CN111325268A - Image classification method and device based on multi-level feature representation and integrated learning - Google Patents

Abstract

The invention discloses an image classification method and device based on multi-level feature representation and integrated learning. The method includes the following steps: Step 1: acquiring multiple images and preprocessing them; image, extract the feature vectors of three different levels of points, edges and networks respectively; Step 3: Perform feature selection on the feature vectors of the three different levels respectively, and obtain the optimal feature vectors of the three different levels; Step 4: Construct the Three weak classifiers based on single-level features are trained using the optimal feature vectors of the three levels of the image sample respectively, and three trained weak classifiers are obtained; Step 5: The optimal features of the three levels of the image to be tested are used for training. The vectors are respectively input to the corresponding three trained weak classifiers to obtain three classification results of the image to be tested; the classification results of the three weak classifiers are fused to obtain the final classification label of the image to be tested. The present invention has high classification accuracy.

Description

Image classification method and device based on multi-level feature representation and ensemble learning

technical field

The invention specifically relates to an image classification method and device based on multi-level feature representation and integrated learning.

Background technique

In recent years, the extensive application of neuroimaging technology in neurological research has accelerated the development of neuroimaging and pushed brain science research into a period of rapid development. The development of magnetic resonance imaging technology provides people with a non-invasive way to study the anatomical structure and functional mechanism of the human brain. Several commonly used brain imaging techniques include structural magnetic resonance imaging (sMRI), functional magnetic resonance imaging Imaging (functionalMagnetic Resonance Imaging, fMRI), diffusion tensor imaging (Diffusion Tensor Imaging, DTI) and positron emission tomography (Positron Emission Tomography, PET) and so on. Different neuroimaging technologies provide different perspectives on different levels of the brain, from the overall structure of the brain to the functional cooperation between brain regions, and even the activity state of neurons, providing a data basis for various brain research.

However, the current methods for image classification based on the features of brain imaging data usually only use the features of a single level of brain imaging data, without considering the complementary advantages of multi-level feature information to improve the accuracy of classification. Therefore, it is necessary to provide a method and device that can make full use of the information complementary advantages of multi-level features to accurately classify images.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an image classification method and device based on multi-level feature representation and integrated learning to improve the accuracy of image classification in view of the shortcomings of the prior art.

For achieving the above object, technical scheme of the present invention is as follows:

In one aspect, an image classification method based on multi-level feature representation and ensemble learning is provided, including the following steps:

Step 1: Acquire multiple images and preprocess them;

Step 2: According to the preprocessed image, extract the feature vectors of three different levels: point, edge and network:

Divide the preprocessed image into multiple regions of interest (ROI);

Extract the functional activation spectral features of different angles of each region of interest on the preprocessed image, and cascade the functional activation spectral features of each region of interest into a feature vector, which is used as the point feature vector RVF of the image. ;

Based on the correlation between each region of interest on the preprocessed image, the whole-brain functional connectivity feature is extracted as the edge feature vector FC of the image;

Based on the correlation between the regions of interest on the preprocessed images, a functional connection network is constructed. Based on the functional connection network, various network attribute features of each region of interest are extracted from different angles. The attributes are concatenated into a feature vector as the network feature vector NET of the image;

Step 3: Perform feature selection on the feature vectors of three different levels of point, edge and network respectively, and obtain the optimal feature vector of different levels;

Step 4: Build three weak classifiers based on single-level features, and use the optimal feature vectors of the three levels of image samples for training respectively, and obtain three trained weak classifiers;

Step 5: Use the weak classifier ensemble learning method based on single-level features to maximize the information complementary advantages of multi-level features, specifically: input the optimal feature vectors of the three levels of the image to be tested into the trained corresponding three Weak classifier is used to obtain three classification results of the image to be tested; and then the classification results of the three weak classifiers are fused through a classifier integration strategy to obtain the final classification label of the image to be tested.

Further, in the step 1, the DPABI tool is used to perform preprocessing operations such as temporal layer correction, head motion correction, spatial normalization, smoothing, filtering, etc. on the image data, to obtain preprocessed image data.

Further, in the step 2, according to the brain network group (Brainnetome, BN246) atlas, the preprocessed image is divided into 246 regions of interest, and then the feature vectors are extracted based on these regions of interest.

