CN106529586A - Image classification method based on supplemented text characteristic - Google Patents

Image classification method based on supplemented text characteristic Download PDF

Info

Publication number
CN106529586A
CN106529586A CN201610939529.9A CN201610939529A CN106529586A CN 106529586 A CN106529586 A CN 106529586A CN 201610939529 A CN201610939529 A CN 201610939529A CN 106529586 A CN106529586 A CN 106529586A
Authority
CN
China
Prior art keywords
image
feature
characteristic
dictionary
vector
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201610939529.9A
Other languages
Chinese (zh)
Inventor
郭继昌
王楠
张帆
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.)
Tianjin University
Original Assignee
Tianjin 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 Tianjin University filed Critical Tianjin University
Priority to CN201610939529.9A priority Critical patent/CN106529586A/en
Publication of CN106529586A publication Critical patent/CN106529586A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/245—Classification techniques relating to the decision surface
    • G06F18/2451—Classification techniques relating to the decision surface linear, e.g. hyperplane
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/467—Encoded features or binary features, e.g. local binary patterns [LBP]

Abstract

The invention relates to the field of digital image processing technology, and provides a method for image classification through supplementing a text characteristic for settling a problem of classification result deterioration caused by image characteristic discriminating property reduction because of text characteristic information loss in a bag of visual words. The image classification method based on the supplemented text characteristic comprises the following steps of (1), respectively extracting an SIFT characteristic, a color characteristic, and a shape-and-boundary characteristic of an image in a data set; (2), randomly selecting 200000 characteristics from the supplemented text characteristics and obtaining a dictionary B through a K-singular value decomposition method (K-SVD); (3), obtaining a sparse characteristic vector C through the studied dictionary B; (4), performing spatial maximal value gathering on the sparse characteristic; and (5), transmitting an image characteristic vector F into a linear SVM classifier, thereby performing training and testing respectively for obtaining an image classification result. The image classification method is mainly applied for digital image processing.

