CN103440501A - Scene classification method based on nonparametric space judgment hidden Dirichlet model - Google Patents

Abstract

The invention discloses a scene classification method based on a non-parametric space judgment hidden Dirichlet model. It mainly solves the shortcomings of existing classification methods that do not contain scene space information. The implementation steps are: (1) input image; (2) extract image block features; (3) initialize non-parametric space decision hidden Dirichlet model parameters; (4) establish non-parametric space decision hidden Dirichlet model; (5) ) image scene classification. The invention utilizes image blocks containing spatial information, can describe image scenes more abundantly, and improves the accuracy rate of image scene classification.

Description

A Scene Classification Method Based on Nonparametric Spatial Decision Hidden Dirichlet Model

technical field

The invention belongs to the technical field of image processing, and further relates to a scene classification method based on a Nonparametric Spatial Discriminative Latent Dirichlet Allocation (NS-DiscLDA) model in the technical field of pattern recognition. The invention can be used for scene classification of natural images and improves the accuracy of scene classification.

Background technique

Scene classification is one of the basic tasks of image understanding, and it plays a very important role in scene recognition. Traditional scene classification is usually based on three methods: first, a scene classification method based on image collections based on graph analysis; second, a scene classification method based on supervised manifold learning; third, a scene classification method based on objects and their spatial relationship characteristics method.

A scene classification method is disclosed in the patent "Scene classification method and device based on image collection based on spectrogram analysis" (application number: 201110221407.3 application date: 2011-08-03 application publication number: CN102542285A) applied by Tsinghua University. This method determines the degree of membership through interaction time, which mainly solves the problem of non-linear data loss in existing methods, thereby improving the classification accuracy. The implementation of the method is as follows: first extract the SIFT feature set of the image set, and obtain K clusters and K codewords, establish a weighted map according to the SIFT features and codewords of any image, and determine the weighted map and the European node of any node. K' nodes with the closest distance to obtain the weight matrix corresponding to the node set, and then obtain the Laplacian operator matrix according to the weight matrix, and perform operations on the Laplacian operator matrix to obtain each SIFT feature of any image and K The interaction time between codewords, the membership degree is determined according to the interaction time, and finally the codeword assignment result is determined according to the membership degree, and the scene is classified according to the assignment result. However, the disadvantage of the method of this patent application is that simply using a classifier to classify image scenes lacks semantic information in the scenes, thereby reducing the accuracy of scene classification.

A scene classification method is disclosed in the patent "Scene Classification Method and Device Based on Supervised Manifold Learning" (application number: 201110202756.0 application date: 2011-07-19 application publication number: CN102254194A) applied by Tsinghua University. This method uses manifold learning to classify image scenes, and mainly solves the problem that existing methods do not consider the manifold characteristics of high-dimensional feature points. The implementation of the method is as follows: first extract image features and obtain the codebook composed of the cluster centers of the features, then obtain the metrics of each feature on each manifold structure to the codeword, and calculate the membership of the feature of the test image to the codeword degree and get the histogram vector, and finally use the support vector machine to learn the histogram vector to get the scene category of the image. However, the disadvantages of the method of this patent application are: the method uses manifold learning technology, but the classification ability of manifold learning technology is weak, which leads to a decrease in the accuracy of scene classification, and the calculation of this method is complicated If the degree is too high, the speed of scene classification will be reduced.

A scene is disclosed in the patent "A method for classifying image scenes based on objects and their spatial relationship characteristics" (application number: 201110214985.4 application date: 2011-07-29 application publication number: CN102902976A) applied by the Institute of Electronics, Chinese Academy of Sciences Classification. This method improves the accuracy of scene classification by fusing the spatial relationship characteristics between topics. The implementation of the method is as follows: firstly define a spatial relationship histogram to characterize the spatial relationship between objects, then adopt a probabilistic implicit semantic analysis model that integrates the spatial relationship characteristics between topics, establish an image model, and finally use a support vector machine to classify scene image. However, the disadvantage of the method of this patent application is that since the method uses the pLSA topic model modeling method, the pLSA topic model lacks prior information, resulting in the loss of detailed information and reducing the accuracy of scene classification.

