CN107133640A - Image classification method based on topography's block description and Fei Sheer vectors - Google Patents

本发明公开了基于局部图像块描述子和费舍尔向量的图像分类方法，步骤依次为，构建基于局部图像块的特征描述子；利用高斯混合模型对训练数据集的特征描述子进行建模；生成训练图像集和测试图像集的费舍尔向量；构建高斯空间金字塔，获取图像多尺度空间信息；计算训练数据集和测试数据集的特征集合；采用互信息方法对数据集的特征集合进行特征选择；训练分类器，实现图像分类。本发明能够精确地获取图像信息，提高了图像的分类准确率，可用于大规模图像分类与检索系统的构建。

The invention discloses an image classification method based on a local image block descriptor and a Fisher vector. The steps are as follows: constructing a feature descriptor based on a local image block; using a Gaussian mixture model to model the feature descriptor of a training data set; Generate the Fisher vectors of the training image set and the test image set; construct a Gaussian space pyramid to obtain the multi-scale spatial information of the image; calculate the feature set of the training data set and the test data set; use the mutual information method to characterize the feature set of the data set Select; train the classifier to achieve image classification. The invention can accurately acquire image information, improves the classification accuracy of images, and can be used in the construction of a large-scale image classification and retrieval system.

基于局部图像块描述子和费舍尔向量的图像分类方法Image Classification Method Based on Local Image Block Descriptor and Fisher Vector

技术领域technical field

本发明属于机器学习、计算机视觉技术领域，涉及了一种图像分类方法。The invention belongs to the technical fields of machine learning and computer vision, and relates to an image classification method.

背景技术Background technique

随着多媒体技术的发展，图像分类已成为计算机视觉领域研究的重点，图像分类是依据图像具有的某种属性而将其划分到预先设定的不同类别中，如何将图像进行有效的表达是提高图像分类准确率的关键，特征的选择与提取问题是图像分类目前存在的难点问题。随着移动互联网的迅速发展，人类社会已进入大数据时代。SIFT、HOG等这些传统的特征学习虽能提取图像的某些特征，在图像分类中也取得了较好的效果，但这种人工设计特征方法存在一定的缺陷。近年来，深度学习模型尤其是卷积神经网络(CNN)在特征提取中取得突破性进展，但是CNN需要调整大量的参数和耗费巨大的计算成本。目前计算成本已经成为对象检测/图像分类的一个核心问题。With the development of multimedia technology, image classification has become the focus of research in the field of computer vision. Image classification is to divide images into different preset categories according to certain attributes of images. How to effectively express images is the key to improving the The key to the accuracy of image classification, the selection and extraction of features are currently difficult problems in image classification. With the rapid development of the mobile Internet, human society has entered the era of big data. Although traditional feature learning methods such as SIFT and HOG can extract some features of images, they have also achieved good results in image classification, but this artificially designed feature method has certain defects. In recent years, deep learning models, especially convolutional neural networks (CNNs), have made breakthroughs in feature extraction, but CNNs need to adjust a large number of parameters and consume huge computational costs. Computational cost has become a core issue in object detection/image classification.

聚合局部特征已被广泛应用于2D图像的分类或检索任务中。许多特征聚合算法已被提出，视觉词袋模型(Bag-of-Visual-Words,BOW)是广泛使用的模型之一。简而言之，该模型从每幅图像提出一组局部特征，然后将这些局部特征量化为离散的视觉单词，最后用一个紧凑的直方图表示图像。虽然BOW模型具有较好的性能，但是该模型受限于以下两点：1、没有考虑位置因素；2、基于特征的0阶统计。Aggregating local features has been widely used in classification or retrieval tasks of 2D images. Many feature aggregation algorithms have been proposed, and the Bag-of-Visual-Words (BOW) model is one of the widely used models. In brief, the model proposes a set of local features from each image, then quantizes these local features into discrete visual words, and finally represents the image with a compact histogram. Although the BOW model has better performance, the model is limited by the following two points: 1. The location factor is not considered; 2. The zero-order statistics based on features.