Description

Image classification method based on supplemental text feature
Technical field
The present invention relates to digital image processing techniques field, and in particular to a kind of scene image based on supplemental text feature Sorting technique.
Background technology
Basis of the image classification as image understanding, plays important role in computer vision field.Although people Many effort are done in terms of feature extraction and training, has been had made great progress, but at present still without unified best The sorting technique suitable for all images.
In image classification, it is exactly vision bag of words (Bag-of-Word) that one of them is the most popular, and the model is one The image classification model based on statistical model is planted, its objective is to obtain the image feature representation of more expressive force.Should in the model It is the feature based on image text information with most commonly used characteristics of image, such as SIFT, HOG etc..After extracting the feature of image, Need the local feature that profit is obtained to build visual dictionary, the feature extracted is carried out mapping with the dictionary of construction then and obtains figure The character representation of picture, is used for image classification and retrieval as the output characteristic of bag of words.
In traditional vision bag of words, the description subcharacter of extraction is usually the text features such as SIFT, and this feature is main The specific details of some of target described in collection, tends to ignore the information such as color in image, shape, border.And these Information is also an indispensable part in image classification, if not having image classification of these information based on bag of words can Can ignore that the nuance between image, so as to cause final characteristics of image to cash deficiency, affects the effect of image classification Really.
The content of the invention
To overcome the deficiencies in the prior art, it is contemplated that：As text feature information lacks in solution vision bag of words Lose and the final problem for affecting classification results of caused characteristics of image ability to see things in their true light decline, proposition is a kind of to supplement to text feature Method be used for image classification.The text features such as SIFT are mainly based upon a kind of local feature of image interest point extraction, the spy Levy the information such as color in easy missing image, shape, border.Therefore it is proposed that a kind of specific compensation process, the method While not being significantly increased amount of calculation, using the color of image, shape and boundary characteristic as the supplement of text feature and its melt Close, and the complementary features after fusion are used for into visual dictionary and build and feature coding, so as to improve the ability to see things in their true light of characteristics of image, carry High-class accuracy rate.The technical solution used in the present invention is that, based on the image classification method of supplemental text feature, step is as follows：
(1) SIFT feature of image, color characteristic, shape and boundary characteristic in data set is extracted respectively, and by color, shape Shape and boundary characteristic are dissolved in SIFT feature as side information, obtain supplementary text feature；
(2) randomly select 200000 from supplementary text feature to train by K- singular value decomposition method K-SVD To dictionary B；
(3) by the dictionary B for learning, all of text feature X of single image is entered using orthogonal matching pursuit method Row sparse coding, obtains sparse features vector C；
(4) space maximum convergence is carried out to sparse levying, and feature after convergence is normalized obtains image Characteristic vector F；
(5) image feature vector F is sent into into Linear SVM grader, is trained respectively and test obtains image classification knot Really.
Step 1 is specially：
(1.1) by each image with 8 pixels as interval, 16 × 16 image block is divided into, and obtains each in image block The gradient magnitude and gradient direction of pixel；
(1.2) vector sum of each pixel in 8 directions in the image block per 4 × 4 pixels is calculated, wherein 8 directions point Not Wei 0 degree, 45 degree, 90 degree, 135 degree, 180 degree, 225 degree, 270 degree, 315 degree, obtain 128 dimensions SIFT feature vector, 16 × 16 such characteristic vectors are had in 16 image block, this feature is used as the basic text feature of image；
(1.3) image is transformed into into hsv color space, and original image block is divided into 4 × 4 grid, in each grid The average and standard deviation of three Color Channels are inside calculated respectively, obtain the color characteristic of the feature of 96 dimensions, i.e. image, as text The side information of feature；
(1.4) dimensionality reduction, half of the dimensionality reduction to text feature dimension are carried out to color characteristic using PCA；
