherein (1)>
Figure QLYQS_65
And is also provided with
Figure QLYQS_68
Figure QLYQS_70
c＝j％C l+1 The method comprises the steps of carrying out a first treatment on the surface of the Here, a->
Figure QLYQS_60
Representation->
Figure QLYQS_63
Column vectors of rows a to b of column c, a% b representing a remainder of b,/-a%>
Figure QLYQS_66
Representation rounding down the real number a, mat m×n (v) represents that an arbitrary column vector is +.>
Figure QLYQS_69
Rearranged into an mxn matrix.
3. The robust image recognition method based on high-dimensional PCANet according to claim 1 or 2, wherein in the step 14, the feature atlas Y of the 1+1th convolution layer is calculated as follows (l+1) :14.1 Will) be
Figure QLYQS_71
Projected to W (l+1) ：
Figure QLYQS_72
14.2 Will->
Figure QLYQS_73
Reorganizing elements in a feature atlas Y (l+1) ：Y (l +1) ＝{Y i (l+1) } i＝1,…,N Wherein->
Figure QLYQS_74
And is also provided with
Figure QLYQS_75
Here, a->
Figure QLYQS_76
The representation connects the matrices in the set in the channel direction.
CN202010147000.XA 2020-03-05 2020-03-05 Robust image recognition method based on high-dimensional PCANet Active CN111488905B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010147000.XA CN111488905B (en) 2020-03-05 2020-03-05 Robust image recognition method based on high-dimensional PCANet

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010147000.XA CN111488905B (en) 2020-03-05 2020-03-05 Robust image recognition method based on high-dimensional PCANet

Publications (2)

Publication Number Publication Date
CN111488905A CN111488905A (en) 2020-08-04
CN111488905B true CN111488905B (en) 2023-07-14

Family

ID=71794288

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010147000.XA Active CN111488905B (en) 2020-03-05 2020-03-05 Robust image recognition method based on high-dimensional PCANet

Country Status (1)

Country Link
CN (1) CN111488905B (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104573729A (en) * 2015-01-23 2015-04-29 东南大学 Image classification method based on kernel principal component analysis network
CN106778554A (en) * 2016-12-01 2017-05-31 广西师范大学 Cervical cell image-recognizing method based on union feature PCANet

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10713563B2 (en) * 2017-11-27 2020-07-14 Technische Universiteit Eindhoven Object recognition using a convolutional neural network trained by principal component analysis and repeated spectral clustering
US10747989B2 (en) * 2018-08-21 2020-08-18 Software Ag Systems and/or methods for accelerating facial feature vector matching with supervised machine learning

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104573729A (en) * 2015-01-23 2015-04-29 东南大学 Image classification method based on kernel principal component analysis network
CN106778554A (en) * 2016-12-01 2017-05-31 广西师范大学 Cervical cell image-recognizing method based on union feature PCANet

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
li xiaoxin，et al.Adaptive Weberfaces for occlusion-robust face representation and recognition.IET Image Processiong.2017,第11卷(第11期),第964-975页. *
韩冰，等.改进的主成分分析网络极光图像分类方法.西安电子科技大学学报.2017,第44卷(第1期),第83-88页. *

Also Published As

Publication number Publication date
CN111488905A (en) 2020-08-04

Similar Documents

Publication Publication Date Title
CN112633382B (en) Method and system for classifying few sample images based on mutual neighbor
CN110738207A (en) character detection method for fusing character area edge information in character image
CN110443128B (en) Finger vein identification method based on SURF feature point accurate matching
CN111027377B (en) Double-flow neural network time sequence action positioning method
CN111652273B (en) Deep learning-based RGB-D image classification method
CN111339924B (en) Polarized SAR image classification method based on superpixel and full convolution network
CN104077742B (en) Human face sketch synthetic method and system based on Gabor characteristic
CN113674334A (en) Texture recognition method based on depth self-attention network and local feature coding
CN112966574A (en) Human body three-dimensional key point prediction method and device and electronic equipment
CN110188864B (en) Small sample learning method based on distribution representation and distribution measurement
CN114387454A (en) Self-supervision pre-training method based on region screening module and multi-level comparison
CN111127407B (en) Fourier transform-based style migration forged image detection device and method
CN110516640B (en) Vehicle re-identification method based on feature pyramid joint representation
CN111488905B (en) Robust image recognition method based on high-dimensional PCANet
CN111488907B (en) Robust image recognition method based on dense PCANet
CN109583584B (en) Method and system for enabling CNN with full connection layer to accept indefinite shape input
CN111539966A (en) Colorimetric sensor array image segmentation method based on fuzzy c-means clustering
CN115937540A (en) Image Matching Method Based on Transformer Encoder
CN112818779B (en) Human behavior recognition method based on feature optimization and multiple feature fusion
CN111488906B (en) Low-resolution image recognition method based on channel correlation PCANet
CN113192076B (en) MRI brain tumor image segmentation method combining classification prediction and multi-scale feature extraction
CN111382703B (en) Finger vein recognition method based on secondary screening and score fusion
CN110599518B (en) Target tracking method based on visual saliency and super-pixel segmentation and condition number blocking
JP3454335B2 (en) Online handwritten character recognition method and apparatus
CN112419373A (en) Large-displacement optical flow field estimation 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