Further, in the step 2, the functional activation spectral features of the different angles include the three functional activation spectral features of ReHo, VMHC and fALFF, and the point feature vector RVF of the extracted image specifically includes the following steps:

First, calculate the three functional activation spectral features of the whole brain voxels on the preprocessed image, namely ReHo value, VMHC value and fALFF value, to form three functional activation spectral maps. These three functional activation spectral features can be viewed from different angles. Reflect the local area activity of image data;

Then, the average ReHo value, VMHC value and fALFF value of all voxels in each region of interest on the image are calculated as the features (ReHo feature, VMHC feature and fALFF feature) of the quantified region of interest, and ReHo with dimension D is obtained. Feature vector, VMHC feature vector and fALFF feature vector; these three function activation spectrum feature vectors are concatenated into a 3D feature vector, which is used as the point feature vector RVF of the image, where D is the number of regions of interest on the image. .

Further, in the step 2, extracting the edge feature vector FC of the image specifically includes the following steps:

First, extract the average time series of each region of interest on the image (the average of the time series of all voxels in each region of interest), calculate the Pearson correlation coefficient between the region of interest and the average time series, and then construct D×D functional connection matrix, each element in the matrix is the connection strength between the regions of interest, the connection strength between the regions of interest is defined as an edge, and the upper triangular matrix elements of the matrix are flattened into a vector, as The edge feature vector FC of the image. The formula for calculating the Pearson correlation coefficient between the region of interest and the average time series is as follows:

和

分别表示感兴趣区域i和感兴趣区域j的平均时间序列均值；n为影像的扫描时间点数；则x_it和x_jt分别为在第t个时间点处的感兴趣区域i和感兴趣区域j的平均时间序列值。where r _ij represents the Pearson correlation coefficient of the average time series of the region of interest i and the region of interest j; x _i and x _j are the average time series of the region of interest i and the region of interest j, respectively;

and

Represents the average time series mean of the region of interest i and region of interest j respectively; n is the number of scanning time points of the image; then x _it and x _jt are the region of interest i and the region of interest j at the t-th time point, respectively The average time series value of .

Further, in the described step 2, the network feature vector NET of the extraction image specifically includes the following steps:

Firstly, the region of interest on the image is defined as a node in its functional connection network, and the correlation between regions of interest is defined as an edge in its functional connection network;

Then, the graph theory method is used to calculate the network attribute characteristics of each node from five different angles in the functional connection network of the image, namely between centrality (Betweenness centrality, BC), node local efficiency (Nodal local efficiency, Eloc), node clustering coefficient ( Nodal clustering coefficient, NCC), participant coefficient (Participant coefficient, PC) and degree (Degree, DG). These network attributes can characterize the information interaction capability of network nodes from different perspectives.

According to the D regions of interest on the image, that is, their functions connect D nodes in the network, the BC feature vector, Eloc feature vector, NCC feature vector, PC feature vector and DG feature vector with dimension D are obtained. The vectors are concatenated into a feature vector with a dimension of 5D, which is used as the network feature vector NET of the image.

Further, in the step 3, the Least absolute shrinkage and selection operator (Lasso) method is used to perform feature selection on the feature vectors of three different levels respectively,

The objective function of Lasso method for feature selection is as follows:

Where X represents the feature matrix of N*P, N is the number of samples, P is the feature vector dimension, Y represents the classification label vector corresponding to the sample, Y=(y ₁ , y ₂ ,...,y _N ) ^T , y _i represents the first The classification labels of i image samples; α represents the weight coefficient vector of P*1, and the non-zero element position in α indicates that the feature at this position will be selected (that is, if the element of a dimension in α is not 0, then the feature vector The elements of the corresponding dimension in α remain unchanged. If the element of a certain dimension in α is 0, the element of the corresponding dimension in the feature vector is set to 0, or the element of the dimension is deleted, thereby obtaining the corresponding optimal feature vector); λ ₁ is a regularization parameter for model sparsity.

Further, in the step 4, the weak classifier is an SVM classifier.

Further, in the step 5, the classifier integration strategy is defined as:

Among them, y represents the classification label of the image to be tested, and f _m (x) represents the classification label of the image to be tested obtained by the m-th weak classifier based on the optimal feature vector of the m-th level of the image to be tested x is predicted to be 1 The probability value of , m=1, 2, 3; _um is the weight of the predicted probability of the optimal feature vector of the mth level, and the sum of all weights is 1.

