CN103336974B - A kind of flowers classification discrimination method based on local restriction sparse representation - Google Patents

Abstract

The invention discloses a kind of flowers classification discrimination method based on local restriction sparse representation, first collect flowers image data base and set up flowers popular science knowledge storehouse, then extracting the various characteristics of image of flowers picture；The linear expression of training data feature and test data characteristics is set up followed by sparse coding theory；When modeling linear expression, not only consider that test characteristics of image is minimum with the linear expression error of training image feature set, increase the weight constraints of the locality characteristic Similarity Structure of test data characteristics and training data feature simultaneously, and the statistical gradient descending method utilizing kernel function to extend effectively solves, complete flowers classification identification learning process；Finally flowers image zooming-out feature to be identified is substituted into flowers classification judgement formula, obtains identification result, and recall, from flowers popular science knowledge storehouse, the explanatory note that this flowers classification is corresponding.The present invention has the advantage that recognition performance is high.

Description

A kind of flowers classification discrimination method based on local restriction sparse representation

Technical field

The present invention relates to pattern recognition and field of artificial intelligence, particularly to one based on local restriction sparse representation Flowers classification discrimination method.

Background technology

Flowers classification identification i.e. refers to utilize computer that the image information of flowers is carried out feature extraction, and computer is according to people's Understanding and mode of thinking in addition kind is sorted out and understands, and then can provide the user the flowers knowledge of some science popularization, and it belongs to The category of computer automatic object identification, in terms of flowers classification identification, patent is fewer at present, has a small amount of paper at academia Deliver, such as paper (Yuning CHAI, Victor LEMPITSKY, Andrew ZISSERMAN.BiCoS:A Bi-level Co-Segmentation Method for Image Classification.ICCV, 2011), it utilizes image partition method Divide the image into into foreground and background, then extract distribution of color and the super-pixel information of image, and calculation is effectively deduced in utilization Method is identified.Paper (Nilsback, M-E.and Zisserman, A.Automated flower Classification over a large number of classes, Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing2008), it extracts image Color histogram, SIFT feature, after three kinds of features of HOG feature, carry out the classification of flowers classification with svm classifier.Paper (Nilsback, M-E.and Zisserman, A, A Visual Vocabulary for Flower Classification, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition2006), it utilizes BOW feature to carry out the classification of flower chart picture.

Summary of the invention

In order to overcome the disadvantages mentioned above of prior art with not enough, it is an object of the invention to provide a kind of based on local restriction The flowers classification discrimination method of sparse representation, recognition performance is high.

The purpose of the present invention is achieved through the following technical solutions: a kind of flowers classification based on local restriction sparse representation is distinguished Knowledge method, comprises the following steps:

(1) collect flowers image data base and set up flowers popular science knowledge storehouse: utilizing existing plants and flowers wikipedia The title of S conventional flowers classification of definition, utilizes network search engines to search for flowers word corresponding to each flowers classification and is situated between Continue and picture, picture is formed flowers image data base；Explanatory note is included into flowers popular science knowledge storehouse；

(2) all pictures in the flowers image data base obtaining step (1) carry out feature extraction, and every pictures extracts Feature be m；

(3) flowers classification identification learning process:

(3-1) from flowers image data base, p pictures is selected, using its feature as test data set Y={y^k, wherein k =1...m represents m the feature that every pictures extracts respectively, and p ＜ N, N are the picture number in flowers image data base；Flower chart picture In data base, the feature of remaining N-p pictures is as training datasetWherein j=1,2 ... S, represent flowers Classification；

(3-2) will test data characteristics y^kUse training data featureLinear expression is:Wherein Being the weight coefficient of training dataset characterization test data set, its value is more than 0；

(3-3) minimum for training dataset and test data set linear expression error, increase test data characteristics simultaneously With the weight constraints of the locality characteristic Similarity Structure of training data feature, set up following optimization learning criterion:

$\underset{W_j^k}{\min }\frac{1}{2}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\left|\right|y^k-\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{X}_j^k{w}_j^k\left|\right|}_2^2+\mathrm{\λ}\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{\left|\right|D_j^k\mathrm{\Θ}{w}_j^k\left|\right|}_2---\left(1\right)$

