CN111666967B - Image classification method based on incoherence combined dictionary learning - Google Patents
Image classification method based on incoherence combined dictionary learning Download PDFInfo
- Publication number
- CN111666967B CN111666967B CN202010316589.1A CN202010316589A CN111666967B CN 111666967 B CN111666967 B CN 111666967B CN 202010316589 A CN202010316589 A CN 202010316589A CN 111666967 B CN111666967 B CN 111666967B
- Authority
- CN
- China
- Prior art keywords
- dictionary
- class
- shared
- training
- updating
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 239000011159 matrix material Substances 0.000 claims description 21
- 238000011478 gradient descent method Methods 0.000 claims description 7
- 230000001427 coherent effect Effects 0.000 claims description 3
- 230000008602 contraction Effects 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 2
- 238000002474 experimental method Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 101150029129 AR gene Proteins 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
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
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/55—Clustering; Classification
-
- 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
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
-
- 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 Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
An image classification method based on incoherence combined dictionary learning is characterized in that a class dictionary is trained for each class of images, a shared dictionary is trained for all images, low rank property of the shared dictionary is guaranteed to prevent the shared dictionary from absorbing class characteristics, and coherence constraint items are added between the low rank shared dictionary and the class dictionary to prevent the shared characteristics from appearing in the class dictionary. The method increases the discriminant of the training dictionary, improves the sparse representation capability of the dictionary, and further improves the accuracy of image classification.
Description
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to an image classification method based on incoherence combined dictionary learning.
Background
In recent years, sparse representation has achieved great success in the field of image processing, such as image classification, image denoising, compressed sensing, etc., which represents a signal as a linear combination of a few atoms in a redundant dictionary. In the sparse representation process, the training dictionary largely determines the quality of the sparse representation capability.
Currently, researchers have proposed various dictionary training methods to improve sparse representation capabilities. The simplest dictionary training directly uses all samples as a dictionary, such as Sparse Representation Classification (SRC), but when the training samples are too large, the complexity of the algorithm is too high and there is great redundancy in training the dictionary. K-SVD is a classical dictionary training algorithm that does not work well in image classification due to the fact that the class characteristics of the training samples are not introduced during the dictionary training process. As an improvement, class-oriented dictionary training algorithms D-KSVD, DLSI and FDDL were proposed. For example: the D-KSVD introduces a linear classifier to make the coding coefficients of the same sample more similar, and the coding coefficient difference of different samples is increased; DLSI introduces coherence constraint among class dictionaries to improve the discrimination capability of the dictionaries; FDDL adds a fisher discriminant term to increase the discriminant ability of the coding coefficients. Furthermore, experimental studies have found that although different classes of dictionaries have their own unique features, they share some features, so researchers have proposed class dictionary and shared dictionary joint training methods LRSDL, COPAR. The LRSDL prevents the shared dictionary from absorbing the characteristics of the class dictionary by ensuring the low rank property of the shared dictionary, so that the discriminant of the dictionary is improved. However, during the joint dictionary training process, shared features may appear in the class dictionary, which may reduce the accuracy of image classification.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides the image classification method based on the incoherence combined dictionary, which prevents shared features from appearing in the class dictionary so as to improve the discrimination capability of the dictionary, optimize the sparse representation capability of the image and further improve the accuracy of image classification.