On the other hand, an image classification device based on multi-level feature representation and ensemble learning is provided, including the following modules:

Data preprocessing module for acquiring images and preprocessing them;

The feature extraction module is used to extract three different levels of feature vectors of points, edges and networks according to the preprocessed images:

The feature selection module is used to perform feature selection on the feature vectors of three different levels of point, edge and network respectively, and obtain the optimal feature vector of different levels;

The weak classifier construction and training module is used to construct three weak classifiers based on single-level features, and use the optimal feature vectors of the three levels of image samples for training, and obtain three trained weak classifiers;

The classification module is used to input the optimal feature vectors of the three levels of the image to be tested into the corresponding three trained weak classifiers to obtain three classification results of the image to be tested; and then fuse them through a classifier integration strategy Three weak classifier classification results to obtain the final classification label of the image to be tested;

The apparatus adopts the above-mentioned method to realize image classification.

In another aspect, an electronic device is provided, comprising a processor and a memory, the memory having a computer program stored thereon, when the computer program is executed by the processor, the processor is made to implement any one of claims 1 to 7 the method described.

In another aspect, there is provided a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, causes the processor to implement the method as claimed in any one of claims 1 to 7.

Beneficial effects:

The above technical solution of the present invention fully exploits the complementary information between these features by extracting features from different angles from different levels, and adopts the integrated learning method of weak classifiers based on single-level features to maximize the complementary advantages of feature information at different levels to further improve the model. Classification performance. Specifically, the present invention first acquires rs-fMRI image data and performs a series of preprocessing operations, and extracts features at different levels and angles from the preprocessed rs-fMRI data, including point feature vector RVF, edge feature vector FC and network Eigenvectors NET. Then, the Lasso method is used to select features at different levels, and multiple optimal subsets of different levels of features are obtained, and an ensemble learning method of weak classifiers based on single-level features is proposed. This method is mainly based on points, edges and The weak classifier of the network feature vector is used to predict, and then the predicted probability of the weak classifier based on single-level features is fused through a classifier integration strategy to determine the final sample classification label. The multi-level and multi-angle feature representation proposed by the above technical solutions of the present invention can describe the structure and functional characteristics of image data from different levels, and improve the classification accuracy through the integrated learning method of weak classifiers based on single-level features.

The present invention can significantly improve the accuracy of image classification.

Description of drawings

FIG. 1 is a flowchart of an image classification method based on multi-level feature representation and ensemble learning in an embodiment of the present invention.

Detailed ways

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

The present invention can be applied to the classification of all image data including brain image time series and from which regions of interest (brain regions) can be extracted, that is, it can classify image data with higher temporal resolution and spatial resolution (including rs -fMRI (resting-state fMRI), ts-fMRI (task-state fMRI), etc.), and from these image data, multi-level and multi-angle brain image features can be simultaneously extracted.

Example 1:

The image data in this embodiment is rs-fMRI data.

Resting-state fMRI (rs-fMRI) technology studies the state of brain activity according to the blood oxygen level-dependent (Blood Oxygenation-Level Dependent, BOLD) signal changes in brain tissue, because it has higher relative to other imaging technologies. The advantages of high spatial and temporal resolution, accurate localization of activated brain regions, and non-invasive non-invasive measurement methods make it widely used in various bioinformatics research fields, and by exploring the characteristics of Rs-fMRI for the study of neuropsychiatric diseases. brings a new perspective.

To study rs-fMRI features, features from different perspectives can be extracted from different levels for classification research. At the point level, the functional activation spectrum that reflects the activity of local regions of the brain based on the fMRI BOLD signal can be extracted, including regional homogeneity (ReHo), ratio of low frequency fluctuations (fALFF) and body Prime mirror homotopy symmetry (Voxel-Mirrored Homotopic Connectivity, VMHC). These functional activation spectra are calculated based on voxel analysis. In order to extract different functional activation spectral features in the region of interest, the average of all voxel functional activation spectral features in these regions is calculated to quantify the region of interest. Further, in order to study the functional cooperative relationship between regions of interest in the brain, extract the time series of the region of interest and calculate the correlation between the time series of the regions, construct a brain functional connectivity matrix, and extract the upper triangular element of the matrix as an edge. Characteristics. A functional connection network is constructed on the basis of functional connection, and some network attributes are further calculated from the network level through graph theory. Common network attributes include degree, betweenness centrality, node clustering coefficient (Nodal clustering). coefficient), participant coefficient (Participation coefficient), node local efficiency (Nodal local efficient), small worldness (Smallworldness), etc. These network attributes can represent the internal functional changes of the network from different perspectives, which is conducive to the study of brain regions of interest. Functional information interaction. The above multi-angle features extracted from different levels are both independent and related to each other, and they can describe the characteristics of fMRI data from different levels. In this embodiment, by extracting features from different angles from different levels, the complementary information between these features is fully exploited, and the ensemble learning method of weak classifiers based on single-level features is used to maximize the complementary advantages of feature information at different levels to further improve the classification performance of the model. .