(1.5) the compact low-dimensional feature of image block is extracted using the matching kernel function based on local binary patterns
Wherein P represents two different image blocks with Q, and a and a ' represents the pixel in two image blocks respectively, S (a) is the standard deviation of pixel value in 3 × 3 fields centered on a, ∈sIt is the constant of a numerical value very little, B (a) is a binary column vector for pixel value difference in the local window around binaryzation a, standardized linear kernelFor weighing the contribution of each local binary patterns, gaussian kernel function kb(b (a), b (a '))=exp (- γb‖b (a)-b(a′)‖2) for measuring the shape similarity between two local binary patterns, Gauss position kernel function kp(a, a ')= exp(-γp‖a-a′‖2) for measuring the space length between two pixels；
(1.6) dimensionality reduction is carried out to low-dimensional feature using core principle component analysis method, dimension is equally chosen for 64 dimensions, obtaining can be with The feature of image local shape and boundary information is obtained effectively；
(1.7) merged color, shape and boundary information as supplementary information and SIFT feature, obtaining dimension is 256 supplemental text feature；
Step 2 is specially：
(2.1) 200000 features, each feature are randomly selected in all of supplemental text feature of image from data set It is 256 dimensions, 200000 image blocks may be constructed the matrix of 256 × 200000, each row of matrix represent a benefit The Text eigenvector for filling；
(2.2) 512 are randomly selected as initial dictionary from 200000 image blocks, then initial dictionary can be with one Individual 256 × 512 matrix is represented；
(2.3) with K-SVD method training dictionaries, that is, solve object function：
Wherein K is degree of rarefication, is set to 5, c hereiI-th row of presentation code C.The step of obtaining dictionary is as follows：
(2.3.1) dictionary is fixed, each y is solved by orthogonal matching pursuit methodiCorresponding rarefaction representation ci, its The span of middle i is 1 200000；
(2.3.2) residual error is calculated, and chooses the residual error item for indexing set that correspondence uses the kth of dictionary to arrange, to residual error item Singular value decomposition is carried out, singular value decomposition matrix is obtained, and is wherein obtained 256 × 256 unitary matrice；
(2.3.3) the kth row of dictionary are updated with 256 × 256 unitary matrice first row；
(2.3.4) iterate, training obtains final dictionary, and iterations is set to 50 here.
Step 3, specially：
(3.1) dictionary B is given, cataloged procedure is just to solve for the process of following formula：
(3.2) first time iteration, selects input data x as current residue, orthogonal matching pursuit algorithm select with it is current The optimal corresponding vision word of residual error；
(3.3) by x orthogonal mappings to the vision word for above selecting, and recalculate residual error；
(3.4) iterate, obtain sparse coding vector.
Step 4 is specially：
(4.1) to all of supplemental text feature in entire image, the maximum of its each dimension is sought, obtains one 512 dimension The characteristic vector of degree；
(4.2) entire image is divided into 4 image blocks, each characteristics of image is in each dimension in seeking each image block respectively Maximum, obtains the characteristic vector of 4 512 dimensions；
(4.3) entire image is divided into 16 image blocks, each characteristics of image is in each dimension in seeking each image block respectively Maximum, obtain the characteristic vector of 16 512 dimensions；
(4.4) the characteristics or every unit vector for producing (4.1) step to (4.3) is normalized respectively：
FsFor the convergence feature that each unit is obtained, S is unit number, ‖ Fs‖pFor p norms；
(4.5) normalization characteristic of obtain 21 units is attached, obtains the character representation of image.
Step 5 is specially：
(5.1) it is random that 60 images are chosen from each class image of data set as such training sample, residual graph As such test sample；
(5.2) characteristic vector of 100 images for choosing step (5.1) is trained in sending into Linear SVM grader, Obtain the SVM classifier for training；
(5.3) test sample chosen step (5.1) is sent in the SVM classifier that step (5.2) is trained, and obtains figure The classification results of picture.
The characteristics of of the invention and beneficial effect are：
1. context of methods is incorporated initial using the color in image, shape and boundary information as supplementary image information In text feature, the problem of text feature loss of learning is compensate for, improve the expressive force of characteristics of image.
2. the side information that context of methods is added all is Jing dimension-reduction treatment, therefore the addition of side information substantially will not increase Plus amount of calculation；
3. context of methods extracts effective information in image and by effective fusion method, is not increasing computing complexity The ability to see things in their true light of feature is increased while spending, image classification accuracy rate is improve.