Wherein Θ represents vector dot；λ is bound term weight, is used to leverage linear and represents between error and weight coefficient Constant variables；For test data characteristics y^kWith training data featureEuclidean distance；

(3-4) statistical gradient descent method pair is usedBeing iterated updating, renewal equation is:

$\left\{\begin{array}{c}{w}_j^{k,t+1}=w_j^{k,t}-{\mathrm{\η}\▿}_w\\ {\▿}_w=-X_j^k{y}^k+X_j^k{X}_j^k{w}_j^t+\mathrm{\λ}{\left|\right|y^k-X_j^k\left|\right|}_2^2\·w_j^{k,t}\end{array}\right.----\left(2\right)$

Wherein, t is the iterations of SGD iterative process, and η is the learning rate that statistical gradient declines iterative process；

(3-5) utilize nonlinear function φ that the feature of training and test data is carried out the nonlinear mapping regeneration to higher-dimension Core Hilbert space (reproducing kernel hilbert space is called for short RKHS), i.e. φ (x_i)^Tφ(x_j)=g (x_i,x_j), wherein g (x_ix_j) it is χ²Kernel function, x_iAnd x_jIt it is data characteristics；Formula (2) is transformed into:

$\left\{\begin{array}{c}{w}_j^{k,t+1}=w_j^{k,t}-{\mathrm{\η}\▿}_w\\ {\▿}_w={-h}^k+G^k{w}_j^{k,t}+\mathrm{\λ}(P^k-{2h}^h+G^k)\·w_j^{k,t}\end{array}\right.---\left(3\right)$

WhereinIt it is training data featureWith test data characteristics y^kDot product kernel function,It it is training data featureWith the dot product kernel function of self, P^k=φ(y^k)φ(y^k) it is that test data are special Levy y^kDot product kernel function with self；Through the successive ignition of formula (3), obtain the sign weight coefficient of optimum

(4) identification flowers: user shoots the image of flowers to be identified, to flowers image zooming-out characteristic Z to be identified^k, its Middle k=1...m represents m the feature that every pictures extracts respectively；

According to Z^k, flowers classification adjudicate formula identification flowers classification, and recall this flowers from flowers popular science knowledge storehouse The explanatory note that classification is corresponding；

Wherein flowers classification judgement formula is:

Wherein j^*Represent the linear expression of the training data utilizing certain j classification and the minimum error values of test data, logical Crossing minimum selection, j is and identifies the flowers classification obtained.

In step (3-4), η reduces along with the increase of iterations t,

Step (3-5) described χ²The expression formula of kernel function is exp (-χ²(x, x)/μ), wherein χ²Card side for symmetrical expression (Chi-squared) distance, μ is the χ of current training dataset²The average of distance.

The described title utilizing existing plants and flowers wikipedia S conventional flowers classification of definition of step (1), utilizes Network search engines searches for flowers character introduction corresponding to each flowers classification and picture, particularly as follows: utilize existing plant flowers The title of grass wikipedia 313 conventional herbage flower classifications of definition, utilizes network search engines to search for each flowers classification Corresponding flowers character introduction and picture, each flowers classification downloads 100 pictures.

The feature that every pictures extracts includes color histogram, SIFT feature (Scale-invariant feature Transform, scale invariant feature converting characteristic), HOG feature (Histogram of oriented gradients, direction Histogram of gradients feature), BOW feature (Bag of word, dictionary bag feature), SSIM feature (Structural Similarity, structural similarity feature), GB feature (Structural Similarity, structural similarity feature).

Compared with prior art, the present invention has the following advantages and beneficial effect:

(1), in flowers classification identification learning process, optimization learning criterion not only allows for current linear and represents that error is Little principle, but also by the bound term of based on local for image similar features structural modeling to learning criterion, thus image is online When property represents, system preferably selects most like characteristics of image and carries out linear expression, then while ensureing that linear coefficient is openness, Improve again the identification performance of system.

(2) frame structure of the present invention can seamlessly incorporate more characteristics of image, provides convenient for follow-up system upgrading, And more characteristics of image having identification also can improve the identification performance of system further.

(3) during the identification method of the present invention can be applied to the flowers Science Popularization System of reality well, due to the identification of the present invention Modeling method recognition performance is high, thus ensures robustness and the stability of flowers Science Popularization System.