Contents of the invention

The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art, and propose a scene analysis method based on a non-parametric spatial decision hidden Dirichlet model, to describe images more comprehensively and improve the accuracy of scene classification.

The technical idea of realizing the present invention is to divide the image evenly into many 8×8 image blocks, extract the SIFT features of the image blocks, obtain the spatial coordinates of the image blocks, and use the features and spatial coordinates of the image blocks to establish a non-parametric spatial decision hidden The Likelet model enables the model to include the spatial information of the image block, so as to achieve a more comprehensive description of the image and improve the accuracy of scene classification.

To achieve the above object, the present invention comprises the following main steps:

(1) Input image: Input the training image manually labeled scene category.

(2) Extract image block features.

Divide the training image into multiple 8×8 image blocks, extract SIFT features for each image block, and record the spatial coordinates of each image block.

(3) Initialize model parameters: Manually initialize the non-parametric spatial decision hidden Dirichlet model to obtain the spatial distribution parameters of scene elements.

(4) Establish a non-parametric spatial decision hidden Dirichlet model.

The parameters of word distribution for each topic in the model are estimated, the features and spatial coordinates of image patches are statistically modeled, and a non-parametric spatial decision latent Dirichlet model is established.

(5) Image scene classification.

According to the non-parametric space decision hidden Dirichlet model, the category label of the test image is predicted, and the classification of the image scene is completed.

Compared with existing methods, the present invention has the following advantages:

First, the present invention records the spatial coordinates of the image block when extracting the features of the image block, which overcomes the disadvantage of not including spatial information in the prior art, makes the image information of the present invention more complete, and improves the accuracy of image scene classification.

Second, the present invention performs statistical modeling on the features and spatial coordinates of the image blocks, so that the features of the image blocks are related to each other through the spatial coordinates, which overcomes the disadvantage that the information of the image blocks in the image representation method of the prior art is irrelevant, and makes the present invention The image representation method keeps the comprehensiveness of image information.

Third, the non-parametric space decision hidden Dirichlet model established by the present invention utilizes the word distribution of the topic for modeling, which makes modeling easier, overcomes the shortcomings of poor image modeling capabilities in the prior art, and makes the present invention Demonstrated better modeling ability.

Description of drawings

Fig. 1 is a flowchart of the present invention;

Fig. 2 is a diagram of the hidden Dirichlet model of non-parametric space decision in the present invention.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings.

With reference to accompanying drawing 1, the step that the present invention is realized is as follows:

Step 1, input image.

Input training images with human-annotated scene categories. The manual labeling in the present invention refers to labeling the natural image category marks on all the training images respectively.

Step 2, extract image block features.

Divide the training image into multiple 8×8 image blocks, extract SIFT features for each image block, and record the spatial coordinates of each image block.

The steps of SIFT feature extraction are as follows:

In the first step, the image blocks to be extracted with SIFT features are composed into an image block set.

In the second step, five scale values of 0.5, 0.8, 1.1, 1.4, and 1.7 are selected for the scale value σ, and the five scale values are respectively brought into the following formula to obtain five Gaussian functions of different scales.

$\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}$

Among them, G(x, y, σ) represents the Gaussian function under the σ scale value, and x and y represent the abscissa and ordinate values corresponding to the pixels of the image block, respectively.

In the third step, each image block in the image block set in the first step is convoluted with five Gaussian functions of different scales to obtain a first-order five-layer image set.

In the fourth step, each image in the first-order five-layer image set is sampled at intervals to obtain a second-order five-layer image set.

In the fifth step, each image in the second-order five-layer image set is sampled at intervals to obtain a third-order five-layer image set.

The sixth step is to subtract images of adjacent stages in the same layer to obtain a second-order five-layer differential image set.

The seventh step is to obtain the second-order five-layer difference image set of all images, and the second-order five-layer difference image set of all images is the Gaussian difference scale space.