Description of the drawings：
Fig. 1 is the flow chart of image classification method herein based on supplemental text feature.
Fig. 2 is this paper experimental data base parts of images.
Fig. 3 image space pyramid partition structure figures.
Specific embodiment
A kind of image classification method based on supplemental text feature is proposed, the shape in image, border and color are believed Breath merges in initial SIFT text features as side information, then carries out the structure and feature coding of dictionary, is had The graphical representation of effect, improves the accuracy rate of image classification.
The basic ideas of this method are：Basic text feature is entered using the shape in image, border and colouring information Row is supplemented, by by these information fusions solving the problems, such as text feature loss of learning in text feature, so as to follow-up The expressive force of feature is improved during feature learning, the accuracy rate of final image classification is improved.Specific method and step is as follows：
(6) SIFT feature of image, color characteristic, shape and boundary characteristic in data set is extracted respectively, and by color, shape Shape and boundary characteristic are dissolved in SIFT feature as side information, obtain supplementary text feature；
(7) randomly select 200000 from supplementary text feature to train by K- singular value decomposition methods (K-SVD) Obtain dictionary B；
(8) by the dictionary B for learning, all of text feature X of single image is entered using orthogonal matching pursuit method Row sparse coding, obtains sparse features vector C；
(9) space maximum convergence is carried out to sparse levying, and feature after convergence is normalized obtains image Characteristic vector F；
(10) image feature vector F is sent into into Linear SVM grader, is trained respectively and test obtains image classification knot Really.
Compared by experiment and shown, this method is compared existing similar image classification method and has more preferable classifying quality.
This experiment adopt Oxford Flower-17 data sets, data set include 17 classes flower image, each classification 80, Open image and there is larger dimensional variation, shooting angle change and illumination variation.Each class training image takes 60 respectively at random It is trained, used as test image, classification results take the mean value of 10 subseries results to residual image.With this method and its other party Method is compared on Oxford Flower-17 data sets, as a result as shown in table 1.
1 this method of table and additive method comparative result on Oxford Flower-17 data sets
Method Average classification accuracy
Gehler[1] 85.5%
LLC[2] 88.24%
Context of methods 89.81%
Can be seen that under same experiment parameter by result in table, this method is in Oxford Flower-17 data sets On classification accuracy be higher than the method in article [1] [2], and be higher by 4.3% and 1.6% respectively.With method ratio in [1] Compared with, the image classification for combining multiple features is, but the accuracy rate of this method is higher, illustrate this method selection supplements letter Breath increased the expressive force of text feature, and the characteristics of image obtained by characteristic study has more ability to see things in their true light；Compare with [2] method, This method increased the side information of text feature, but not be significantly increased amount of calculation, and classification accuracy is improved, Illustrate the validity and feasibility of this method.
[1]Gehler P,Nowozin S.On feature combination for multiclass object classification[C]//Computer Vision,2009IEEE 12th International Conference on.IEEE,2009:221-228。
[2]Wang J,Yang J,Yu K,et al.Locality-constrained linear coding for image classification[C]//Computer Vision and Pattern Recognition(CVPR), 2010IEEE Conference on.IEEE,2010:3360-3367。
In order that the scheme and advantage of this method become more apparent, with reference to accompanying drawing, this method is specifically described：
Step 1, extracts SIFT feature, color characteristic, shape and the boundary characteristic of image in data set respectively.
(1.1) by each image with 8 pixels as interval, 16 × 16 image block is divided into, and obtains each in image block The gradient magnitude and gradient direction of pixel；
(1.2) vector sum of each pixel in 8 directions in the image block per 4 × 4 pixels is calculated, wherein 8 directions point Not Wei 0 degree, 45 degree, 90 degree, 135 degree, 180 degree, 225 degree, 270 degree, 315 degree, obtain 128 dimensions SIFT feature vector, 16 × 16 such characteristic vectors are had in 16 image block, this feature is used as the basic text feature of image；
(1.3) image is transformed into into hsv color space, and original image block is divided into 4 × 4 grid, in each grid The average and standard deviation of three Color Channels are inside calculated respectively, obtain the color characteristic of the feature of 96 dimensions, i.e. image, as text The side information of feature；