发明内容Contents of the invention

为了解决上述背景技术提出的技术问题，本发明旨在提供基于局部图像块描述子和费舍尔向量的图像分类方法，克服现有图像分类方法存在的缺陷，减少计算成本，提高分类精度。In order to solve the technical problems raised by the above-mentioned background technology, the present invention aims to provide an image classification method based on local image block descriptors and Fisher vectors, which overcomes the defects of existing image classification methods, reduces computing costs, and improves classification accuracy.

为了实现上述技术目的，本发明的技术方案为：In order to realize above-mentioned technical purpose, technical scheme of the present invention is:

基于局部图像块描述子和费舍尔向量的图像分类方法，包括以下步骤：An image classification method based on local image block descriptors and Fisher vectors, comprising the following steps:

(1)选定训练数据集和测试数据集，分别构建两个数据集的基于局部图像块的特征描述子；(1) Select the training data set and the test data set, and construct the feature descriptors based on local image blocks of the two data sets respectively;

(2)利用高斯混合模型对训练数据集的特征描述子进行聚类，获取高斯混合模型参数；(2) clustering the feature descriptors of the training data set using the Gaussian mixture model to obtain the parameters of the Gaussian mixture model;

(3)生成训练数据集和测试数据集中每幅图像的费舍尔向量；(3) Generate the Fisher vector of each image in the training data set and the test data set;

(4)对训练数据集和测试数据集中每幅图像构建高斯空间金字塔；(4) Construct a Gaussian spatial pyramid for each image in the training data set and the testing data set;

(5)计算数据训练集和测试数据集中每幅图像的高斯空间金字塔中每幅图像的费舍尔向量，从而构成数据训练集和测试数据集的特征集合；(5) Calculate the Fisher vector of each image in the Gaussian space pyramid of each image in the data training set and the test data set, thereby forming the feature set of the data training set and the test data set;

(6)采用互信息方法对训练数据集、测试数据集的特征集合进行特征选择；(6) Using the mutual information method to perform feature selection on the feature set of the training data set and the test data set;

(7)采用选择后的训练数据集的特征值来训练分类器，将测试数据集的特征值输入训练好的分类器中，实现图像分类。(7) Use the eigenvalues of the selected training data set to train the classifier, and input the eigenvalues of the test data set into the trained classifier to realize image classification.

进一步地，步骤(1)的具体步骤如下：Further, the specific steps of step (1) are as follows:

(11)选定训练数据集和测试数据集；(11) selected training data set and test data set;

(12)通过随机采样的方法分别在训练数据集和测试数据集中的每幅图像中提取T个图像块；(12) Extract T image blocks in each image in the training data set and the test data set respectively by random sampling method;

(13)对提取的图像块进行亮度、对比度归一化和白化预处理操作，将其转换为基于像素值的列向量特征描述子，分别构成训练数据集的特征描述子和测试数据集的特征描述子；(13) Perform brightness, contrast normalization and whitening preprocessing operations on the extracted image blocks, and convert them into column vector feature descriptors based on pixel values, which constitute the feature descriptors of the training data set and the characteristics of the test data set respectively descriptor;

(14)对训练数据集的特征描述子和测试数据集的特征描述子进行PCA降维处理。(14) Perform PCA dimensionality reduction processing on the feature descriptors of the training data set and the feature descriptors of the test data set.

进一步地，步骤(2)的具体步骤如下：Further, the specific steps of step (2) are as follows:

(21)将K个高斯单元的高斯混合模型记为其中，p_i(x)代表第i个高斯单元，D表示特征描述子x的维数，π_i,μ_i,∑_i分别表示第i个高斯单元的权重、均值和协方差矩阵；(21) Record the Gaussian mixture model of K Gaussian units as Among them, p _i (x) represents the i-th Gaussian unit, D represents the dimensionality of the feature descriptor x, π _i , μ _i , ∑ _i represent the weight, mean and covariance matrix of the i-th Gaussian unit, respectively;

(22)通过训练数据集特征描述子组成的集合C来估计高斯混合模型的所有参数θ＝{π_i,μ_i,∑_i,i＝1,...K}，构造集合C的概率公式：(22) Estimate all the parameters of the Gaussian mixture model θ={π _i ,μ _i ,∑ _i ,i=1,...K} through the set C composed of the feature descriptors of the training data set, and construct the probability formula of the set C :

其中，S为集合C的特征描述子个数；Among them, S is the number of feature descriptors of the set C;

(23)采用期望值最大算法，根据步骤(22)中的概率公式，对高斯混合模型进行参数估计。(23) Using the maximum expected value algorithm, according to the probability formula in step (22), estimate the parameters of the Gaussian mixture model.