Accompanying drawing explanation

Fig. 1 is the flow chart of the flowers classification discrimination method based on local restriction sparse representation of the present embodiment.

Fig. 2 is the statistical gradient descending method broad flow diagram of the present embodiment.

Detailed description of the invention

Below in conjunction with embodiment, the present invention is described in further detail, but embodiments of the present invention are not limited to this.

Embodiment

As it is shown in figure 1, the flowers classification discrimination method based on local restriction sparse representation of the present embodiment, including following step Rapid:

(1) collect flowers image data base and set up flowers popular science knowledge storehouse: utilizing existing plants and flowers wikipedia Define the title of 313 conventional herbage flower classifications, utilize the network search engines such as *** and Baidu to search for each flowers Flowers character introduction that classification is corresponding and picture, each flowers classification downloads 100 pictures.Picture is formed flowers view data Storehouse；Explanatory note is included into flowers popular science knowledge storehouse.

(2) all pictures in the flowers image data base obtaining step (1) carry out feature extraction, and every pictures extracts Feature be 6, (realize details with reference to Lowe, D.G., Distinctive Image including color histogram, SIFT feature Features from Scale-Invariant Keypoints,International Journal of Computer Vision, 60,2, pp.91-110,2004), HOG feature (realizes details with reference to N.Dalal and B.Triggs.Histogram of oriented gradients for human detection.In Computer Vision and Pattern Recognition, pp.887 893.IEEE, 2005), BOW feature (realize details reference S.Lazebnik,C.Schmid,and J.Ponce.Beyond bag of features:Spatial pyramid matching for recognizing natural scene categories.In Computer Vision and Pattern Recognition, pp.2169 2178.IEEE, 2006), SSIM feature (realizes details with reference to E.Shechtman and M.Irani.,Matching local self-similarities across images and videos.In Proc.CVPR, 2007), GB feature (realizes details with reference to A.C.Berg, T.L.Berg, and J.Malik., Shape matching and object recognition using low distortion correspondences.In Proc.CVPR, 2005).

(3) flowers classification identification learning process:

(3-1) from flowers image data base, 30 pictures are selected, using its feature as test data set Y={y^k, wherein 6 kinds of features of k=1...6 representative image respectively: color histogram, SIFT feature, HOG feature, BOW feature, SSIM feature and GB Feature；The feature of remaining 31270 pictures in flowers image data base is as training datasetWherein j= 1...313, represent the classification of flowers；

(3-2) will test data characteristics y^kUse training data featureLinear expression is:Wherein Being the weight coefficient of training dataset characterization test data set, its value is more than 0；

(3-3) minimum for training dataset and test data set linear expression error, increase test data characteristics simultaneously With the weight constraints of the locality characteristic Similarity Structure of training data feature, set up following optimization learning criterion:

$\underset{W_j^k}{\min }\frac{1}{2}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\left|\right|y^k-\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{X}_j^k{w}_j^k\left|\right|}_2^2+\mathrm{\λ}\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{\left|\right|D_j^k\mathrm{\Θ}{w}_j^k\left|\right|}_2---\left(1\right)$

Wherein Θ represents vector dot；λ is bound term weight, is used to leverage linear and represents between error and weight coefficient Constant variables (the present embodiment is set as 0.01)；Represent test data characteristics y^kThe training data feature similar with itLocality describe, be specifically defined and beFor test data characteristics y^kWith training data featureEuclidean distance；Under optimization learning criterion, it is possible to ensure that similar test feature can select similar training characteristics to enter Line linearity represents；In order to make the final linear expression coefficient can be sparse, typically can be by wherein weight coefficientIt is less than Certain threshold value (the present embodiment 0.005) is set to 0.

(3-4) use statistical gradient descent method (Stochastic gradient descent Method is called for short SGD) rightBeing iterated updating, renewal equation is:

$\left\{\begin{array}{c}{w}_j^{k,t+1}=w_j^{k,t}-{\mathrm{\η}\▿}_w\\ {\▿}_w=-X_j^k{y}^k+X_j^k{X}_j^k{w}_j^t+\mathrm{\λ}{\left|\right|y^k-X_j^k\left|\right|}_2^2\·w_j^{k,t}\end{array}\right.----\left(2\right)$