In the eighth step, each pixel of the image in the Gaussian difference scale space is compared with the 8 pixels adjacent to the pixel and the 18 pixels adjacent to the upper and lower images of the same order to compare the gray value , to determine whether the pixel is an extreme point, if the pixel is an extreme point, it is marked as a feature point, otherwise, it is not marked.

The ninth step is to calculate the principal curvature ratio of each feature point in the Gaussian difference scale space according to the following formula.

$\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;\&beta;</span>}}$

Among them, C represents the principal curvature ratio of the feature points in the Gaussian difference scale space, and α and β respectively represent the gradient values of the feature points in the Gaussian difference scale space in the horizontal and vertical coordinate directions of the image pixel.

The tenth step is to judge whether the principal curvature ratio of each feature point in the Gaussian difference scale space is less than the principal curvature ratio threshold 10, and if so, mark the feature point as a key point, otherwise, do not mark it.

In the eleventh step, calculate the gradient magnitude of each key point of the image in the Gaussian difference scale space according to the following formula.

$\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Among them, m(x, y) represents the gradient magnitude of the key point, x and y represent the horizontal and vertical coordinate values corresponding to the key point, respectively, and L represents the scale of the key point in the scale space.

In the twelfth step, calculate the gradient direction of each key point of the image in the Gaussian difference scale space according to the following formula.

θ(x,y)=αtan2{[L(x,y+1)-L(x,y-1)]/[L(x+1,y)-L(x-1,y)]}

Among them, θ(x, y) represents the gradient direction of the key point, x and y represent the horizontal and vertical coordinate values corresponding to the key point respectively, α represents the gradient value of the key point in the horizontal coordinate direction, and L represents the key point in the scale space scale.

The thirteenth step is to count the gradient direction and magnitude of 8×8 pixels around each key point in the Gaussian difference scale space to obtain a gradient histogram, where the horizontal axis of the gradient histogram is the gradient direction angle, and the vertical axis is The magnitude corresponding to the gradient orientation angle.

The fourteenth step is to rotate the coordinate axis of the Gaussian difference scale space to the direction of the key point, calculate the 8-dimensional vector representation of each key point sub-region, combine the 8-dimensional vectors of all key point sub-regions, and obtain the value of each key point in 128-dimensional SIFT features.

Step 3, initialize the model parameters.

The non-parametric spatial decision latent Dirichlet model is manually initialized to obtain the spatial distribution parameters of scene elements.

The steps of model initialization are as follows:

The first step is to find out whether the spatial layout information of scene elements exists in the training image, if it exists, go to the second step, otherwise, go to the third step.

In the second step, the spatial layout information of the scene elements of the training image is used as the spatial distribution parameter of the scene elements.

In the third step, the training image is evenly divided into multiple 8×8 image blocks, and the spatial distribution parameters of scene elements are obtained by the method of statistical image blocks.

Step 4, establishing a non-parametric spatial decision hidden Dirichlet model.

The parameters of word distribution for each topic in the model are estimated, the features and spatial coordinates of image patches are statistically modeled, and a non-parametric spatial decision latent Dirichlet model is established.

The process of establishing the hidden Dirichlet model of non-parametric space decision in the present invention will be further described in detail below in conjunction with FIG. 2 .

The steps of the present invention to establish the hidden Dirichlet model of non-parametric space decision are as follows:

In the first step, the probability distribution R of image subjects is estimated according to the distribution parameters of the scene element space.

The second step is to randomly sample the probability distribution θ _d of the image d (d=1,2,...,D) on the topic, and obtain the sampling sample z _dn of the image block topic, and α represents the distribution parameter of the variable θ _d , k (k=1,2,...,K ₁ ) represents a topic.

The third step is to estimate the word distribution parameter φ _k of the image block topic according to the sampling sample z _dn of the image block topic, and β represents the distribution parameter of the variable φ _k .

The fourth step is to establish a non-parametric spatial decision hidden Dirichlet model according to the word distribution parameter φ _k of the subject of the image block, where π represents the distribution of category labels, which is uniform distribution in the present invention, and y _d represents the category label of image d , T represents the mapping matrix of the scene, N _d represents the image block, u _dn represents the hidden theme of the image block dn, w _dn represents the word number of the image block dn, l _dn represents the space coordinates of the image block dn.

