CN108564111A - A kind of image classification method based on neighborhood rough set feature selecting - Google Patents

Abstract

An Image Classification Method Based on Neighborhood Rough Set Feature Selection. On the spatial pyramid model, firstly, the image features are extracted by SURF and HOG, which has scale-invariant characteristics and can describe the appearance and shape of local objects in the image, and the image feature selection algorithm based on neighborhood rough sets is used to eliminate SURF and HOG features The redundant features in; secondly, use the k-means clustering algorithm to generate a visual dictionary for the reduced image feature set; then, count the number of visual feature words in each scale image of the spatial pyramid model, and obtain the histogram All are connected in series, and corresponding weights are assigned to the features of different scales; finally, the weighted histogram is put into the linear SVM classifier for training and prediction. The present invention overcomes the defects in the existing image classification method that single feature extraction easily causes image information loss, while multi-feature fusion will generate a large number of redundant features, resulting in low image classification accuracy.

Description

A Method of Image Classification Based on Neighborhood Rough Set Feature Selection

technical field

The invention belongs to an image classification method in the field of computer vision.

Background technique

Image classification is an important research topic in the field of computer vision. The purpose of image classification is to enable computers to have the same ability to quickly and accurately identify complex visual images as humans. With the rapid development of artificial intelligence and pattern recognition, image classification is widely used in image understanding, object recognition and image retrieval.

Spatial Pyramid Matching (SPM) is one of the main image classification methods at present. Based on the bag of words (BOW), the spatial position and shape of the image are increased by multi-level division of the image. information. The image classification method based on the spatial pyramid model mainly includes four parts: feature extraction, generation of visual dictionary, construction of spatial pyramid and generation of image visual description by merging histogram statistics. Feature extraction and feature selection are the key prerequisites for image classification. Although the traditional spatial pyramid matching model (SPM) has made great breakthroughs in image classification, the classification performance is still not good because the extracted features cannot effectively express the image information. high. At present, feature extraction methods based on local descriptors are widely used in constructing visual dictionaries. The SIFT feature descriptor has better stability on images such as translation, rotation, and uneven illumination. The HOG descriptor can well represent the shape of the object in the image. Similar to SIFT features, SURF has the characteristics of scale and rotation invariance, but the calculation amount and calculation time are greatly reduced compared with SIFT, and it has the robustness of acceleration.

In the image feature extraction method, single feature extraction can describe the image one-sidedly, and cannot express the content of the image well. Multi-features can be used to describe image information more comprehensively. At present, there are a large number of image content expressed by combining multiple feature descriptors. While multi-features fully describe images, they also bring a lot of redundant information. How to remove unimportant or even redundant features from many features without affecting the expression of image content is crucial in image classification. Neighborhood rough set feature selection has a good effect in eliminating redundant features of continuous knowledge representation systems. After feature extraction, a more effective feature subset is formed through feature selection, the image information is simplified, and the basic information expressed by the image is not lost.

At present, in the feature extraction of image classification, there is still a single feature that cannot fully describe the image, but there will be a lot of redundant information in describing the image through multi-feature extraction. When clustering, a large number of visual dictionaries need to be generated, which will not only lead to low accuracy of image classification. High and time-consuming for image classification. Therefore, in image classification, feature selection after image feature extraction is also crucial.

Contents of the invention

The technical problem to be solved by the present invention is to provide an image classification method based on neighborhood rough set feature selection, which uses the complementary advantages between HOG and SURF to extract image features, and uses neighborhood rough set feature selection algorithm to eliminate Redundant features overcome the shortcomings of the existing image classification technology that there will be a lot of redundant information in the multi-feature extraction description image, resulting in low accuracy of image classification and large time-consuming image classification.

For solving the problems of the technologies described above, the invention provides a kind of image classification method of neighborhood rough set feature selection, and it comprises the following steps:

(1) Extract the features of the training sample and the test set sample image respectively;

(2) Build an image feature expression system;

(3) Use the feature selection algorithm based on the neighborhood rough set to eliminate the redundant features in the image knowledge expression system to obtain a new image feature set;

(4) clustering generates a visual feature dictionary;

(5) construct the spatial pyramid model, according to the visual characteristic dictionary of generation, count and merge the weighted visual characteristic histogram of each image spatial pyramid of training set and test set;

(6) Train a linear SVM classifier to classify the test image.

Extracting the features of the training sample and the test set sample image specifically includes: the SURF feature and the HOG feature of the image.