(1.4) dimensionality reduction, half of the dimensionality reduction to text feature dimension are carried out to color characteristic using PCA；
(1.5) the compact low-dimensional feature of image block is extracted using the matching kernel function based on local binary patterns
Wherein P represents two different image blocks with Q, and a and a ' represents the pixel in two image blocks respectively, S (a) is the standard deviation of pixel value in 3 × 3 fields centered on a, ∈sIt is the constant of a numerical value very little, B (a) is a binary column vector for pixel value difference in the local window around binaryzation a, standardized linear kernelFor weighing the contribution of each local binary patterns, gaussian kernel function kb(b (a), b (a '))=exp (- γb‖b (a)-b(a′)‖2) for measuring the shape similarity between two local binary patterns, Gauss position kernel function kp(a, a ')= exp(-γp‖a-a′‖2) for measuring the space length between two pixels.
(1.6) dimensionality reduction is carried out to low-dimensional feature using core principle component analysis method, dimension is equally chosen for 64 dimensions, obtaining can be with The feature of image local shape and boundary information is obtained effectively；
(1.7) merged color, shape and boundary information as supplementary information and SIFT feature, obtaining dimension is 256 supplemental text feature.
Step 2, is randomly selected 200000 from supplementary text feature and is instructed by K- singular value decomposition methods (K-SVD) Get dictionary B.
(2.1) 200000 features, each feature are randomly selected in all of supplemental text feature of image from data set It is 256 dimensions, 200000 image blocks may be constructed the matrix of 256 × 200000, each row of matrix represent a benefit The Text eigenvector for filling；
(2.2) 512 are randomly selected as initial dictionary from 200000 image blocks, then initial dictionary can be with one Individual 256 × 512 matrix is represented；
(2.3) with K-SVD method training dictionaries, that is, solve object function：
Wherein K is degree of rarefication, is set to 5, c hereiI-th row of presentation code C.The step of obtaining dictionary is as follows：
(2.3.1) dictionary is fixed, each y is solved by orthogonal matching pursuit methodiCorresponding rarefaction representation ci, its The span of middle i is 1 200000；
(2.3.2) residual error is calculated, and chooses the residual error item for indexing set that correspondence uses the kth of dictionary to arrange, to residual error item Singular value decomposition is carried out, singular value decomposition matrix is obtained, and is wherein obtained 256 × 256 unitary matrice；
(2.3.3) the kth row of dictionary are updated with 256 × 256 unitary matrice first row；
(2.3.4) iterate, training obtains final dictionary, and iterations is set to 50 here.
Step 3, by the dictionary B for learning, using all of text feature X of the orthogonal matching pursuit method to single image Sparse coding is carried out, sparse features vector C is obtained.
(3.1) dictionary B is given, cataloged procedure is just to solve for the process of following formula：
(3.2) first time iteration, selects input data x as current residue, orthogonal matching pursuit algorithm select with it is current The optimal corresponding vision word of residual error；
(3.3) by x orthogonal mappings to the vision word for above selecting, and recalculate residual error；
(3.4) iterate, obtain sparse coding vector.
Step 4, carries out spatial pyramid division as shown in Figure 3, and sparse the levying to obtaining carries out space maximum to image Value is converged, and feature after convergence is normalized characteristic vector F for obtaining image.
(4.1) to all of supplemental text feature in entire image, the maximum of its each dimension is sought, obtains one 512 dimension The characteristic vector of degree；
(4.2) entire image is divided into 4 image blocks, each characteristics of image is in each dimension in seeking each image block respectively Maximum, obtains the characteristic vector of 4 512 dimensions；
(4.3) entire image is divided into 16 image blocks, each characteristics of image is in each dimension in seeking each image block respectively Maximum, obtain the characteristic vector of 16 512 dimensions；
(4.4) the characteristics or every unit vector for producing (4.1) step to (4.3) is normalized respectively：
FsFor the convergence feature that each unit is obtained, S is unit number, here S=21, ‖ Fs‖pFor p norms, p takes 2 here；
(4.5) normalization characteristic of obtain 21 units is attached, obtains the character representation of image.
Image feature vector F is sent into Linear SVM grader, is trained respectively and test obtains image classification by step 5 As a result.
(5.1) it is random that 60 images are chosen from each class image of data set as such training sample, residual graph As such test sample；
(5.2) characteristic vector of 100 images for choosing step (5.1) is trained in sending into Linear SVM grader, Obtain the SVM classifier for training；
(5.3) test sample chosen step (5.1) is sent in the SVM classifier that step (5.2) is trained, and obtains figure The classification results of picture.