Referring to FIG. 1 , the image classification method based on multi-level feature representation and ensemble learning provided in this embodiment specifically includes the following steps:

Step 1: Acquire multiple rs-fMRI image data, and use the Data Processing & Analysis for Brain Imaging (DPABI) tool to complete a series of preprocessing processes on the image data, including deleting the first ten frames of images, temporal layer correction, head motion correction, spatial standardization, Smoothing and filtering operations to obtain preprocessed image data.

Step 2: According to the preprocessed image, extract the feature vectors of three different levels: point, edge and network:

(1) Extract the point feature vector of the image

According to the preprocessed image data, the three functional activation spectrum features of ReHo, VMHC and fALFF are extracted from different angles based on the region of interest. Specifically, the ReHo value, VMHC value and fALFF value of the whole brain voxels on the preprocessed image are calculated. Three functional activation spectra are formed, and these three functional activation spectral features can reflect the local area activities of image data from different angles. Then, the brain network group (Brainnetome, BN246) map was used to divide the three functional activation spectra into 246 regions of interest, and the average ReHo value, VMHC value and fALFF value of all voxels under each region of interest were calculated as the quantification of the interest. The features of the region are obtained to obtain ReHo feature vector, VMHC feature vector and fALFF feature vector with dimensions of 246 respectively. These three functional activation spectral feature vectors are concatenated into a point feature vector of dimension 738, named RVF.

The ReHo value for any voxel is usually obtained by calculating the Kendall's Coefficient Concordance (KCC) of the time series of the voxel. The calculation formula is as follows:

表示R_i的平均。KCC的取值范围在0到1之间，度量了相邻体素时间序列的一致性与相似性，当某个脑部区域的ReHo值越接近1时，表示该大脑局部区域的时间序列一致性越高，反映了该局部区域的脑部神经元的活动具有很高的同步性，反之亦然。where K represents the total number of specific voxels and the nearest neighbor voxels, usually 27, that is, 27 voxels form a voxel cluster; n is the number of time points; R _i represents the rank sum of the i-th time point;

represents the average of R _i . The value range of KCC is between 0 and 1, which measures the consistency and similarity of the time series of adjacent voxels. When the ReHo value of a certain brain area is closer to 1, it means that the time series of the local area of the brain are consistent The higher the sex, the higher the synchronization of the activity of the brain neurons in the local area, and vice versa.

For the calculation of the VMHC value of any voxel, that is, by calculating the Pearson correlation coefficient between a voxel and the corresponding voxel time series in the symmetric hemisphere of the brain, the calculation formula is as follows:

where r _ij represents the Pearson correlation between the time series of two symmetrical voxels i, j in the cerebral hemisphere; _xi and x _j represent the time series of two symmetrical voxels i, j, respectively. cov(x _i ,x _j ) is the covariance of voxel i,j time series;

分别表示体素i，j时间序列的标准差。

are the standard deviations of the time series of voxels i and j, respectively.

The formula for calculating the fALFF value for any voxel is as follows:

where f _k represents the frequency, N represents the number of voxels, and a _k and b _k represent the coefficients corresponding to different frequencies, respectively. The increase or decrease of the fALFF value is consistent with the blood oxygen level-dependent signal strength. When the fALFF value in a certain area of the brain increases, it indicates that the neuronal activity in that area is increased, and vice versa.