The SURF feature is a local feature descriptor. The SURF algorithm is similar to the SIFT algorithm and also has scale-invariant properties, but its calculation speed and robustness are better than SIFT.

The SURF feature extraction steps are:

Step 1. Use the box filter to construct a pyramid scale space;

In order to find feature points in different scale spaces, SURF introduces the concept of box filter to establish the scale space of the image. The division of the scale space of the SURF algorithm includes different orders and layers. Each stage consists of many different layers which are box filter responses of different sizes to the original image. The box filter size of the lowest-order first layer is 9×9, and the corresponding Gaussian scale is σ=1.2, and then the size will continue to increase to 15×15, 21×21, and 27×27.

Step 2. Establish a fast feature point detector through the Hessian matrix to obtain extreme stable points;

The detector of the SURF algorithm is based on the Hessian matrix. For a point X=(x, y) in the image I, the Hessian matrix of the point at the scale σ is H(p, σ):

where L _xx (P,σ) represents the second order partial derivative Convolved with image I at point P, L _xy (P,σ) represents Convolved with image I at point P, L _yx (P,σ) represents Convolved with image I at point P, L _yy (P,σ) represents Convolution with image I at point P; .

By calculating the determinant of the H matrix: det(H _approx )=D _xx D _yy -(ωD _xy ) ²

D _xx , D _yy , and D _xy are the approximations to L _xx , L _yy , and L _xy respectively after using filters, and ω is the weight, which generally takes a value of 0.9.

The value of the determinant represents the speckle response at X=(x, y, σ), and this function can be used to find the feature points in the image in space and scale. In order to obtain the feature points, the method of non-maximum value suppression is adopted here, and the feature point of a certain scale is compared with 26 adjacent points of the surrounding scale to observe whether it is a maximum value point.

Step 3. Describe the main direction of the feature point;

In order to make the feature points acquired in the image invariant to rotation, each feature point can be assigned a main direction. Taking the feature point as the center, calculate the area with a radius of 6s (s is the scale value of the layer where the feature point is located), select a fan-shaped area of 60°, and calculate the wavelet response vector superposition in the Harr wavelet response area in the x and y directions, Traversing the entire circle, the maximum vector direction is the main direction of the feature point.

Step 4. When constructing the feature vector, take a square frame with a side length of 20s around the feature point, and the direction of the frame is the main direction of the feature point. Divide the square frame into 4×4 sub-areas, and count the Harr wavelet responses in the x-direction and y-direction of 25 sampling points in each sub-area, and a 64-dimensional feature description vector can be obtained.

The HOG feature uses the overlapping local contrast normalization technology to characterize the appearance and shape of the local target in the image, and is one of the best features to describe the edge and shape information.

HOG feature extraction includes the following steps:

Step 1. Normalize the color image to remove the influence of lighting.

Step 2. Divide the image into pixel blocks (cells) of the same size, and calculate the horizontal and vertical gradients of each pixel point (x, y) in each cell as follows:

Gx(x,y)=G(x+1,y)-G(x-1,y)

G _y (x, y) = G(x, y+1)-G(x, y-1)

Then the size and direction of the gradient of the pixel point are obtained as:

Step 3. Concatenate all the features of the entire image to obtain the HOG features of the image. The HOG feature descriptor dimension is 36 (the parameters of HOG feature extraction in the present invention are set as follows: the cell size is 8×8 pixels, 2×2 cells form a block block, and the histogram of 9 bins is used to count each cell gradient information).

The image feature expression system of the construction is:

Construct image feature expression system NTD=<U,C,D>, U={u ₁ ,u ₂ ...,u _m }, u _i represents the set of feature vectors of the i-th image u _i =[X ₁ ,X ₂ ,…,X _n ], C={c ₁ ,c ₂ ,…,c _n } is the conditional attribute of the image knowledge representation system, c _l represents the lth feature vector of the image, and the dimension is the length of the image feature descriptor . The category D of the image is used as a decision attribute.

The relevant definitions of the neighborhood rough set feature selection algorithm specifically include:

Definition 1. The δ neighborhood of image u _i in the image feature expression system is: δ(u|Δ(u,u _i )≤δ). Δ is the distance function, and the distance function used in this paper is Chebyshev distance (infinite norm):

Under the same neighborhood radius, Chebyshev distance (infinite norm) represents the largest neighborhood range, and the calculation is simple.