Claims (6)

1. a kind of image classification method based on supplemental text feature, is characterized in that, step is as follows：
(1) extract the SIFT feature of image, color characteristic, shape and boundary characteristic in data set respectively, and by color, shape and Boundary characteristic is dissolved in SIFT feature as side information, obtains supplementary text feature；
(2) 200000 are randomly selected from supplementary text feature and word is obtained by K- singular value decomposition methods K-SVD training Allusion quotation B；
(3) by the dictionary B for learning, all of text feature X of single image is carried out using orthogonal matching pursuit method dilute Coding is dredged, sparse features vector C is obtained；
(4) space maximum convergence is carried out to sparse levying, and feature after convergence is normalized the feature for obtaining image Vectorial F；
(5) image feature vector F is sent into into Linear SVM grader, is trained respectively and test obtains image classification result.
2. the image classification method based on supplemental text feature as claimed in claim 1, is characterized in that, step 1 is specially：
(1.1) by each image with 8 pixels as interval, 16 × 16 image block is divided into, and obtains each pixel in image block The gradient magnitude and gradient direction of point；
(1.2) vector sum of each pixel in 8 directions in the image block per 4 × 4 pixels is calculated, wherein 8 directions are respectively 0 Degree, 45 degree, 90 degree, 135 degree, 180 degree, 270 degree, 315 degree, obtains the SIFT feature vector of 128 dimensions, 16 × 16 figure by 225 degree The characteristic vector as 16 are had in block, this feature is used as the basic text feature of image；
(1.3) image is transformed into into hsv color space, and original image block is divided into 4 × 4 grid, in each grid point Not Ji Suan three Color Channels average and standard deviation, obtain the color characteristic of the feature of 96 dimensions, i.e. image, as text feature Side information；
(1.4) dimensionality reduction, half of the dimensionality reduction to text feature dimension are carried out to color characteristic using PCA；
(1.5) the compact low-dimensional feature of image block is extracted using the matching kernel function based on local binary patterns
K s h a p e ( P , Q ) = Σ a ∈ P Σ a ′ ∈ Q s ~ ( a ) s ~ ( a ′ ) k b ( b ( a ) , b ( a ′ ) ) k p ( a , a ′ )
Wherein P represents two different image blocks with Q, and a and a ' represents the pixel in two image blocks respectively, S (a) is the standard deviation of pixel value in 3 × 3 fields centered on a, ∈sIt is the constant of a numerical value very little, B (a) is a binary column vector for pixel value difference in the local window around binaryzation a, standardized linear kernelFor weighing the contribution of each local binary patterns, gaussian kernel function kb(b (a), b (a '))=exp (- γb‖b (a)-b(a′)‖2) for measuring the shape similarity between two local binary patterns, Gauss position kernel function kp(a, a ')= exp(-γp‖a-a′‖2) for measuring the space length between two pixels；
(1.6) dimensionality reduction is carried out to low-dimensional feature using core principle component analysis method, dimension is equally chosen for 64 dimensions, obtaining can be effective Obtain the feature of image local shape and boundary information；
(1.7) merged color, shape and boundary information as supplementary information and SIFT feature, dimension is obtained for 256 Supplemental text feature.
3. the image classification method based on supplemental text feature as claimed in claim 1, is characterized in that, step 2 is specially：
(2.1) 200000 features are randomly selected in all of supplemental text feature of image from data set, each feature is 256 dimensions, 200000 image blocks may be constructed the matrix of 256 × 200000, each row of matrix represent one it is supplementary Text eigenvector；
(2.2) 512 are randomly selected as initial dictionary from 200000 image blocks, then initial dictionary can be with one 256 × 512 matrix is represented；
(2.3) with K-SVD method training dictionaries, that is, solve object function：
Wherein K is degree of rarefication, is set to 5, c hereiI-th row of presentation code C.The step of obtaining dictionary is as follows：
(2.3.1) dictionary is fixed, each y is solved by orthogonal matching pursuit methodiCorresponding rarefaction representation ci, wherein i's Span is 1 200000；
(2.3.2) residual error is calculated, and chooses the residual error item for indexing set that correspondence uses the kth of dictionary to arrange, residual error item is carried out Singular value decomposition, obtains singular value decomposition matrix, and wherein obtains 256 × 256 unitary matrice；
(2.3.3) the kth row of dictionary are updated with 256 × 256 unitary matrice first row；
(2.3.4) iterate, training obtains final dictionary, and iterations is set to 50 here.
4. the image classification method based on supplemental text feature as claimed in claim 1, is characterized in that, step 3 is specially：
(3.1) dictionary B is given, cataloged procedure is just to solve for the process of following formula：
arg m i n c | | x - B c | | 2 2 s . t . | | c | | 0 ≤ K
(3.2) first time iteration, selects input data x as current residue, and orthogonal matching pursuit algorithm is selected and current residue Optimal corresponding vision word；
(3.3) by x orthogonal mappings to the vision word for above selecting, and recalculate residual error；
(3.4) iterate, obtain sparse coding vector.
5. the image classification method based on supplemental text feature as claimed in claim 1, is characterized in that, step 4 is specially：
(4.1) to all of supplemental text feature in entire image, the maximum of its each dimension is sought, 512 dimensions are obtained Characteristic vector；
(4.2) entire image is divided into 4 image blocks, seeks maximum of each characteristics of image in each dimension in each image block respectively Value, obtains the characteristic vector of 4 512 dimensions；
(4.3) entire image is divided into 16 image blocks, ask in each image block respectively each characteristics of image in each dimension most It is big to be worth, obtain the characteristic vector of 16 512 dimensions；
(4.4) the characteristics or every unit vector for producing (4.1) step to (4.3) is normalized respectively：
F s ~ = F s | | F s | | p ∀ s = 1 , 2 , ... , S
FsFor the convergence feature that each unit is obtained, S is unit number, ‖ Fs‖pFor p norms；
(4.5) normalization characteristic of obtain 21 units is attached, obtains the character representation of image.
6. the image classification method based on supplemental text feature as claimed in claim 1, is characterized in that, step 5 is specially：
(5.1) random that 60 images are chosen from each class image of data set as such training sample, residual image is made For such test sample；
(5.2) characteristic vector of 100 images for choosing step (5.1) is trained in sending into Linear SVM grader, is obtained The SVM classifier for training；
(5.3) test sample chosen step (5.1) is sent in the SVM classifier that step (5.2) is trained, and obtains image Classification results.
CN201610939529.9A 2016-10-25 2016-10-25 Image classification method based on supplemented text characteristic Pending CN106529586A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610939529.9A CN106529586A (en) 2016-10-25 2016-10-25 Image classification method based on supplemented text characteristic