进一步地，步骤(3)的具体步骤如下：Further, the specific steps of step (3) are as follows:

(31)利用高斯混合模型参数，计算训练数据集和测试数据集中每个特征描述子在第i个高斯单元上的梯度向量；(31) Using the parameters of the Gaussian mixture model, calculate the gradient vector of each feature descriptor in the i-th Gaussian unit in the training data set and the test data set;

(32)计算第i个高斯单元的概率密度函数p_i的费舍尔信息矩阵；(32) Calculate the Fisher information matrix of the probability density function p _i of the i Gaussian unit;

(33)对信息矩阵F_i进行柯列斯基分解，得到柯列斯基分量，将此分量与步骤(31)中得到的梯度向量相乘，得到数据集中每幅图像的费舍尔向量F₀，从而构成训练数据集和测试数据集的费舍尔向量集。(33) Carry out Colesky decomposition on the information matrix F _i to obtain the Colesky component, multiply this component with the gradient vector obtained in step (31), and obtain the Fisher vector F of each image in the data set ₀ , thus forming the Fisher vector set of the training data set and the testing data set.

进一步地，步骤(4)的具体步骤如下：Further, the specific steps of step (4) are as follows:

(41)利用高斯卷积函数生成数据集中图像I的尺度空间：L(q,y,σ)＝G(q,y,σ)*I(q,y)，其中，G(q,y,σ)是尺度可变的高斯卷积函数，“*”表示卷积运算，(q,y)是图像I上的空间坐标，σ是尺度坐标，k是平滑系数；通过变换不同的平滑系数k形成多幅图像，将这多幅图像构成图像I的第一组金字塔的若干层图像；(41) Use the Gaussian convolution function to generate the scale space of the image I in the dataset: L(q,y,σ)=G(q,y,σ)*I(q,y), where G(q,y, σ) is a scale-variable Gaussian convolution function, "*" means convolution operation, (q, y) is the spatial coordinate on the image I, σ is the scale coordinate, and k is the smoothing coefficient; multiple images are formed by transforming different smoothing coefficients k, and these multiple images are composed of images Several layer images of the first group of pyramids of I;

(42)将步骤(41)得到的第一组金字塔的倒数第三层图像进行比例因子为2的下采样操作，得到第二组金子塔的第一层图像，通过变换不同的平滑系数k，得到第二组金字塔的若干层图像；(42) carry out the down-sampling operation that scale factor is 2 with the penultimate third layer image of the first group of pyramids that step (41) obtains, obtain the first layer image of the second group of pyramids, by transforming different smoothing coefficients k, Obtain several layer images of the second group of pyramids;

(43)反复执行步骤(41)-(42)，一共得到O组金字塔，每组金字塔包含S层图像，共计O×S幅图像，这些图像构成图像I的高斯空间金字塔。(43) Steps (41)-(42) are repeatedly performed to obtain a total of O groups of pyramids, each group of pyramids contains S layers of images, a total of O×S images, and these images constitute the Gaussian space pyramid of image I.

进一步地，步骤(5)的具体步骤如下：Further, the specific steps of step (5) are as follows:

(51)计算图像I高斯空间金字塔中每幅图像的费舍尔向量F_i，i＝1,2,…,O×S，将O×S个费舍尔向量以及步骤(33)求得的向量F₀进行串联，形成图像I的特征表示F＝[F₀,F₁,...,F_O×S]；(51) Calculate the Fisher vector F _i of each image in the image I Gaussian space pyramid, i=1, 2,..., O × S, with O × S Fisher vectors and step (33) obtained The vector F ₀ is concatenated to form the feature representation of the image I F=[F ₀ , F ₁ ,...,F _O×S ];

(52)将训练数据集和测试数据集中每幅图像进行步骤(51)的操作，得到训练数据集和测试数据集的特征集合。(52) Perform the operation of step (51) on each image in the training data set and the test data set to obtain the feature set of the training data set and the test data set.