The technical scheme of the invention is as follows:
an image classification method based on an incoherence joint dictionary is characterized in that a class dictionary is trained for each class of images, a shared dictionary is trained for all images, the low rank property of the shared dictionary is guaranteed to avoid the shared dictionary from absorbing the characteristics of the class dictionary, a coherent constraint item is added between the low rank shared dictionary and the class dictionary to prevent the shared characteristics from appearing in the class dictionary, and the model is as follows:
wherein the training sampleComprises class C->Represents a class c training sample,/-> Is a training dictionary,/->Representing a class c training dictionary,/->Representing a shared dictionary->Is a class dictionary that is used to store the data,is a matrix of coding coefficients>Is Y c Coding coefficients of the corresponding dictionary D, < >>Is Y c Corresponding dictionary D c Coding coefficient of>Is X c Remove->Given a matrix a and a natural number n, let μ (a, n) be a matrix of n identical columns, each column being the average vector of all columns in a, let>
Further, updating variables in the dictionary model with alternating iterative solutions includes: random gradient descent method for updating class dictionaryIterative updating of shared dictionary D by alternate direction multiplier method C+1 The fast soft threshold iterative algorithm updates the coding coefficient matrix X as follows:
taking into account the advantages of low complexity and fast running time of SGD by random gradient descent method, SGD update is used
Here alpha 1 Is the step size of the gradient descent;
2) Updating shared dictionary D C+1 :
Let x= [ X ] 1 ;...;X c ;...;X C ;X C+1 ], Andcoding coefficients of training samples under class c dictionary and shared dictionary, respectively, when +.>X fixed time, update shared dictionary D C+1 :
I=S(X C+1 ) T ;J=X C+1 (X C+1 ) T
Updating D using alternate direction multiplier method C+1 Iteratively updating formulas (6) - (8) until convergence, and solving a nuclear norm minimization problem by a singular value threshold algorithm;
U=U+D C+1 -V (8)
θ is a singular value contraction operator, E represents an identity matrix, α 2 Representing the step size of the gradient descent. Updating D in equation (6) by adopting random gradient descent method C+1 ;
3) Updating coding coefficient matrix X
When dictionary D is fixed, a fast soft threshold iterative algorithm is used to update X, assumingIn each iterative update, the derivative of s (X) is required;
The beneficial effects of the invention are as follows: a new dictionary learning method is provided for image classification, the shared features are prevented from appearing in the class dictionary by the coherence constraint items between the low-rank shared dictionary and the class dictionary, the discriminant of the training dictionary is improved, and the sparse representation capability of the image is optimized, so that the accuracy of the image classification is increased.
Drawings
FIG. 1 is a dictionary sparse representation model in accordance with the present invention.
Fig. 2 is a flow chart of the classification of experimental images according to the present invention.
FIG. 3 is a graph of experimental data set described in the present invention, wherein (a) represents an AR generator data set and (b) is a COIL-100 image of the data set.
FIG. 4 is a graph showing the effect of shared dictionary size on the overall classification accuracy of AR geneder data in the present invention.
FIG. 5 is a graph showing the effect of shared dictionary size on overall classification accuracy of COIL-100 data in the present invention.
Detailed Description
The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto.
Referring to fig. 1 to 5, in an image classification method based on incoherence combined dictionary learning, since sharing features are included between different types of images, the invention trains one class dictionary for each type of image and trains one shared dictionary for all images. In addition, the low rank property of the shared dictionary is ensured to avoid the shared dictionary from absorbing the characteristics of the class dictionary, and a coherent constraint item is added between the low rank shared dictionary and the class dictionary to prevent the shared characteristics from appearing in the class dictionary.
