CN104616013A - Method for acquiring low-dimensional local characteristics descriptor - Google Patents

Abstract

Translated fromChinese

本发明提供一种获取低维局部特征描述子的方法，其中，所述方法包括：获取待处理图像的局部特征描述子；将获取的局部特征描述子形成描述子集合；根据降维矩阵将所述描述子集合中的每一个局部特征描述子进行降维，获得与每一个局部特征描述子对应的低维局部特征描述子；其中，所述降维矩阵为训练预设的图像数据集得到的矩阵。上述方法可以降低现有技术中局部特征描述子的维度，并去除现有技术中局部特征描述子的冗余信息。

The present invention provides a method for acquiring low-dimensional local feature descriptors, wherein the method includes: acquiring local feature descriptors of an image to be processed; forming the acquired local feature descriptors into a descriptor set; Each local feature descriptor in the descriptor set is dimensionally reduced to obtain a low-dimensional local feature descriptor corresponding to each local feature descriptor; wherein, the dimensionality reduction matrix is obtained by training a preset image data set matrix. The above method can reduce the dimension of the local feature descriptor in the prior art, and remove the redundant information of the local feature descriptor in the prior art.

Description

Translated fromChinese

获取低维局部特征描述子的方法A method to obtain low-dimensional local feature descriptors

技术领域technical field

本发明实施例涉及计算机领域，尤其涉及一种获取低维局部特征描述子的方法。Embodiments of the present invention relate to the field of computers, and in particular to a method for obtaining low-dimensional local feature descriptors.

背景技术Background technique

目前，移动视觉搜索应用越来越多，业内人士通常采用局部特征描述子聚合成全局特征描述子以实现图像检索或分类。举例来说，局部特征描述子聚合成全局特征描述子如Fisher向量。At present, there are more and more mobile visual search applications, and people in the industry usually use local feature descriptors to aggregate into global feature descriptors to achieve image retrieval or classification. For example, local feature descriptors are aggregated into global feature descriptors such as Fisher vectors.

现有技术中采用局部特征描述子聚合成全局特征描述子以实现图像检索或分类的具体实现方式为：首先，提取图像的局部特征描述子，并直接利用局部特征描述子聚合Fisher向量。然而，由于提取图像的局部特征描述子的维度较高，使得聚合Fisher向量的时间和空间复杂度较高，进一步地，由于局部特征描述子的维度较高，使得获取的Fisher向量维度较高，使得全局特征描述子占用空间非常大，容易造成传输延迟，进而影响了图像检索或图像分类的响应时间。In the prior art, local feature descriptors are aggregated into global feature descriptors to achieve image retrieval or classification. The specific implementation method is as follows: first, the local feature descriptors of the image are extracted, and the Fisher vectors are directly aggregated using the local feature descriptors. However, due to the high dimension of the local feature descriptor of the extracted image, the time and space complexity of aggregating the Fisher vector is high. Further, due to the high dimension of the local feature descriptor, the obtained Fisher vector has a high dimension. This makes the global feature descriptor occupy a very large space, which is easy to cause transmission delay, and then affects the response time of image retrieval or image classification.

另外，直接使用局部特征描述子聚合Fisher向量，使得聚合成的Fisher向量的判别力下降，不具有鲁棒性，进而降低了图像检索的准确度。In addition, the direct use of local feature descriptors to aggregate Fisher vectors reduces the discriminative power of the aggregated Fisher vectors, which is not robust and reduces the accuracy of image retrieval.

发明内容Contents of the invention

为解决现有技术中的缺陷，本发明提供一种用于获取低维局部特征描述子的方法，用于降低现有技术中局部特征描述子的维度，并去除现有技术中局部特征描述子的冗余信息。In order to solve the defects in the prior art, the present invention provides a method for obtaining low-dimensional local feature descriptors, which is used to reduce the dimension of the local feature descriptors in the prior art and remove the local feature descriptors in the prior art redundant information.