进一步地，步骤(6)的具体步骤如下：Further, the specific steps of step (6) are as follows:

(71)将训练数据集的特征集合和测试数据集的特征集合进行转置，设置训练样本标签为l₁，测试样本标签为l₂，l₁和l₂分别为训练样本和测试样本中每幅图像种类的一组数据；(71) Transpose the feature set of the training data set and the feature set of the test data set, set the label of the training sample as l ₁ , and set the label of the test sample as l ₂ , where l ₁ and l ₂ are each A set of data of image type;

(72)建立训练数据集的特征集合中每一维特征与标签l₁的互信息值模型：(72) Establish the mutual information value model of each dimension feature and label _l in the feature set of the training data set:

I(x_:i,l₁)＝H(l₁)+H(x_:i)-H(x_:i,l₁)I(x _:i ,l ₁ )＝H(l ₁ )+H(x _:i )-H(x _:i ,l ₁ )

其中，H(x)为变量x的离散熵值，x_:i为训练数据集的特征集合中第i维特征，则H(x_:i,l₁)表示在已知l₁的情况下，变量x_:i的离散熵值；Among them, H(x) is the discrete entropy value of the variable x, x _:i is the i-th dimension feature in the feature set of the training data set, then H(x _:i ,l ₁ ) means that when l ₁ is known, variable x _{: discrete entropy value of i} ;

(73)使用1-bit量化方法求解离散熵值H(x)：(73) Use 1-bit quantization method to solve discrete entropy value H(x):

其中，p_j表示变量x落入第j个离散箱的概率值，离散箱如下式所示：Among them, p _j represents the probability value of the variable x falling into the jth discrete box, and the discrete box is shown in the following formula:

根据离散熵值和互信息值模型计算训练数据集的特征集合中每一维特征与标签l₁的互信息值；Calculate the mutual information value of each dimension feature and label _l in the feature set of the training data set according to the discrete entropy value and the mutual information value model;

(74)对互信息值按照由大到小的顺序进行排序，选取前d个特征，并记录前d个特征索引值index；(74) Sort the mutual information values in descending order, select the first d features, and record the first d feature index values index;

(75)根据步骤(74)得到的索引值index选取出测试数据集的特征集合中的d个特征。(75) Select d features in the feature set of the test data set according to the index value index obtained in step (74).

采用上述技术方案带来的有益效果：The beneficial effect brought by adopting the above-mentioned technical scheme:

(1)本发明直接采用局部图像块向量构建局部特征描述子，有效地减少了计算成本、降低了计算复杂度；(1) The present invention directly uses local image block vectors to construct local feature descriptors, effectively reducing computational costs and computational complexity;

(2)本发明构建基于图像尺度空间理论的高斯金字塔，尺度空间方法将传统的单尺度图像信息处理技术纳入尺度不断变化的动态分析框架中，更容易获取图像的本质特征。此外，尺度空间中各尺度图像的模糊程度逐渐变大，能够模拟人在距离目标由近到远时目标在视网膜上的形成过程；(2) The present invention builds a Gaussian pyramid based on the image scale space theory, and the scale space method incorporates the traditional single-scale image information processing technology into the dynamic analysis framework with constantly changing scales, which makes it easier to obtain the essential features of the image. In addition, the blurring degree of each scale image in the scale space gradually increases, which can simulate the formation process of the target on the retina when the distance from the target is close to far;

(3)本发明使用费舍尔向量(Fisher Vector)取代传统的BOW，采用互信息方法对最终数据集的高维特征进行特征选择，从而达到理想的分类结果。(3) The present invention uses Fisher Vector to replace the traditional BOW, and adopts the mutual information method to select the high-dimensional features of the final data set, so as to achieve an ideal classification result.

附图说明Description of drawings

图1是本发明的基本流程图。Figure 1 is a basic flow chart of the present invention.

具体实施方式detailed description

以下将结合附图，对本发明的技术方案进行详细说明。The technical solutions of the present invention will be described in detail below in conjunction with the accompanying drawings.