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610939529.9A CN106529586A (en) 2016-10-25 2016-10-25 Image classification method based on supplemented text characteristic

Publications (1)

Publication Number Publication Date
CN106529586A true CN106529586A (en) 2017-03-22

Family

ID=58292931

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610939529.9A Pending CN106529586A (en) 2016-10-25 2016-10-25 Image classification method based on supplemented text characteristic

Country Status (1)

Country Link
CN (1) CN106529586A (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107729812A (en) * 2017-09-18 2018-02-23 南京邮电大学 A kind of vehicle color suitable for monitoring scene knows method for distinguishing
CN108520539A (en) * 2018-03-13 2018-09-11 中国海洋大学 A kind of image object detection method based on sparse study variable model
CN108764275A (en) * 2018-04-10 2018-11-06 甘肃农业大学 The recognition methods of blade disease and system
CN108959322A (en) * 2017-05-25 2018-12-07 富士通株式会社 Information processing method and device based on text generation image
CN109741377A (en) * 2018-11-30 2019-05-10 四川译讯信息科技有限公司 A kind of image difference detection method
CN109858577A (en) * 2019-04-01 2019-06-07 盐城工学院 Soybean appearance quality detection device and detection method
CN110119776A (en) * 2019-05-10 2019-08-13 长沙理工大学 Recognition methods and its system based on Multiple Kernel Learning K-SVD
WO2020093423A1 (en) * 2018-11-09 2020-05-14 Hong Kong Applied Science and Technology Research Institute Company Limited Systems and methods for super-resolution synthesis based on weighted results from random forest classifier

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020647A (en) * 2013-01-08 2013-04-03 西安电子科技大学 Image classification method based on hierarchical SIFT (scale-invariant feature transform) features and sparse coding
CN103295242A (en) * 2013-06-18 2013-09-11 南京信息工程大学 Multi-feature united sparse represented target tracking method
CN103699902A (en) * 2013-12-24 2014-04-02 南京信息工程大学 Sorting method of ground-based visible light cloud picture
CN104346630A (en) * 2014-10-27 2015-02-11 华南理工大学 Cloud flower identifying method based on heterogeneous feature fusion
CN105005786A (en) * 2015-06-19 2015-10-28 南京航空航天大学 Texture image classification method based on BoF and multi-feature fusion

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020647A (en) * 2013-01-08 2013-04-03 西安电子科技大学 Image classification method based on hierarchical SIFT (scale-invariant feature transform) features and sparse coding
CN103295242A (en) * 2013-06-18 2013-09-11 南京信息工程大学 Multi-feature united sparse represented target tracking method
CN103699902A (en) * 2013-12-24 2014-04-02 南京信息工程大学 Sorting method of ground-based visible light cloud picture
CN104346630A (en) * 2014-10-27 2015-02-11 华南理工大学 Cloud flower identifying method based on heterogeneous feature fusion
CN105005786A (en) * 2015-06-19 2015-10-28 南京航空航天大学 Texture image classification method based on BoF and multi-feature fusion