本发明提供一种获取低维局部特征描述子的方法，包括：The present invention provides a method for obtaining low-dimensional local feature descriptors, including:

获取待处理图像的局部特征描述子；Obtain the local feature descriptor of the image to be processed;

将获取的局部特征描述子形成描述子集合；Form the obtained local feature descriptors into a descriptor set;

根据降维矩阵，将所述描述子集合中的每一个局部特征描述子进行降维，获得与每一个局部特征描述子对应的低维局部特征描述子；其中，所述降维矩阵为训练预设的图像数据集得到的矩阵。According to the dimensionality reduction matrix, each local feature descriptor in the descriptor set is dimensionally reduced to obtain a low-dimensional local feature descriptor corresponding to each local feature descriptor; wherein, the dimensionality reduction matrix is the pre-training Given an image dataset to get the matrix.

可选地，根据降维矩阵，将所述描述子集合中的每一个局部特征描述子进行降维，获得与每一个局部特征描述子对应的低维局部特征描述子，包括：Optionally, according to the dimensionality reduction matrix, each local feature descriptor in the descriptor set is subjected to dimensionality reduction to obtain a low-dimensional local feature descriptor corresponding to each local feature descriptor, including:

所述描述子集合中的每一局部特征描述子减去预设的均值向量，得到转换后的局部特征描述子；Subtracting a preset mean value vector from each local feature descriptor in the descriptor set to obtain a converted local feature descriptor;

将转换后的局部特征描述子组成数据矩阵；Compile the transformed local feature descriptors into a data matrix;

将所述降维矩阵和所述数据矩阵相乘，得到结果矩阵；multiplying the dimensionality reduction matrix and the data matrix to obtain a result matrix;

拆分所述结果矩阵，获得低维局部特征描述子；Splitting the result matrix to obtain a low-dimensional local feature descriptor;

其中，预设的均值向量为训练预设的图像数据集得到的向量，且预设的均值向量的维度与所述局部特征描述子的维度相同。Wherein, the preset mean vector is a vector obtained by training a preset image data set, and the dimension of the preset mean vector is the same as that of the local feature descriptor.

可选地，将转换后的局部特征描述子组成数据矩阵，包括：Optionally, the converted local feature descriptors are formed into a data matrix, including:

在每一转换后的局部特征描述子的维度为N时，将每一局部特征描述子的每一维度上的元素组成所述数据矩阵中对应一行上的数值，以获取M*N维的数据矩阵；When the dimension of each converted local feature descriptor is N, the elements on each dimension of each local feature descriptor are formed into the values on the corresponding row in the data matrix to obtain M*N-dimensional data matrix;

或者，or,

在每一转换后的局部特征描述子的维度为N时，将每一局部特征描述子的每一维度上的元素组成所述数据矩阵的中对应一列上的数值，以获取N*M维的数据矩阵；When the dimension of each converted local feature descriptor is N, the elements on each dimension of each local feature descriptor are composed of the values on the corresponding column of the data matrix to obtain N*M-dimensional data matrix;

其中，M为转换后的局部特征描述子的个数，且M为自然数，N等于128。Among them, M is the number of converted local feature descriptors, and M is a natural number, and N is equal to 128.

可选地，所述降维矩阵为采用主成分分析方式从所述图像数据集中获取的矩阵，所述降维矩阵的维度为N*K，或者，所述降维矩阵的维度为K*N；Optionally, the dimensionality reduction matrix is a matrix obtained from the image data set by principal component analysis, and the dimensionality of the dimensionality reduction matrix is N*K, or, the dimensionality of the dimensionality reduction matrix is K*N ;

在所述降维矩阵的维度为N*K，所述数据矩阵的维度为M*N时，所述结果矩阵的维度为M*K；或者，When the dimensionality of the dimensionality reduction matrix is N*K and the dimensionality of the data matrix is M*N, the dimensionality of the result matrix is M*K; or,

在所述降维矩阵的维度为K*N，所述数据矩阵的维度为N*M时，所述结果矩阵的维度为K*M；When the dimension of the dimension reduction matrix is K*N, and the dimension of the data matrix is N*M, the dimension of the result matrix is K*M;

其中，K等于32。Among them, K is equal to 32.