(2) Extract the edge feature vector of the image

The whole brain functional connectivity feature is extracted based on the correlation between the regions of interest. Specifically, the preprocessed image data is firstly divided into 246 regions of interest (brain regions) by using the BN246 atlas, and then the average value of each region of interest is extracted. Time series, calculate the Pearson correlation coefficient between the region of interest and the average time series, and finally construct a 246*246 functional connection matrix. Each element in the matrix is the Pearson correlation between a region of interest and the average time series. Correlation coefficient, that is, the connection strength between two regions of interest, defines the connection strength between regions of interest as an edge, and flattens the upper triangular matrix elements of the matrix as the edge feature vector of each image, named as fc. The formula for calculating the Pearson correlation coefficient between the region of interest and the average time series is as follows:

和

分别表示感兴趣区域i和感兴趣区域j的平均时间序列均值；n为rs-fMRI影像数据的扫描时间点数；则x_it和x_jt分别为在某个时间点处的感兴趣区域i和感兴趣区域j的平均时间序列值。where r _ij represents the Pearson correlation coefficient between the average time series of the region of interest i and the region of interest j; x _i and x _j are the average time series of the region of interest i and the region of interest j, respectively;

and

Represents the average time series mean value of region of interest i and region of interest j, respectively; n is the number of scanning time points of rs-fMRI image data; then x _it and x _jt are the region of interest i and the sense of interest at a certain time point, respectively. Average time series value for region of interest j.

(3) Extract the network feature vector of the image

The functional connection network is constructed based on the correlation between the regions of interest. Specifically, the region of interest is first defined as the node of the network, the correlation between the regions of interest is defined as the edge in the network, and the betweenness centrality is calculated by using the graph theory method. (Betweeness centrality, BC), node local efficiency (Nodal local efficiency, Eloc), node clustering coefficient (Nodal clustering coefficient, NCC), participant coefficient (Participant coefficient, PC) and degree (Degree, DG) five different angles characteristics of the network properties. Betweenness centrality is a measure of the importance of a node's transmission in the network. It is defined as the ratio of the number of paths passing through the node to the total number of paths in the shortest path between any two nodes in the network. The calculation formula is:

where BC _i represents the betweenness centrality of node i; n _mj represents the number of shortest paths between node m and node j, and n _mj (i) represents the number of nodes passing through node i in the shortest path between node m and node j.

The local efficiency of a node is used to measure the information transmission capability of the network in the local area, and it is related to the shortest path of the node. . The calculation formula of node local efficiency is as follows:

为子网络G_i的节点数量；l_jh为子网络中节点j与节点h的最短路径长度。where E _loc (i) represents the local efficiency of node i; G _i is the sub-network formed by node i and its neighbors;

is the number of nodes in the sub-network _Gi ; l _jh is the shortest path length between node j and node h in the sub-network.

The node clustering coefficient is used to measure the possibility that any two neighboring nodes of a given node are also neighbors, and it is one of the metrics of network function segmentation. Its calculation formula is as follows:

where C _i represents the clustering coefficient of node i; K _i is the degree of node i; s _i represents the weight sum of edges between all neighbor nodes of node i.

Participant coefficient is a measure of the diversity of nodes in the network modular connection under the network modularity. A node with a low participation coefficient but a high degree indicates that the node plays an important role in the module division. The calculation formula is:

where PC _i represents the participant coefficient of node i; M represents the module set divided by the network; _ki (m) represents the sum of the connection weights between node i and all nodes in module m; _ki represents the degree of node i.

Degree is one of the indicators used to measure the centrality of a node in the network. The higher the degree of a node, the stronger the importance of the node in the network. Its calculation formula is as follows:

where K _i is the degree of node i; N is the number of nodes in the network; a _ij =1 means that node i and node j are connected by an edge, otherwise a _ij =0.

The five network attributes of each node in the functional connection network are calculated according to the above formula, and these network attributes can characterize the information interaction ability of network nodes from different angles. According to the 246 regions of interest in the network, BC eigenvectors, Eloc eigenvectors, NCC eigenvectors, PC eigenvectors, and DG eigenvectors with a dimension of 246 can be calculated for each image respectively, and these 5 eigenvectors can be cascaded. into a network feature vector with dimension 1230, named NET.