Wherein, t is the iterations of SGD iterative process, and η is the learning rate that statistical gradient declines iterative process；In order to protect Card learning criterion can effectively restrain, and the present embodiment sets learning rate η and reduces along with the increase of iterations t, $\mathrm{\η}=\frac{1}{1+100t};$

First statistical gradient descending method main flow process as in figure 2 it is shown, be to initialize weight coefficientThen with Machine selects part training datasetCalculate the kernel function that its training dataset is corresponding with test dataWithConnect And allow i=1 ..., n, utilize formula (3) to carry out n iteration of weight coefficient, after iteration is complete, computing formula (2) optimization object function Value, if it is less than last iteration process duration, randomization training dataset carries out next iteration circulation again, or shows Iteration has searched out optimal weight coefficient, and iteration terminates, and output optimum linearity represents weight coefficient；

(3-5) utilize nonlinear function φ that the feature of training and test data is carried out the nonlinear mapping regeneration to higher-dimension Core Hilbert space (reproducing kernel hilbert space is called for short RKHS), i.e. φ (x_i)^Tφ(x_j)=g (x_i,x_j), wherein g (x_i,x_j) it is χ²Kernel function, x_iAnd x_jIt it is data characteristics；χ²The expression formula of kernel function is exp(-χ²(x,x)/ μ, wherein χ²For card side (Chi-squared) distance of symmetrical expression, μ is the χ of current training dataset²The average of distance.

Formula (2) is transformed into:

$\left\{\begin{array}{c}{w}_j^{k,t+1}=w_j^{k,t}-{\mathrm{\η}\▿}_w\\ {\▿}_w=-X_j^k{y}^k+X_j^k{X}_j^k{w}_j^t+\mathrm{\λ}{\left|\right|y^k-X_j^k\left|\right|}_2^2\·w_j^{k,t}\end{array}\right.----\left(2\right)$

WhereinIt it is training data featureWith test data characteristics y^kDot product kernel function,It it is training data featureWith the dot product kernel function of self, P^k=φ(y^k)φ(y^k) it is that test data are special Levy y^kDot product kernel function with self；Through the successive ignition of formula (3), obtain the sign weight coefficient of optimum

(4) identification flowers: user shoots the image of flowers to be identified, to flowers image zooming-out characteristic Z to be identified^k, its Middle k=1...6 represents color histogram, SIFT feature, HOG feature, BOW feature, SSIM feature, GB feature respectively；

According to Z^k, flowers classification adjudicate formula identification flowers classification, and recall this flowers from flowers popular science knowledge storehouse The explanatory note that classification is corresponding；

Wherein flowers classification judgement formula is:

Wherein j^*Represent the linear expression of the training data utilizing certain j classification and the minimum error values of test data, logical Crossing minimum selection, j is and identifies the flowers classification obtained.

Above-described embodiment is the present invention preferably embodiment, but embodiments of the present invention are not by described embodiment Limit, the change made under other any spirit without departing from the present invention and principle, modify, substitute, combine, simplify, All should be the substitute mode of equivalence, within being included in protection scope of the present invention.

Claims (5)

1. a flowers classification discrimination method based on local restriction sparse representation, it is characterised in that comprise the following steps:

(1) collect flowers image data base and set up flowers popular science knowledge storehouse: utilizing existing plants and flowers wikipedia to define S The title of individual conventional flowers classification, utilizes network search engines to search for flowers character introduction corresponding to each flowers classification and figure Sheet, forms flowers image data base by picture；Explanatory note is included into flowers popular science knowledge storehouse；S is natural number；

(2) all pictures in the flowers image data base obtaining step (1) carry out feature extraction, the spy that every pictures extracts Levy as m；M is natural number；

(3) flowers classification identification learning process:

(3-1) from flowers image data base, p pictures is selected, using the feature of p pictures as test data set Y, every figure Test data characteristics y of sheet^kRepresent m the feature that every pictures extracts respectively as k=1...m, p ＜ N, N are flower chart picture Picture number in data base；

In flowers image data base, the feature of remaining N-p pictures is as training dataset X, and the training data of every pictures is special LevyWork as j=1,2 ... S, represent the classification of flowers, represent m the feature that every pictures extracts as k=1...m respectively；

(3-2) will test data characteristics y^kUse training data featureLinear expression is:WhereinIt it is training Data set characterizes the weight coefficient of test data set, and its value is more than 0；