可选地，所述拆分所述结果矩阵，获得低维局部特征描述子，包括：Optionally, the splitting the result matrix to obtain a low-dimensional local feature descriptor includes:

若所述结果矩阵的维度为M*K，则提取所述结果矩阵中的每一行中的数值，将提取的每一行的数值作为一个低维局部特征描述子；If the dimension of the result matrix is M*K, then extract the value in each row in the result matrix, and use the value of each row extracted as a low-dimensional local feature descriptor;

或者，or,

若所述结果矩阵的维度为K*M，则提取所述结果矩阵中的每一列中的数值，将提取的每一列的数值作为一个低维局部特征描述子；If the dimension of the result matrix is K*M, then extract the value in each column in the result matrix, and use the value of each column extracted as a low-dimensional local feature descriptor;

其中，M为转换后的局部特征描述子的个数，且M为自然数，K等于32。Among them, M is the number of converted local feature descriptors, and M is a natural number, and K is equal to 32.

可选地，所述拆分所述结果矩阵，获得低维局部特征描述子，包括：Optionally, the splitting the result matrix to obtain a low-dimensional local feature descriptor includes:

提取所述结果矩阵中的每一行中的数值，将提取的每一行的数值作为一个低维局部特征描述子，得到M个低维局部特征描述子，且每一低维局部特征描述子的维度为K；Extracting the value in each row in the result matrix, using the extracted value in each row as a low-dimensional local feature descriptor to obtain M low-dimensional local feature descriptors, and the dimension of each low-dimensional local feature descriptor is K;

其中，M为转换后的局部特征描述子的个数，且M为自然数，K等于32；Among them, M is the number of transformed local feature descriptors, and M is a natural number, and K is equal to 32;

或者，or,

提取所述结果矩阵中的每一列中的数值，将提取的每一列的数值作为一个低维局部特征描述子，得到M个低维局部特征描述子，且每一低维局部特征描述子的维度为K；Extract the value in each column in the result matrix, use the extracted value of each column as a low-dimensional local feature descriptor to obtain M low-dimensional local feature descriptors, and the dimension of each low-dimensional local feature descriptor is K;

其中，M为转换后的局部特征描述子的个数，且M为自然数，K等于32。Among them, M is the number of converted local feature descriptors, and M is a natural number, and K is equal to 32.

可选地，所述根据降维矩阵将所述描述子集合中的局部特征描述子进行降维，获得低维局部特征描述子之前，所述方法还包括：Optionally, performing dimensionality reduction on the local feature descriptors in the descriptor set according to the dimensionality reduction matrix, and before obtaining the low-dimensional local feature descriptors, the method further includes:

获取所述图像数据集的样本矩阵；Obtain a sample matrix of the image data set;

根据样本矩阵获得均值向量；Obtain the mean vector from the sample matrix;

利用所述均值向量对样本矩阵进行中心化，得到中心化后的样本矩阵；centering the sample matrix by using the mean value vector to obtain a centered sample matrix;

计算所述中心化后的样本矩阵的协方差矩阵；Calculating the covariance matrix of the centered sample matrix;

获取所述协方差矩阵的特征值，以及与所述特征值对应的特征向量；Obtaining the eigenvalues of the covariance matrix and the eigenvectors corresponding to the eigenvalues;

根据所述特征值的大小对所述特征向量进行由大到小排序，选取前K个所述特征向量；Sorting the eigenvectors from large to small according to the size of the eigenvalues, and selecting the first K eigenvectors;

将所述前K个特征向量组成所述降维矩阵；Composing the first K eigenvectors into the dimensionality reduction matrix;

其中，K等于32。Among them, K is equal to 32.