Wherein the training sampleComprises class C->Represents a class c training sample,/-> Is a training dictionary,/->Representing a class c training dictionary,/->Representing a shared dictionary->Is a class dictionary that is used to store the data,is a matrix of coding coefficients>Is Y c Coding coefficients of the corresponding dictionary D, < >>Is Y c Corresponding dictionary D c Coding coefficient of>Is X c Remove->Given a matrix a and a natural number n, let μ (a, n) be a matrix of n identical columns, each column being the average vector of all columns in a, let>
In order to minimize dictionary objective functions, the class dictionaries are preferably updated separately using an alternating iterative solution methodShared dictionary D C+1 A sparse coding matrix X;
taking into account the advantages of low complexity and fast running time of SGD by random gradient descent method, SGD update is used
Here alpha 1 Is the step size of the gradient descent;
2) Updating shared dictionary D C+1 :
Let x= [ X ] 1 ;...;X c ;...;X C ;X C+1 ], Andcoding coefficients of training samples under class c dictionary and shared dictionary, respectively, when +.>X fixed time, update shared dictionary D C+1 :
I=S(X C+1 ) T ;J=X C+1 (X C+1 ) T
Updating D using alternate direction multiplier method C+1 Iteratively updating formulas (6) - (8) until convergence, and solving a nuclear norm minimization problem by a singular value threshold algorithm;
U=U+D C+1 -V (8)
θ is a singular value contraction operator, E represents an identity matrix, α 2 Step size representing gradient descent, updating D in formula (6) by adopting random gradient descent method C+1 ;
3) Updating coding coefficient matrix X
When dictionary D is fixed, a fast soft threshold iterative algorithm is used to update X, assumingIn each iterative update, the derivative of s (X) is required;
Classification was performed on two classes of image dataset, including the AR gener database, the COIL-100 database. The experimental procedure is divided into five steps: firstly, preprocessing all data sets, and reducing the interference of training samples on experiments; secondly, extracting the characteristics of each image in order to extract useful information of the image and reduce the dimension of the characteristics of the image; furthermore, the dictionary learning method is adopted to train the discriminant dictionary; then, sparsely coding the image on the trained dictionary; and finally, classifying the image according to the reconstruction error and the dictionary coding item to obtain the Overall Classification Accuracy (OCA).
AR gene experiment: AR gender face dataset images are divided into two categories, each containing 700 pictures. The dimension of each image after feature extraction is 300, the size of each class of dictionary in the invention is 300, the size of the shared dictionary is set to be 3, and the influence of the size of the shared dictionary on classification accuracy is shown in figure 4. The classification accuracy obtained by various dictionary learning methods is shown in the table one.
Dictionary learningMethod | Classification accuracy (%) |
SRC | 91 |
COPAR | 93.56 |
LRSDL | 92.86 |
DLSI | 93.86 |
FDDL | 91.71 |
Proposed Method | 94.56 |
TABLE 1
As shown in Table 1, the classification accuracy obtained by adopting the dictionary training method of the invention on the data set AR gener is 94.56%, which is improved by 0.7% compared with the suboptimal algorithm DLSI, and is improved by 3.56%, 1%, 1.7% and 2.85% compared with the SRC, COPAR, LRSDL, FDDL dictionary learning method.
COIL-100 experiment: the COIL-100 image dataset contained 100 classes, each containing 72 pictures, with each image having dimensions 324 after feature extraction. In the invention, the size of each class of dictionary is 45, the size of the shared dictionary is 3, and the influence of the size of the shared dictionary on the whole classification precision is shown in figure 5.
The overall classification accuracy obtained by various dictionary learning methods is shown in a second table.
Dictionary learning method | Overall classification accuracy (%) |
SRC | 89.61 |
COPAR | 90.29 |
LRSDL | 91.76 |
DLSI | 92.94 |
FDDL | 88.82 |
Proposed Method | 93.53 |
TABLE 2
As can be seen from Table 2, the overall classification accuracy obtained by the dictionary training method of the invention on the COIL-100 data set is 93.53%, which is improved by 0.59% compared with the suboptimal algorithm DLSI, and is respectively improved by 3.92%, 3.24%, 1.77% and 4.71% compared with the SRC, COPAR, LRSDL, FDDL dictionary learning method.
The foregoing is considered as illustrative of the principles of the present invention, and has been described herein before with reference to the accompanying drawings, in which the invention is not limited to the specific embodiments shown.
What is not described in detail in this specification is prior art known to those skilled in the art.