(3-3) minimum for training dataset and test data set linear expression error, increase test data characteristics and instruction simultaneously Practice the weight constraints of locality characteristic Similarity Structure of data characteristics, set up following optimization learning criterion:

$\underset{w_j^k}{\min \frac{1}{2}}\underset{k=1}{\overset{m}{\Σ}}\left|\right|y^k-\underset{j=1}{\overset{S}{\Σ}}{X}_j^k{w}_j^k|{|}_2^2+\mathrm{\λ}\underset{j=1}{\overset{S}{\Σ}}|\left|D_j^k{\mathrm{\Θw}}_j^k\right|{|}_2---\left(1\right)$

Wherein Θ represents vector dot；λ is bound term weight, and be used to that leverage linear represents between error and weight coefficient is normal Number variable；For test data characteristics y^kWith training data featureEuclidean distance；

(3-4) statistical gradient descent method pair is usedBeing iterated updating, renewal equation is:

$\left\{\begin{array}{c}{w}_j^{k,t+1}=w_j^{k,t}-\mathrm{\η}{\▿}_w\\ {\▿}_w=-X_j^k{y}^k+X_j^k{X}_j^k{w}_j^{k,t}+\mathrm{\λ}\left|\right|y^k-X_j^k|{|}_2^2\·w_j^{k,t}\end{array}\right.---\left(2\right)$

Wherein, t is the iterations of SGD iterative process, and η is the learning rate that statistical gradient declines iterative process；

(3-5) utilize nonlinear function φ that the feature of training and test data is carried out nonlinear mapping to wish to the reproducing kernel of higher-dimension That Bert space, i.e. φ (x_i)^Tφ(x_j)=g (x_i,x_j), wherein g (x_i,x_j) it is χ²Kernel function, x_iAnd x_jIt it is data characteristics；Public Formula (2) is transformed into:

$\left\{\begin{array}{c}{w}_j^{k,t+1}=w_j^{k,t}-\mathrm{\η}{\▿}_w\\ {\▿}_w=-h^k+G^k{w}_j^{k,t}+\mathrm{\λ}(P^k-2h^k+G^k)\·w_j^{k,t}\end{array}\right.---\left(3\right)$

WhereinIt it is training data featureWith test data characteristics y^kDot product kernel function,It it is training data featureWith the dot product kernel function of self, P^k=φ (y^k)φ(y^k) it is test data Feature y^kDot product kernel function with self；Through the successive ignition of formula (3), obtain the sign weight coefficient of optimum

(4) identification flowers: user shoots the image of flowers to be identified, to flowers image zooming-out characteristic Z to be identified^k, wherein k= Represent m the feature that every pictures extracts the most respectively；

According to Z^k, flowers classification adjudicate formula identification flowers classification, and from flowers popular science knowledge storehouse, recall this flowers classification pair The explanatory note answered；

Wherein flowers classification judgement formula is:

Wherein j^*Represent the linear expression of the training data utilizing certain j classification and the minimum error values of test data, by minimum Selecting, j is and identifies the flowers classification obtained.

Flowers classification discrimination method based on local restriction sparse representation the most according to claim 1, it is characterised in that step Suddenly, in (3-4), η reduces along with the increase of iterations t,

Flowers classification discrimination method based on local restriction sparse representation the most according to claim 1, it is characterised in that step Suddenly (3-5) described χ²The expression formula of kernel function is exp (-χ²(x, x)/μ), wherein χ²For card side's distance of symmetrical expression, μ is current The χ of training dataset²The average of distance.

Flowers classification discrimination method based on local restriction sparse representation the most according to claim 1, it is characterised in that step Suddenly (1) the described title utilizing existing plants and flowers wikipedia S conventional flowers classification of definition, utilizes web search to draw Hold up and search for flowers character introduction corresponding to each flowers classification and picture, particularly as follows: utilize existing plants and flowers wikipedia Define the title of 313 conventional herbage flower classifications, utilize network search engines to search for the flowers that each flowers classification is corresponding Character introduction and picture, each flowers classification downloads 100 pictures.

Flowers classification discrimination method based on local restriction sparse representation the most according to claim 1, it is characterised in that every The feature that pictures extracts includes color histogram, SIFT feature, HOG feature, BOW feature, SSIM feature, GB feature.