实施例以STL-10数据库为例，该数据库包含10类RGB图像，每幅图像的大小为96*96。其中用于有监督训练的训练样本数共为5000，将5000个训练样本划分为十折，每次用于监督训练的训练样本数为1000，测试样本数为8000。The embodiment takes the STL-10 database as an example, the database contains 10 types of RGB images, and the size of each image is 96*96. The number of training samples used for supervised training is 5000 in total, and the 5000 training samples are divided into ten folds. The number of training samples used for supervised training each time is 1000, and the number of test samples is 8000.

基于局部图像块描述子和费舍尔向量的图像分类方法，如图1所示，具体步骤如下。The image classification method based on local image block descriptors and Fisher vectors is shown in Figure 1, and the specific steps are as follows.

步骤1，构建基于局部图像块的描述子：Step 1, build descriptors based on local image blocks:

(1a)将待分类的图像数据集分为各种类对应的训练数据集和测试数据集，其中训练数据集的数目为1000，测试数据集的数目为8000；(1a) divide the image data set to be classified into training data sets and test data sets corresponding to various classes, wherein the number of training data sets is 1000, and the number of test data sets is 8000;

(1b)采取随机采样的方法分别对训练数据集和测试数据集中的每幅图像中提取1000个图像块，图像块的大小为8*8；(1b) Take a random sampling method to extract 1000 image blocks from each image in the training data set and the test data set, and the size of the image block is 8*8;

(1c)对采集的图像块进行亮度和对比度归一化和白化预处理操作，将其转换为基于像素值的列向量特征描述子，分别构成训练数据集的特征描述子和测试数据集的特征描述子；(1c) Perform brightness and contrast normalization and whitening preprocessing operations on the collected image blocks, and convert them into column vector feature descriptors based on pixel values, which constitute the feature descriptors of the training data set and the characteristics of the test data set respectively descriptor;

(1d)对步骤(1c)中训练数据集的特征描述子和测试数据集的特征描述子进行PCA降维处理，获取最终训练集和测试集特征描述子组成的集合，分别记为C、E；(1d) Perform PCA dimensionality reduction processing on the feature descriptors of the training data set and the feature descriptors of the test data set in step (1c), and obtain a set composed of the final training set and test set feature descriptors, which are respectively denoted as C and E ;

步骤2，利用高斯混合模型对训练数据集的特征描述子进行建模：利用高斯混合模型对训练数据集特征描述子集C进行聚类，获取高斯混合模型参数θ＝{π_i,μ_i,∑_i,i＝1,...256}，其中高斯单元的个数取为256，π_i,μ_i,∑_i分别表示第i个高斯单元的权重、均值和协方差矩阵：Step 2, use the Gaussian mixture model to model the feature descriptors of the training data set: use the Gaussian mixture model to cluster the feature description subset C of the training data set, and obtain the parameters of the Gaussian mixture model θ={π _i ,μ _i , ∑ _i , i=1,...256}, where the number of Gaussian units is 256, π _i , μ _i , ∑ _i represent the weight, mean and covariance matrix of the i-th Gaussian unit respectively:

(2a)将256个高斯单元的高斯混合模型记为其中p_i(x)代表第i个高斯单元，D表示特征描述子x的维数；(2a) Record the Gaussian mixture model of 256 Gaussian units as where p _i (x) represents the ith Gaussian unit, D represents the dimensionality of the feature descriptor x;

(2b)通过训练集特征描述子组成的集合C来估计高斯混合模型的所有参数θ＝{π_i,μ_i,∑_i,i＝1,...256}。构造集合C的概率公式为：(2b) Estimate all parameters of the Gaussian mixture model θ={π _i ,μ _i ,∑ _i ,i=1,...256} through the set C composed of feature descriptors in the training set. The probability formula for constructing a set C is:

其中，S为集合C的特征描述子个数，S的值为1000； Among them, S is the number of feature descriptors of the set C, and the value of S is 1000;

(2c)采用期望值最大算法根据步骤(2b)中的概率公式对高斯混合模型进行参数估计；(2c) using the expected value maximum algorithm to estimate the parameters of the Gaussian mixture model according to the probability formula in step (2b);