Definition 2. The consistent neighborhood of image sample u is defined as: images of the same category in the neighborhood of image sample u, that is: for δ _c (u)∩δ _D (u); Conversely, the inconsistent neighborhood of image sample u is defined as: images of different categories in the neighborhood of image sample x, that is: for δ _c (u) - δ _D (u).

Definition 3: The information entropy and conditional entropy of image feature expression system NTD=<U,C,D> are defined as follows, information entropy and conditional entropy can represent the degree of uncertainty of image information:

Information entropy:

Conditional entropy:

The conditional entropy of an image feature representation system is related to the inconsistent neighborhood of its image samples. The larger the number of inconsistent neighborhoods of the image under each feature, the greater the conditional entropy, the less relevant the feature is to the image category; conversely, the smaller the conditional entropy, the more relevant the feature is to the image category. The degree of correlation between features and images can be reflected by conditional entropy.

The neighborhood rough set feature selection includes the following steps:

Step 1: Calculate the conditional entropy E(D|C) of the image knowledge expression system NTD=<U,C,D> according to the conditional entropy formula (3), reduce the feature set red, and initially X _i ∈ C-red;

Step 2: Calculate the conditional entropy of E(D|red∪{X _i }) according to formula (3), find out the feature _Xi with the smallest conditional entropy, and add _Xi to red;

Step 3: Calculate whether E(D|red∪{b _i }) is equal to E(D|C), if they are equal, output the feature reduction set red, if not, return to step 2.

The clustering generation visual feature dictionary includes:

The features obtained through feature selection are regarded as "visual words" and clustered using the k-means clustering algorithm to obtain a "bag of visual words" with a number of M.

Steps to build a spatial pyramid model:

1) Divide the image into three layers, the 0th layer takes the whole image as a region, the 1st layer evenly divides the image into 4 regions, the 2nd layer evenly divides the image into 16 regions, different layers are given different Weight, the weight is [1/4,1/4,1/2] in turn;

2) Count the pyramid statistics of different layers from left to right and from top to bottom to the frequency of "visual words" in each area of each layer in the "visual word bag", and obtain the image histogram of each area in different layers of the image Figure representation;

3) Assign the weights in 1) to the image histogram representations of each region in the above different layers, and connect them in series to obtain the final image representation

The training linear SVM classifier classifies the test images: that is, the training set images and the test set images are randomly selected, and the training set images are obtained through feature extraction, feature selection, generation of visual dictionaries, and construction of image pyramid spaces to obtain visual feature histograms. Get the trained classifier in the linear SVM; perform feature extraction on the test set image, and match the visual dictionary of the training set image to construct an image space pyramid model, obtain the histogram representation of the test set image, and input it to the trained linear In the SVM classifier, the category of the output test set image.

Compared with the prior art, the outstanding substantive features and notable features of the present invention are as follows:

(1) The present invention uses the combination of SURF and HOG features which are more complementary. The extracted features can not only describe the appearance and shape of the local target in the image, but also have scale invariance, fast calculation speed and good robustness Features.

(2) The present invention builds an image feature expression system, uses the feature selection algorithm of the neighborhood rough set to eliminate a large amount of redundant information brought by the combination of SURF and HOG, improves the accuracy of image classification and reduces the classification time.

Description of drawings

Fig. 1 is a flow chart of the present invention.

Detailed ways

Below in conjunction with accompanying drawing and embodiment, the specific embodiment of the present invention is described in further detail:

The invention provides an image classification method for neighborhood rough set feature selection, comprising the following steps:

1) Extract the features of training samples and test set sample images respectively;

2) Build an image feature expression system;

3) Use the feature selection algorithm based on the neighborhood rough set to eliminate the redundant features in the image knowledge expression system to obtain a new image feature set;

4) Clustering generates a visual feature dictionary;

5) construct the spatial pyramid model, according to the generated visual feature dictionary, count the weighted visual feature histogram of each image space pyramid of the combined training set and test set;

6) Train a linear SVM classifier to classify the test image.

The features extracted in step 1) specifically include: SURF features and HOG features of the image.

SURF is a local feature descriptor. The SURF algorithm is similar to the SIFT algorithm and also has scale-invariant properties, but its calculation speed and robustness are better than SIFT. The four basic steps of the SURF feature extraction algorithm are:

Step1. Use the box filter to construct the pyramid scale space;

In order to find feature points in different scale spaces, SURF introduces the concept of box filter to establish the scale space of the image. The division of the scale space of the SURF algorithm includes different orders and layers. Each stage consists of many different layers which are box filter responses of different sizes to the original image. The box filter size of the lowest-order first layer is 9×9, and the corresponding Gaussian scale is σ=1.2, and then the size will continue to increase to 15×15, 21×21, and 27×27.