Step 5, image scene classification.

According to the non-parametric space decision hidden Dirichlet model, the category label of the test image is predicted, and the classification of the image scene is completed.

In the first step, the test image is brought into the non-parametric space decision hidden Dirichlet model, and the probability distribution of the test image category is obtained.

In the second step, the category with the highest probability in the category probability distribution of the test image is used as the category label of the test image.

The effects of the present invention can be further illustrated by the following simulation experiments.

1. Simulation conditions

The present invention is the emulation that uses MATLAB software to carry out on the central processor being Intel (R) Core i3-5302.93GHZ, internal memory 4G, WINDOWS7 operating system. The database uses LabelMe database and UIUC-Sports database.

2. Simulation content

The invention performs scene classification simulation experiments on the image scene database. The test images use LabelMe database and UIUC-Sports database. The LabelMe database contains 8 scene classifications, namely "highway", "city", "high building", "street", "forest", "coast", "mountain" and "country". The UIUC-Sports database contains 8 scene categories, namely "badminton", "boccia", "croquet", "polo", "rock climbing", "rowing", "sailing" and "skiing".

The present invention evaluates the performance of the method with the classification accuracy rate as an index, and compares the accuracy rate of different scene classification methods for classifying image scenes in simulation. The comparison of multiple scene classification methods includes spatial latent Dirichlet Allocation (SLDA) ) method, Discriminative Latent Dirichlet Allocation (DiscLDA) method, Spatial Discriminative Latent Dirichlet Allocation (S-DiscLDA) method and the method of the present invention. The comparison experiment results are shown in the table below.

It can be seen from the above table that the classification accuracy rate of the present invention is the highest among the four methods in the scene classification experiments on the two databases. This is because the present invention highlights the image blocks containing spatial information, so it can better describe the image scene, thereby obtaining the effect of classification accuracy superior to other scene classification methods, which further verifies the advanced nature of the present invention.

Claims (6)