Step 3: Use the Lasso method to perform feature selection on the three feature vectors extracted from different levels: RVF, FC and NET. The objective function of the Lasso feature selection method is as follows:

Where X represents the feature matrix of N*P, N is the number of samples, P is the feature vector dimension, Y represents the classification label vector corresponding to the sample, Y=(y ₁ , y ₂ ,...,y _N ) ^T , y _i represents the first The classification labels of i image samples; α represents the weight coefficient vector of P*1, and the non-zero element position in α indicates that the feature at this position will be selected (that is, if the element of a dimension in α is not 0, then the feature vector The elements of the corresponding dimension in α remain unchanged. If the element of a certain dimension in α is 0, the element of the corresponding dimension in the feature vector is set to 0, or the element of the dimension is deleted, thereby obtaining the corresponding optimal feature vector); λ ₁ is a regularization parameter for model sparseness and is an empirical parameter.

Step 4: First, use the optimal feature vectors of the three levels obtained by the image samples through the feature selection in Step 3 to train an SVM classifier respectively, which is used for prediction by using single-level features;

Step 5: A weak classifier ensemble learning method based on single-level features is used to maximize the information complementary advantages of multi-level features. Specifically, a classifier ensemble strategy is used to fuse multiple SVM classifiers based on single-level features to predict the probability , which determines the final classification label of the image to be tested. The classifier integration strategy is defined as:

Among them, y represents the classification label of the image to be tested, and f _m (x) represents the classification label of the image to be tested obtained by the m-th weak classifier based on the optimal feature vector of the m-th level of the image to be tested x is predicted to be 1 The probability value of , m=1, 2, 3; _um is the weight of the predicted probability of the optimal feature vector of the mth level, which is an empirical parameter, and the sum of all weights is 1.

Example 2:

This embodiment discloses an image classification device based on multi-level feature representation and integrated learning, including the following modules:

Data preprocessing module for acquiring images and preprocessing them;

The feature extraction module is used to extract three different levels of feature vectors of points, edges and networks according to the preprocessed images:

The feature selection module is used to perform feature selection on the feature vectors of three different levels of point, edge and network respectively, and obtain the optimal feature vector of different levels;

The weak classifier construction and training module is used to construct three weak classifiers based on single-level features, and use the optimal feature vectors of the three levels of image samples for training, and obtain three trained weak classifiers;

The classification module is used to input the optimal feature vectors of the three levels of the image to be tested into the corresponding three trained weak classifiers to obtain three classification results of the image to be tested; and then fuse them through a classifier integration strategy Three weak classifier classification results to obtain the final classification label of the image to be tested;

The apparatus adopts the method described in Embodiment 1 to realize image classification.

Example 3:

This embodiment discloses an electronic device, including a processor and a memory, where a computer program is stored in the memory, and when the computer program is executed by the processor, the processor is made to implement the method described in Embodiment 1.

Example 4:

This embodiment discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the processor is made to implement the method described in Embodiment 1.

The scheme described in the above embodiment simultaneously extracts features from different levels and angles of rs-fMRI image data. Considering that these different types of features can describe the characteristics of rs-fMRI image data from different levels, the ensemble learning based on weak classifiers based on single-level features The method maximizes the potential complementary information between these different levels of features, and further improves the classification performance of the model. The Lasso method is used to select the features at different levels, and the optimal subsets of the different levels of features are obtained, and an ensemble learning method of weak classifiers based on single-level features is proposed to maximize the use of the optimal subsets of these different levels of features. The complementary advantages of information can greatly improve the classification performance.

Claims (10)

1. An image classification method based on multi-level feature representation and integrated learning is characterized by comprising the following steps:

step 1: acquiring a plurality of images and preprocessing the images;

step 2: respectively extracting feature vectors of three different levels of points, edges and networks of the preprocessed images;

and step 3: respectively selecting features of the feature vectors of three different levels, namely, a point, an edge and a network to obtain optimal feature vectors of different levels;

and 4, step 4: constructing three weak classifiers based on single-level features, and training by respectively utilizing optimal feature vectors of three levels of an image sample to obtain three trained weak classifiers;

and 5: firstly, respectively inputting the optimal feature vectors of three levels of the image to be detected into the corresponding trained three weak classifiers to obtain three classification results of the image to be detected; and then, fusing the classification results of the three weak classifiers through a classifier integration strategy to obtain a final classification label of the image to be detected.

2. The method for classifying images based on multi-level feature representation and integrated learning of claim 1, wherein in the step 2, the preprocessed image is divided into 246 regions of interest according to the brain network group atlas, and then the feature vectors of the regions of interest are extracted based on the regions of interest.