Step2. Establish a fast feature point detector through the Hessian matrix to obtain extreme stable points;

The detector of the SURF algorithm is based on the Hessian matrix. For a point X=(x, y) in the image I, the Hessian matrix of the point at the scale σ is H(p, σ):

where L _xx (P,σ) represents the second order partial derivative Convolved with image I at point P, L _xy (P,σ) represents Convolved with image I at point P, L _yx (P,σ) represents Convolved with image I at point P, L _yy (P,σ) represents Convolution with image I at point P; .

By calculating the determinant of the H matrix: det(H _approx )=D _xx D _yy -(ωD _xy ) ²

D _xx , D _yy , and D _xy are the approximations to L _xx , L _yy , and L _xy respectively after using filters, and ω is the weight, which is generally 0.9.

The value of the determinant represents the speckle response at X=(x, y, σ), and this function can be used to find the feature points in the image in space and scale. In order to obtain the feature points, the method of non-maximum value suppression is adopted here, and the feature point of a certain scale is compared with 26 adjacent points of the surrounding scale to observe whether it is a maximum value point.

Step3. Describe the main direction of the feature point;

In order to make the feature points acquired in the image invariant to rotation, each feature point can be assigned a main direction. Taking the feature point as the center, calculate the area with a radius of 6s (s is the scale value of the layer where the feature point is located), select a fan-shaped area of 60°, and calculate the wavelet response vector superposition in the Harr wavelet response area in the x and y directions, Traversing the entire circle, the maximum vector direction is the main direction of the feature point.

Step4. When constructing the feature vector, take a square frame with a side length of 20s around the feature point, and the direction of the frame is the main direction of the feature point. Divide the square frame into 4×4 sub-areas, and count the Harr wavelet responses in the x-direction and y-direction of 25 sampling points in each sub-area, and a 64-dimensional feature description vector can be obtained.

The HOG feature uses the overlapping local contrast normalization technology to characterize the appearance and shape of the local target in the image, and is one of the best features to describe the edge and shape information. HOG feature extraction can be divided into the following steps:

Step 1. Normalize the color image to eliminate the influence of lighting.

Step 2. Divide the image into pixel blocks (cells) of the same size, and calculate the horizontal and vertical gradients of each pixel point (x, y) in each cell as follows:

G _x (x,y)=G(x+1,y)-G(x-1,y)

G _y (x,y)=G(x,y+1)-G(x,y-1)

Then the size and direction of the gradient of the pixel point are obtained as:

Step3. Concatenate all the features of the entire image to obtain the HOG features of the image. The HOG feature descriptor dimension is 36 (the parameters of HOG feature extraction in the present invention are set as follows: the cell size is 8×8 pixels, 2×2 cells form a block block, and the histogram of 9 bins is used to count each cell gradient information).

Wherein step 2) the image feature expression system described in construction is:

Construct image feature expression system NTD=<U,C,D>, U={u ₁ ,u ₂ ...,u _m }, u _i represents the set of feature vectors of the i-th image u _i =[X ₁ ,X ₂ ,…,X _n ], C={c ₁ ,c ₂ ,…,c _n } is the conditional attribute of the image knowledge representation system, c _l represents the lth feature vector of the image, and the dimension is the length of the image feature descriptor . The category D of the image is used as a decision attribute.

Wherein the relevant definitions of the neighborhood rough set feature selection algorithm described in step 3) specifically include:

Definition 1. The δ neighborhood of image u _i in the image feature expression system is: δ(u|Δ(u,u _i )≤δ). Δ is the distance function, and the distance function used in this paper is Chebyshev distance (infinite norm):

Under the same neighborhood radius, Chebyshev distance (infinite norm) represents the largest neighborhood range, and the calculation is simple.

Definition 2. The consistent neighborhood of image sample u is defined as: images of the same category in the neighborhood of image sample u, that is: for δ _c (u)∩δ _D (u); Conversely, the inconsistent neighborhood of image sample u is defined as: images of different categories in the neighborhood of image sample x, that is: for δ _c (u) - δ _D (u).