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
KOEN E.A. VAN DE SANDE 等: "Evaluating Color Descriptors for Object and Scene Recognition", 《EVALUATING COLOR DESCRIPTORS》 *
LIEFENG BO 等: "Kernel Descriptors for Visual Recognition", 《PROCEEDINGS OF ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS》 *
罗会兰 等: "一种基于多级空间视觉词典集体的图像分类方法", 《电子学报》 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108959322A (en) * 2017-05-25 2018-12-07 富士通株式会社 Information processing method and device based on text generation image
CN108959322B (en) * 2017-05-25 2021-09-10 富士通株式会社 Information processing method and device for generating image based on text
CN107729812A (en) * 2017-09-18 2018-02-23 南京邮电大学 A kind of vehicle color suitable for monitoring scene knows method for distinguishing
CN107729812B (en) * 2017-09-18 2021-06-25 南京邮电大学 Method suitable for recognizing vehicle color in monitoring scene
CN108520539A (en) * 2018-03-13 2018-09-11 中国海洋大学 A kind of image object detection method based on sparse study variable model
CN108520539B (en) * 2018-03-13 2021-08-31 中国海洋大学 Image target detection method based on sparse learning variable model
CN108764275A (en) * 2018-04-10 2018-11-06 甘肃农业大学 The recognition methods of blade disease and system
WO2020093423A1 (en) * 2018-11-09 2020-05-14 Hong Kong Applied Science and Technology Research Institute Company Limited Systems and methods for super-resolution synthesis based on weighted results from random forest classifier
US10685428B2 (en) 2018-11-09 2020-06-16 Hong Kong Applied Science And Technology Research Institute Co., Ltd. Systems and methods for super-resolution synthesis based on weighted results from a random forest classifier
CN109741377A (en) * 2018-11-30 2019-05-10 四川译讯信息科技有限公司 A kind of image difference detection method
CN109858577A (en) * 2019-04-01 2019-06-07 盐城工学院 Soybean appearance quality detection device and detection method
CN110119776A (en) * 2019-05-10 2019-08-13 长沙理工大学 Recognition methods and its system based on Multiple Kernel Learning K-SVD

Similar Documents

Publication Publication Date Title
CN106529586A (en) Image classification method based on supplemented text characteristic
Yang et al. Learning face age progression: A pyramid architecture of gans
Song et al. Region-based quality estimation network for large-scale person re-identification
CN108154194B (en) Method for extracting high-dimensional features by using tensor-based convolutional network
CN104268593B (en) The face identification method of many rarefaction representations under a kind of Small Sample Size
CN105426919B (en) The image classification method of non-supervisory feature learning is instructed based on conspicuousness
CN105608690B (en) A kind of image partition method being combined based on graph theory and semi-supervised learning
CN109190643A (en) Based on the recognition methods of convolutional neural networks Chinese medicine and electronic equipment
CN105956560A (en) Vehicle model identification method based on pooling multi-scale depth convolution characteristics
CN110533024B (en) Double-quadratic pooling fine-grained image classification method based on multi-scale ROI (region of interest) features
Sun et al. Combining feature-level and decision-level fusion in a hierarchical classifier for emotion recognition in the wild
CN108804397A (en) A method of the Chinese character style conversion based on a small amount of target font generates
CN109389045A (en) Micro- expression recognition method and device based on mixing space-time convolution model
CN107133640A (en) Image classification method based on topography's block description and Fei Sheer vectors
CN105550712B (en) Aurora image classification method based on optimization convolution autocoding network
CN105654122B (en) Based on the matched spatial pyramid object identification method of kernel function
CN107085731A (en) A kind of image classification method based on RGB D fusion features and sparse coding
CN111881716A (en) Pedestrian re-identification method based on multi-view-angle generation countermeasure network
CN112818764A (en) Low-resolution image facial expression recognition method based on feature reconstruction model
CN109344898A (en) Convolutional neural networks image classification method based on sparse coding pre-training
CN108460400A (en) A kind of hyperspectral image classification method of combination various features information
Zhang et al. Video-based action recognition using rate-invariant analysis of covariance trajectories
CN104077742B (en) Human face sketch synthetic method and system based on Gabor characteristic
Kindiroglu et al. Temporal accumulative features for sign language recognition
CN110796022A (en) Low-resolution face recognition method based on multi-manifold coupling mapping

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20170322

WD01 Invention patent application deemed withdrawn after publication