Claims (1)
1. An image classification method based on incoherence joint dictionary learning is characterized by comprising the following steps:
firstly, preprocessing all data sets;
secondly, extracting the characteristics of each image;
furthermore, training a discriminant dictionary by adopting a dictionary learning method;
then, sparsely coding the image on the trained dictionary;
finally, classifying the images according to the reconstruction errors and dictionary coding items;
the process of training the discriminant dictionary by adopting the dictionary learning method is as follows:
training a class dictionary for each class of images, training a shared dictionary for all images, ensuring the low rank property of the shared dictionary, and adding a coherent constraint term between the low rank shared dictionary and the class dictionary, wherein the model is as follows:
wherein the training sampleComprises class C->Represents a class c training sample,/-> Is a training dictionary,/->Representing a class c training dictionary,/->Representing a shared dictionary->Is a class dictionary that is used to store the data,is a matrix of coding coefficients>Is Y c Coding coefficients of the corresponding dictionary D, < >>Is Y c Corresponding dictionary D c Coding coefficient of>Is X c Remove->Given a matrix a and a natural number n, let μ (a, n) be a matrix of n identical columns, each column being the average vector of all columns in a, let
Alternating iterative solutions to update variables in the dictionary model includes: the dictionary is updated by adopting a random gradient descent method, the shared dictionary is iteratively updated by adopting an alternate direction multiplier method, and the encoding coefficient matrix is updated by adopting a rapid soft threshold iterative algorithm, wherein the method comprises the following steps:
Here alpha 1 Is the step size of the gradient descent;
2) Updating shared dictionary D C+1 :
Let x= [ X ] 1 ;...;X c ;...;X C ;X C+1 ], Andcoding coefficients of training samples under class c dictionary and shared dictionary, respectively, when +.>X fixed time, update shared dictionary D C+1 :
Updating D using alternate direction multiplier method C+1 Iteratively updating formulas (6) - (8) until convergence, and solving a nuclear norm minimization problem by a singular value threshold algorithm;
V=θ(D C+1 +U) (7)
U=U+D C+1 -V (8)
θ is a singular value contraction operator, E represents an identity matrix, α 2 Step size representing gradient descent, updating D in formula (6) by adopting random gradient descent method C+1 ;
3) Updating coding coefficient matrix X
When dictionary D is fixed, a fast soft threshold iterative algorithm is used to update X, assumingIn each iterative update, the derivative of s (X) is required,
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010316589.1A CN111666967B (en) | 2020-04-21 | 2020-04-21 | Image classification method based on incoherence combined dictionary learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010316589.1A CN111666967B (en) | 2020-04-21 | 2020-04-21 | Image classification method based on incoherence combined dictionary learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111666967A CN111666967A (en) | 2020-09-15 |
CN111666967B true CN111666967B (en) | 2023-06-13 |
Family
ID=72382643
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010316589.1A Active CN111666967B (en) | 2020-04-21 | 2020-04-21 | Image classification method based on incoherence combined dictionary learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111666967B (en) |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106778807A (en) * | 2016-11-22 | 2017-05-31 | 天津大学 | The fine granularity image classification method of dictionary pair is relied on based on public dictionary pair and class |
CN106815876A (en) * | 2016-12-30 | 2017-06-09 | 清华大学 | Image sparse characterizes the combined optimization training method of many dictionary learnings |
US9865036B1 (en) * | 2015-02-05 | 2018-01-09 | Pixelworks, Inc. | Image super resolution via spare representation of multi-class sequential and joint dictionaries |
CN108573263A (en) * | 2018-05-10 | 2018-09-25 | 西安理工大学 | A kind of dictionary learning method of co-ordinative construction rarefaction representation and low-dimensional insertion |
CN108985177A (en) * | 2018-06-21 | 2018-12-11 | 南京师范大学 | A kind of facial image classification method of the quick low-rank dictionary learning of combination sparse constraint |