3. The method for classifying images based on multi-level feature representation and ensemble learning of claim 1, wherein the step 2 of extracting the point feature vector of the image specifically comprises the following steps:

firstly, calculating function activation spectrum characteristics of a whole brain voxel at three different angles on a preprocessed image, namely a ReHo value, a VMHC value and a fALFF value, wherein the three function activation spectrum characteristics can reflect local region activities of image data from different angles;

then, calculating the average ReHo value, VMHC value and fALFF value of all voxels in each region of interest on the image as the features of the region of interest to obtain a ReHo feature vector, a VMHC feature vector and an fALFF feature vector with the dimensions of D respectively; and cascading the three feature vectors into a feature vector of one dimension 3D, wherein D is the number of interested areas on the image and is used as a point feature vector of the image.

4. The method of claim 1, wherein the step 2 of extracting edge feature vectors of the image comprises the steps of first extracting an average time sequence of each region of interest on the image, calculating a Pearson correlation coefficient between the region of interest and the average time sequence, and then constructing a functional connection matrix D × D, wherein each element in the matrix is the Pearson correlation coefficient between one region of interest and the average time sequence, and flattening an upper triangular matrix element of the matrix into a vector as the edge feature vector of the image, wherein D is the number of the regions of interest on the image.

5. The method for classifying images based on multi-level feature representation and integrated learning of claim 1, wherein the step 2 of extracting the network feature vector of the image specifically comprises the following steps:

firstly, defining an interested area on an image as a node in a functional connection network, and defining the correlation between the interested areas as an edge in the functional connection network;

then, calculating network attribute characteristics of each node at five different angles in the functional connection network of the image by using a graph theory method, namely medium centrality BC, node local efficiency Eloc, node clustering coefficient NCC, participant coefficient PC and degree DG; according to the D interested areas on the image, namely the D nodes in the network are functionally connected, a BC eigenvector, an Eloc eigenvector, an NCC eigenvector, a PC eigenvector and a DG eigenvector with the dimension of D are obtained, and the 5 eigenvectors are cascaded into one eigenvector with the dimension of 5D and used as the network eigenvector of the image.

6. The method for classifying images based on multi-level feature representation and integrated learning of claim 1, wherein in the step 3, feature selection is performed on feature vectors of three different levels by using a Lasso method;

the objective function for feature selection by the Lasso method is as follows:

wherein, X represents a feature matrix of N × P, N is the number of image samples, P is the dimension of the feature vector, Y represents the classification label vector corresponding to the sample, and Y is (Y ═ P)₁,y₂,…,y_N)^T，y_iClass label representing the ith image sample, α representing a weight coefficient vector of P1, the position of the element α other than 0 indicating that the feature at that position is to be selected, and λ₁Is a regularization parameter for model sparsity.

7. The method for classifying images based on multi-level feature representation and ensemble learning of claim 1, wherein in the step 5, the classifier ensemble policy is defined as:

wherein y represents the classification label of the image to be measured, f_m(x) The probability value of the classification label of the image to be detected, which is obtained by predicting the m-th weak classifier based on the optimal feature vector of the m-th level of the image to be detected x, is 1,2 and 3; u. of_mIs the weight of the prediction probability of the optimal feature vector of the mth level.

8. The apparatus for classifying images based on multi-level feature representation and integrated learning of claim 1, comprising:

the data preprocessing module is used for acquiring and preprocessing images;

the feature extraction module is used for respectively extracting feature vectors of points, edges and networks of the images according to the preprocessed images:

the characteristic selection module is used for respectively selecting characteristics of the characteristic vectors of three different levels, namely points, edges and networks to obtain optimal characteristic vectors of different levels;

the weak classifier building and training module is used for building three weak classifiers based on single-level features, and training the weak classifiers by respectively utilizing the optimal feature vectors of three levels of the image sample to obtain three trained weak classifiers;

the classification module is used for respectively inputting the optimal feature vectors of the three layers of the image to be detected into the corresponding trained three weak classifiers to obtain three classification results of the image to be detected; then, fusing the classification results of the three weak classifiers through a classifier integration strategy to obtain a final classification label of the image to be detected;

the device realizes image classification by adopting the method of any one of claims 1 to 7.

9. An electronic device, comprising a processor and a memory, the memory having stored thereon a computer program which, when executed by the processor, causes the processor to carry out the method of any one of claims 1 to 7.

10. A computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, causes the processor to carry out the method according to any one of claims 1 to 7.

Images