Definition 3: Literature [11] defines the information entropy and conditional entropy of NTD=<U,C,D>, which can represent the degree of uncertainty of image information:

Information entropy:

Conditional entropy:

The conditional entropy of an image feature representation system is related to the inconsistent neighborhood of its image samples. The larger the number of inconsistent neighborhoods of the image under each feature, the greater the conditional entropy, the less relevant the feature is to the image category; conversely, the smaller the conditional entropy, the more relevant the feature is to the image category. The degree of correlation between features and images can be reflected by conditional entropy.

Wherein step 3) described neighborhood rough set feature selection algorithm comprises the following steps:

Step 1: According to the conditional entropy formula (3), we can calculate the conditional entropy E(D|C) of the image knowledge expression system NTD=<U,C,D>, reduce the feature set red, and initially X _i ∈ C-red;

Step 2: Also calculate the conditional entropy of E(D|red∪{X _i }) according to the formula (3), find out the feature X _i with the smallest conditional entropy, and add Xi _i to red;

Step 3: Calculate whether E(D|red∪{b _i }) is equal to E(D|C), if they are equal, output the feature reduction set red, if not, return to step 2.

Wherein step 4) specifically includes:

The features obtained through feature selection are regarded as "visual words" and clustered with the k-means clustering algorithm to obtain a "visual dictionary" with a number of M.

Wherein step 5) based on the spatial pyramid model SPM specific steps include:

1) Divide the image into three layers, the 0th layer takes the whole image as a region, the 1st layer evenly divides the image into 4 regions, the 2nd layer evenly divides the image into 16 regions, different layers are given different Weight, the weight is [1/4,1/4,1/2] in turn;

2) Count the pyramid statistics of different layers from left to right and from top to bottom to the frequency of "visual words" in each area of each layer in the "visual word bag", and obtain the image histogram of each area in different layers of the image Figure representation;

3) Assign the weights in 1) to the image histogram representations of each region in the above different layers, and connect them in series to obtain the final image representation

Wherein step 6) specifically comprises:

1) Randomly select a training set and a test set, carry out the first 5 steps of claim 1 to the training set image, obtain the histogram representation of the training set image, and input the trained classifier in the linear SVM;

2) Perform feature extraction on the test set image, and match the visual dictionary of the training set image to construct an image space pyramid model, obtain the histogram representation of the test set image, input it into the trained linear SVM classifier, and output the test set image category.

Claims (8)

1. a kind of image classification method based on neighborhood rough set feature selecting, which is characterized in that include the following steps：

(1) feature of training sample and test sample image is extracted；

(2) training sample image feature representation system is built；

(3) redundancy feature in image knowledge expression system is rejected with the feature selecting algorithm based on neighborhood rough set, Obtain new set of image characteristics；

(4) cluster generates visual dictionary；

(5) spatial pyramid model is built, according to the visual dictionary of generation, statistics merges the weighting of each image spatial pyramid Visual signature histogram；

(6) training Linear SVM grader classifies to test image.

2. the image classification method as described in claim 1 based on neighborhood rough set feature selecting, which is characterized in that extraction instruction Practice sample and test sample image refers to the SURF features and HOG features for extracting image respectively；

The SURF characteristic extraction steps are：

Step 1. utilizes cassette filter construction pyramid scale space；

Step 2. establishes rapid characteristic points detector acquisition extreme value stable point by Hessian matrixes；

The detector of SURF algorithm is based on Hessian matrixes, and for certain point X=(x, y) in image I, then the point is at scale σ Hessian matrixes be H (p, σ)：

Wherein L_xx(P, σ) is indicatedWith convolution of the image I at point P, L_xy (P, σ) is indicatedWith convolution of the image I at point P, L_yx(P, σ) is indicatedWith volumes of the image I at point P Product, L_yy(P, σ) is indicatedWith convolution of the image I at point P；

Calculate the determinant of H-matrix：det(H_approx)=D_xxD_yy-(ωD_xy)²

D_xx, D_yy, D_xyIt is after having used filter respectively to L_xx, L_yy, L_xyApproximation, ω is weight, and ω values are 0.9；

Step 3. Expressive Features point principal direction；

I.e. centered on characteristic point, the region that radius is 6s is calculated, s is characterized a place figure layer scale-value, selects 60 ° of sector Region calculates small echo response vector superposition in the Harr small echo response regions of sector region in the x and y direction, traverses entire circle, The principal direction of maximum vector direction, that is, characteristic point；