步骤3，分别生成训练数据集和测试数据集中每幅图像的Fisher Vector：Step 3, generate the Fisher Vector of each image in the training dataset and test dataset respectively:

(3a)设X＝{x_t,t＝1,...,1000}是数据集中图像I的特征描述子集合，p_θ是对X进行生成过程建模的概率密度函数，利用高斯混合模型参数θ，计算X中每个特征描述子在参数中第i个高斯单元上的梯度向量；(3a) Let X={x _t ,t=1,...,1000} be the feature descriptor set of image I in the data set, p _θ is the probability density function for modeling the generation process of X, using the Gaussian mixture model Parameter θ, calculate the gradient vector of each feature descriptor in X on the i-th Gaussian unit in the parameter;

(3b)计算概率密度函数p_i的Fisher信息矩阵F_i；(3b) Calculate the Fisher information matrix F _{i of the probability density function p i ;}

(3c)对信息矩阵F_i进行柯列斯基分解从而得到柯列斯基分量，将此分量与步骤(3a)中得到的梯度向量相乘，由此可得到图像I的Fisher Vector F₀；(3c) Colesky decomposition is carried out to the information matrix F _i to obtain the Colesky component, and this component is multiplied with the gradient vector obtained in step (3a), thus obtaining the Fisher Vector F ₀ of the image I;

(3d)计算数据集所有图像的Fisher Vector，从而构成训练数据集和测试数据集的Fisher Vector集；(3d) Calculate the Fisher Vector of all images in the data set, thereby forming the Fisher Vector set of the training data set and the test data set;

步骤4，构建高斯空间金字塔：Step 4, build a Gaussian space pyramid:

(4a)利用高斯卷积函数生成图像I尺度空间：L(q,y,σ)＝G(q,y,σ)*I(q,y)。G(q,y,σ)是尺度可变高斯函数，记为其中(q,y)是空间坐标，σ是尺度坐标。通过变换不同的平滑系数σ,kσ,k²σ...形成多幅图像，将这些多幅图像构成第一组O₁金字塔的若干层图像；(4a) Generate image I scale space by using Gaussian convolution function: L(q,y,σ)=G(q,y,σ)*I(q,y). G(q,y,σ) is a scale-variable Gaussian function, denoted as where (q, y) is the spatial coordinate and σ is the scale coordinate. Multiple images are formed by transforming different smoothing coefficients σ, kσ, k ² σ..., and these multiple images constitute the first group of O ₁ pyramid layers of images;

(4b)将步骤(4a)一组金字塔的倒数第三层图像作比例因子为2的下采样操作，得到第二组金子塔的第一层图像，通过变换不同的平滑系数2σ,2kσ,2k²σ...，得到第二组O₂金字塔的若干层图像；(4b) Perform a downsampling operation on the third-to-last layer images of a group of pyramids in step (4a) with a scale factor of 2 to obtain the first-layer images of the second group of pyramids, by transforming different smoothing coefficients 2σ, 2kσ, 2k ² σ..., get the second group of O ₂ Pyramid layers of images;

(4c)反复执行步骤(4a)和(4b)，可得到一共O组，每组S层图像，共计O*S幅图像，这些图像构成高斯空间金字塔，优选地，O＝3,S＝4，其中，令可得σ＝σ₀2°·k^s，取σ₀＝1.6·2^1/S；(4c) Steps (4a) and (4b) are repeatedly executed to obtain a total of O groups, each group of S layer images, a total of O*S images, these images form a Gaussian space pyramid, preferably, O=3, S=4 , among them, order It can be obtained that σ=σ ₀ 2°·k ^s , take σ ₀ =1.6·2 ^1/S ;

步骤5，计算训练数据集和测试数据集的最终特征描述：Step 5, calculate the final feature description of the training data set and the test data set:

(5a)计算图像I高斯空间金字塔中每幅图像的Fisher Vector F_j,j＝1,...,12。将12个Fisher Vector与步骤(3c)得到的向量F₀进行串联，形成图像I最终的特征表示F＝[F₀,F₁,...,F₁₂]；(5a) Calculate the Fisher Vector F _j of each image in the Gaussian space pyramid of image I, j=1,...,12. Concatenate 12 Fisher Vectors with the vector F ₀ obtained in step (3c) to form the final feature representation of image I F=[F ₀ , F ₁ ,...,F ₁₂ ];