CN109409201A (en) * | 2018-09-05 | 2019-03-01 | 昆明理工大学 | A kind of pedestrian's recognition methods again based on shared and peculiar dictionary to combination learning |
CN110705343A (en) * | 2019-08-20 | 2020-01-17 | 西南科技大学 | Face recognition method and system for structure-incoherent projection dictionary pair learning |
-
2020
- 2020-04-21 CN CN202010316589.1A patent/CN111666967B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9865036B1 (en) * | 2015-02-05 | 2018-01-09 | Pixelworks, Inc. | Image super resolution via spare representation of multi-class sequential and joint dictionaries |
CN106778807A (en) * | 2016-11-22 | 2017-05-31 | 天津大学 | The fine granularity image classification method of dictionary pair is relied on based on public dictionary pair and class |
CN106815876A (en) * | 2016-12-30 | 2017-06-09 | 清华大学 | Image sparse characterizes the combined optimization training method of many dictionary learnings |
CN108573263A (en) * | 2018-05-10 | 2018-09-25 | 西安理工大学 | A kind of dictionary learning method of co-ordinative construction rarefaction representation and low-dimensional insertion |
CN108985177A (en) * | 2018-06-21 | 2018-12-11 | 南京师范大学 | A kind of facial image classification method of the quick low-rank dictionary learning of combination sparse constraint |
CN109409201A (en) * | 2018-09-05 | 2019-03-01 | 昆明理工大学 | A kind of pedestrian's recognition methods again based on shared and peculiar dictionary to combination learning |
CN110705343A (en) * | 2019-08-20 | 2020-01-17 | 西南科技大学 | Face recognition method and system for structure-incoherent projection dictionary pair learning |
Non-Patent Citations (2)
Title |
---|
Discriminative Dictionary Learning with Low-Rank Regularization for Face Recognition;Liangyue Li,Sheng Li,Yun Fu;《FG 2013》;全文 * |
Fast Low-Rank Shared Dictionary Learning for Image Classification;Tiep Huu Vu,and Vishal Monga;《TIP 2017》;正文第1-3节 * |
Also Published As
Publication number | Publication date |
---|---|
CN111666967A (en) | 2020-09-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CA3085033A1 (en) | Methods and systems for multi-label classification of text data | |
EP3435247A1 (en) | Electronic device and method for text processing | |
CN113806527A (en) | Cross-language unsupervised classification with multi-view migration learning | |
Christlein et al. | ICDAR 2019 competition on image retrieval for historical handwritten documents | |
CN114791958B (en) | Zero sample cross-modal retrieval method based on variational self-encoder | |
CN113793319A (en) | Fabric image flaw detection method and system based on class constraint dictionary learning model | |
CN109636722A (en) | A method of the online dictionary learning super-resolution rebuilding based on rarefaction representation | |
CN106803105B (en) | Image classification method based on sparse representation dictionary learning | |
CN111666967B (en) | Image classification method based on incoherence combined dictionary learning | |
CN114841176A (en) | Neural machine translation robustness enhancing method, device and storage medium | |
CN108388918B (en) | Data feature selection method with structure retention characteristics | |
CN110378356B (en) | Fine-grained image identification method based on multi-target Lagrangian regularization | |
CN111310807B (en) | Feature subspace and affinity matrix joint learning method based on heterogeneous feature joint self-expression | |
CN109190645B (en) | High-order high-dimensional image data representation and classification method | |
CN111931665B (en) | Under-sampling face recognition method based on intra-class variation dictionary modeling | |
KR20190068205A (en) | System and Method for Automatic Agent Assignment of Bug Reports | |
CN110109994B (en) | Automobile financial wind control system containing structured and unstructured data | |
CN109299260B (en) | Data classification method, device and computer readable storage medium | |
Zhang et al. | Deep residual network based medical image reconstruction | |
CN112613282A (en) | Text generation method and device and storage medium | |
CN111797732A (en) | Video motion identification anti-attack method insensitive to sampling | |
Wang et al. | Seanet: A deep learning architecture for data series similarity search | |
CN107958229B (en) | Face recognition method, device and equipment based on neighbor keeping low-rank representation | |
US11763094B2 (en) | Cascade pooling for natural language processing | |
CN108182429B (en) | Method and device for extracting facial image features based on symmetry |
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 |