For step 4. in construction feature vector, it is the square-shaped frame of 20s that the length of side is taken around characteristic point, and the direction of square-shaped frame is Characteristic point principal direction；Square-shaped frame is divided into 4 × 4 sub-regions, counts the directions x and the side y per 25 sampled points of sub-regions To Harr small echos response, obtain one 64 dimension feature description vector；

The HOG characteristic extraction steps:

Step 1. normalized color image, eliminates the influence of illumination；

Step 2. divides the image into the block of pixels of same size, calculates the horizontal direction of each pixel (x, y) in each block of pixels Gradient with vertical direction is respectively：

G_x(x, y)=G (x+1, y)-G (x-1, y)

G_y(x, y)=G (x, y+1)-G (x, y-1)

The size and Orientation for then obtaining the gradient of the pixel is respectively：

Step 3. connects all features of entire image to obtain the HOG features of image.

3. the image classification method as described in claim 1 based on neighborhood rough set feature selecting, which is characterized in that structure instruction Practicing sample image feature representation system is：

Build image feature representation system NTD=<U,C,D>, U={ u₁,u₂…,u_m},u_i=[X₁,X₂,…,X_n] indicate i-th The set of the feature vector of image, C={ c₁,c₂,…,c_nBe image knowledge expression system conditional attribute, c_lIndicate image First of feature vector, dimension are the length of image feature descriptor.

4. the image classification method as described in claim 1 based on neighborhood rough set feature selecting, which is characterized in that based on neighbour Domain rough set image feature selection includes：

(1) image u in image feature representation system_iδ neighborhoods be：δ(u|Δ(u,u_i)≤δ), Δ is distance function；

(2) the consistent neighborhood definition of image pattern u is：The identical image of classification in the neighborhood of image pattern u, i.e.,：Forδ_c(u)∩δ_D(u)；Otherwise the inconsistent neighborhood definition of image pattern u is：Class in the neighborhood of image pattern x Not different images, i.e.,：Forδ_c(u)-δ_D(u)。

(3) document [11] defines NTD=<U,C,D>Comentropy and conditional entropy, comentropy and conditional entropy：

Comentropy：

Conditional entropy： 。

5. the image classification method as described in claim 1 based on neighborhood rough set feature selecting, which is characterized in that based on neighbour The image feature selection step of domain rough set：

Step 1：According to conditional entropy formula

Calculate image knowledge expression system NTD=<U,C,D>Conditional entropy E (D | C), yojan feature set red, initiallyX_i∈C-red；

Step 2：According to formula

Calculating E (D | red ∪ { X_i) conditional entropy, find out the feature X of conditional entropy minimum_i, by X_iIt is added in red；

Step 3：Calculating E (D | red ∪ { b_i) whether it is equal to E (D | C), if E (D | red ∪ { b_i)=E (D | C), then feature is exported Yojan collection red；If E (D | red ∪ { b_i) ≠ E (D | C), then 2 are returned to step.

6. the image classification method as described in claim 1 based on neighborhood rough set feature selecting, which is characterized in that cluster life Include at visual dictionary：

The feature obtained by feature selecting be considered as " vision word " using k-means clustering algorithms it is carried out cluster to Obtain " the vision bag of words " that quantity is M.

7. the image classification method as described in claim 1 based on neighborhood rough set feature selecting, which is characterized in that based on sky Between pyramid model SPM the specific steps are：

(1) image is divided into three layers, the 0th layer using entire image as a region, image uniform is divided into 4 areas by the 1st floor Domain, image uniform is divided into 16 regions by the 2nd layer, and different layers are assigned to different weights, and weight is followed successively by [1/4,1/4,1/ 2]；

(2) " vision list of the sequence of the pyramid statistics of statistics different layers from left to right, from top to bottom to each region in each layer The frequency that word " occurs in " vision bag of words ", the image histogram for obtaining each region in image different layers indicate；

(3) in above-mentioned different layers each region image histogram indicate be assigned to 1 respectively) in weight, be together in series to obtain most Whole graphical representation.

8. the image classification method as described in claim 1 based on neighborhood rough set feature selecting, which is characterized in that training line Property SVM classifier to test image carry out classification include：

(1) training set and test set are randomly selected, preceding 5 step process of claim 1 are carried out to training set image, are trained Collect the histogram expression of image, the grader after being trained in input linear SVM；

(2) feature extraction is carried out to test set image, and matches the visual dictionary structure image space pyramid of training set image Model, the histogram for obtaining test set image indicates, is entered into trained Linear SVM grader, exports test set The classification of image.