(5b)分别计算训练数据集和测试数据集的每幅图像最终的特征表示，形成训练数据集和测试数据集最终特征描述集合为trainfeature、testfeature；(5b) Calculate the final feature representation of each image of the training data set and the test data set respectively, forming the final feature description set of the training data set and the test data set as trainfeature, testfeature;

步骤6，采用互信息方法对最终的数据集特征进行特征选择，对于步骤(5b)中数据集最终的高维特征向量，使用基于重要性排序的互信息方法来进行特征选择。Step 6, using the mutual information method to perform feature selection on the final data set features, and for the final high-dimensional feature vector of the data set in step (5b), use the mutual information method based on importance ranking to perform feature selection.

具体实现步骤如下：The specific implementation steps are as follows:

(6a)将步骤(5b)中的最终特征描述集合trainfeature、testfeature进行转置，记训练集样本标签和测试集样本标签分别为l₁和l₂，训练图像特征集合中每一维特征的重要性即为所需获取的互信息，计算表达式为：(6a) Transpose the final feature description sets trainfeature and testfeature in step (5b), record the training set sample label and test set sample label as l ₁ and l ₂ respectively, and the importance of each dimension feature in the training image feature set is the mutual information to be obtained, and the calculation expression is:

I(x_:i,l₁)＝H(l₁)+H(x_:i)-H(x_:i,l₁)，其中x_:i表示trainfeature的第i维特征，H表示随机变量的熵值；I(x _:i ,l ₁ )=H(l ₁ )+H(x _:i )-H(x _:i ,l ₁ ), where x _:i represents the i-th dimension feature of the trainfeature, and H represents the random variable entropy value;

(6b)使用1-bit量化方法取代估计概率密度函数。1-bit量化将实数x量化至两个离散箱(bin)，可通过如下公式进行计算：(6b) Use a 1-bit quantization method instead of estimating the probability density function. 1-bit quantization quantizes the real number x to two discrete bins, which can be calculated by the following formula:

离散熵值H(x_:i)和H(x_:i,l₁)可进一步通过公式H(x)＝-∑_jp_jlog₂(p_j)计算，p_j表示x落入第j个离散箱的概率值，其中j＝1,2；The discrete entropy values H(x _:i ) and H(x _:i ,l ₁ ) can be further calculated by the formula H(x)=-∑ _j p _j log ₂ (p _j ), p _j means that x falls into the jth Probability values for discrete bins, where j=1,2;

(6c)计算trainfeature中每一维特征互信息值I(x_:i,l₁)，然后对其进行排序，选取前d个特征，优选地，d＝10000，并记录前d个特征索引值index；(6c) Calculate the mutual information value I(x _:i ,l ₁ ) of each dimension feature in the trainfeature, and then sort it, select the first d features, preferably, d=10000, and record the first d feature index values index;

(6d)根据步骤(6b)(6c)计算testfeature中每一维特征互信息值I(x'_:i,l₂)，x'_:i表示testfeature的第i维特征。根据步骤(6c)得到的索引值index选取前d个特征；(6d) Calculate the mutual information value I(x' _:i ,l ₂ ) of each dimension feature in the testfeature according to steps (6b)(6c), where x' _:i represents the i-th dimension feature of the testfeature. Select the first d features according to the index value index obtained in step (6c);

步骤7，采用步骤6中经过特征选择后训练样本的特征值来训练SVM分类器，将测试样本的特征值输入已训练的SVM分类器中实现图像分类。In step 7, the feature value of the training sample after feature selection in step 6 is used to train the SVM classifier, and the feature value of the test sample is input into the trained SVM classifier to realize image classification.

实施例仅为说明本发明的技术思想，不能以此限定本发明的保护范围，凡是按照本发明提出的技术思想，在技术方案基础上所做的任何改动，均落入本发明保护范围之内。The embodiment is only to illustrate the technical idea of the present invention, and can not limit the scope of protection of the present invention with this. All technical ideas proposed in the present invention, any changes made on the basis of technical solutions, all fall within the scope of protection of the present invention .