1. A scene classification method based on non-parametric space judgment hidden Dirichlet model, it is characterized in that, comprising the following steps: (1) Input image: input the training image manually marked with the scene category; (2) Extract image block features: Divide the training image into multiple 8×8 image blocks, extract SIFT features for each image block, and record the spatial coordinates of each image block; (3) Initialize model parameters: Manually initialize the non-parametric spatial decision hidden Dirichlet model to obtain the spatial distribution parameters of scene elements; (4) Establish a non-parametric spatial decision hidden Dirichlet model: Estimate the parameters of the word distribution for each topic in the model, statistically model the features and spatial coordinates of image patches, and build a non-parametric spatial decision latent Dirichlet model; (5) Image scene classification: According to the non-parametric space decision hidden Dirichlet model, the category label of the test image is predicted, and the classification of the image scene is completed. 2. the scene classification method based on non-parametric space judgment hidden Dirichlet model according to claim 1, is characterized in that, the labeling scene category described in step (1) refers to labeling natural image category mark respectively to all training images . 3. the scene classification method based on non-parametric space judgment hidden Dirichlet model according to claim 1, is characterized in that, the described step of extracting SIFT feature in step (2) is as follows: In the first step, the image blocks to be extracted with SIFT features are composed into an image block set; In the second step, five scale values of 0.5, 0.8, 1.1, 1.4, and 1.7 are selected for the scale value σ, and the five scale values are respectively brought into the following formula to obtain five Gaussian functions of different scales; $\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}$ Among them, G(x, y, σ) represents the Gaussian function under the σ scale value, and x and y represent the horizontal and vertical coordinate values corresponding to the pixels of the image block, respectively; In the third step, each image block in the image block set in the first step is convoluted with five Gaussian functions of different scales to obtain a first-order five-layer image set; In the fourth step, each image of the first-order five-layer image set is sampled at intervals to obtain a second-order five-layer image set; In the fifth step, each image of the second-order five-layer image set is sampled at intervals to obtain a third-order five-layer image set; The sixth step is to subtract images of adjacent stages in the same layer to obtain a second-order five-layer differential image set; The seventh step is to obtain the second-order five-layer difference image set of all images, and the second-order five-layer difference image set of all images is the Gaussian difference scale space; The eighth step is to compare the gray value of each pixel of the image in the Gaussian difference scale space, the 8 pixels adjacent to the pixel and the 18 pixels adjacent to the upper and lower images of the same order , to determine whether the pixel is an extreme point, if the pixel is an extreme point, mark it as a feature point, otherwise, do not mark it; The ninth step is to calculate the principal curvature ratio of each feature point in the Gaussian difference scale space according to the following formula; $\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;\&beta;</span>}}$ Among them, C represents the principal curvature ratio of the feature points in the Gaussian difference scale space, and α and β respectively represent the gradient values of the feature points in the Gaussian difference scale space in the horizontal and vertical coordinate directions of the image pixels; The tenth step is to judge whether the principal curvature ratio of each feature point in the Gaussian difference scale space is less than the principal curvature ratio threshold 10, and if it is less than, mark the feature point as a key point, otherwise, do not mark; In the eleventh step, calculate the gradient magnitude of each key point of the image in the Gaussian difference scale space according to the following formula; $\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ Among them, m(x, y) represents the gradient amplitude of the key point, x and y represent the horizontal and vertical coordinate values corresponding to the key point respectively, and L represents the scale of the key point in the scale space; In the twelfth step, calculate the gradient direction of each key point of the image in the Gaussian difference scale space according to the following formula; θ(x,y)=αtan2{[L(x,y+1)-L(x,y-1)]/[L(x+1,y)-L(x-1,y)]} Among them, θ(x, y) represents the gradient direction of the key point, x and y represent the horizontal and vertical coordinate values corresponding to the key point respectively, α represents the gradient value of the key point in the horizontal coordinate direction, and L represents the key point in the scale space scale; The thirteenth step is to count the gradient direction and magnitude of 8×8 pixels around each key point in the Gaussian difference scale space to obtain a gradient histogram, where the horizontal axis of the gradient histogram is the gradient direction angle, and the vertical axis is The magnitude corresponding to the gradient direction angle; The fourteenth step is to rotate the coordinate axis of the Gaussian difference scale space to the direction of the key point, calculate the 8-dimensional vector representation of each key point sub-region, combine the 8-dimensional vectors of all key point sub-regions, and obtain the value of each key point in 128-dimensional SIFT features. 4. the scene classification method based on nonparametric space decision hidden Dirichlet model according to claim 1, is characterized in that, the step of the described model initialization of step (3) is as follows: The first step is to find out whether there is scene element space layout information in the training image, if it exists, go to the second step, otherwise, go to the third step; In the second step, the spatial layout information of the scene elements of the training image is used as the spatial distribution parameter of the scene elements; In the third step, the training image is evenly divided into multiple 8×8 image blocks, and the spatial distribution parameters of scene elements are obtained by the method of statistical image blocks. 5. the scene classification method based on non-parametric space judgment hidden Dirichlet model according to claim 1, is characterized in that, the step of setting up non-parametric space judgment hidden Dirichlet model described in step (4) is as follows: In the first step, the probability distribution of image topics is estimated according to the distribution parameters of the scene element space; In the second step, the probability distribution of the image theme is randomly sampled to obtain sampling samples of the image block theme; The third step is to estimate the word distribution parameters of the image block topic according to the sampling samples of the image block topic; The fourth step is to establish a non-parametric spatial decision hidden Dirichlet model according to the word distribution parameters of the image block topic. 6. the scene classification method based on non-parametric space judgment hidden Dirichlet model according to claim 1, is characterized in that, the step of the prediction test image classification mark described in step (5) is as follows: In the first step, the test image is brought into the non-parametric space decision hidden Dirichlet model to obtain the test image category probability distribution; In the second step, the category with the highest probability in the category probability distribution of the test image is used as the category label of the test image.

Images