可选地，所述图像数据集中每一图像的每一个局部特征描述子对应所述样本矩阵中的一行数值，所述图像数据集中每一图像对应所述样本矩阵中的若干行样本数值，所述样本矩阵中每一行的样本数值有N个；Optionally, each local feature descriptor of each image in the image data set corresponds to a row of values in the sample matrix, and each image in the image data set corresponds to several rows of sample values in the sample matrix, so There are N sample values in each row in the sample matrix;

所述根据样本矩阵获得均值向量，包括：The obtaining the mean vector according to the sample matrix includes:

对所述样本矩阵每一列上的所有数值求平均值，所述均值向量的第i个维度的数值等于所述样本矩阵第i列的平均值，其中i＝1,…,N；Calculate the average value of all values on each column of the sample matrix, the value of the i-th dimension of the mean vector is equal to the average value of the i-th column of the sample matrix, where i=1,...,N;

所述利用所述均值向量对样本矩阵进行中心化，得到中心化后的样本矩阵，包括：The centering of the sample matrix by using the mean vector to obtain the centered sample matrix includes:

所述样本矩阵的每一行上的第i个数值减去所述均值向量的第i个维度的数值，得到中心化后的样本矩阵,其中i＝1,…,N；Subtracting the value of the i-th dimension of the mean vector from the i-th value on each row of the sample matrix to obtain a centered sample matrix, where i=1,...,N;

所述协方差矩阵的维度为N*N；The dimension of the covariance matrix is N*N;

所述特征向量的维度为N；The dimension of the feature vector is N;

所述前K个特征向量中所有特征向量的所有维度的元素组成所述降维矩阵中行/列的数值；Elements of all dimensions of all eigenvectors in the first K eigenvectors form the values of rows/columns in the dimensionality reduction matrix;

或者，or,

所述图像数据集中每一图像的每一个局部特征描述子对应所述样本矩阵中的一列数值，所述图像数据集中每一图像对应所述样本矩阵中的若干列样本数值，所述样本矩阵中每一列的样本数值有N个；Each local feature descriptor of each image in the image data set corresponds to a column of values in the sample matrix, each image in the image data set corresponds to several columns of sample values in the sample matrix, and in the sample matrix There are N sample values in each column;

所述根据样本矩阵获得均值向量，包括：The obtaining the mean vector according to the sample matrix includes:

对所述样本矩阵每一行上的所有数值求平均值，所述均值向量的第i个维度的数值等于所述样本矩阵第i行的平均值，其中i＝1,…,N；Calculate the average value of all values on each row of the sample matrix, the value of the i-th dimension of the mean vector is equal to the average value of the i-th row of the sample matrix, where i=1,...,N;

所述利用所述均值向量对样本矩阵进行中心化，得到中心化后的样本矩阵，包括：The centering of the sample matrix by using the mean vector to obtain the centered sample matrix includes:

所述样本矩阵的每一列上的第i个数值减去所述均值向量的第i个维度的数值，得到中心化后的样本矩阵，其中i＝1,…,N；Subtracting the i-th value of each column of the sample matrix from the value of the i-th dimension of the mean vector to obtain a centered sample matrix, where i=1,...,N;

所述协方差矩阵的维度为N*N；The dimension of the covariance matrix is N*N;

所述特征向量的维度为N；The dimension of the feature vector is N;

所述前K个特征向量中所有特征向量的所有维度的元素组成所述降维矩阵中列/行的数值；Elements of all dimensions of all eigenvectors in the first K eigenvectors form the values of columns/rows in the dimensionality reduction matrix;

其中，N等于128，K等于32。Among them, N is equal to 128, and K is equal to 32.

可选地，所述图像数据集包括：平面物体图像和三维物体图像。Optionally, the image data set includes: a planar object image and a three-dimensional object image.

由上述技术方案可知，本发明的获取低维局部特征描述子的方法，通过获取待处理图像的局部特征描述子，进而将获取的所有局部特征描述子形成描述子集合，采用降维矩阵对每一局部特征描述子进行降维，获得每一局部特征描述子的低维局部特征描述子，由此可以降低现有技术中局部特征描述子的维度，并去除现有技术中局部特征描述子的冗余信息。It can be seen from the above technical solution that the method for obtaining low-dimensional local feature descriptors of the present invention obtains the local feature descriptors of the image to be processed, and then forms all the obtained local feature descriptors into a descriptor set, and uses a dimensionality reduction matrix for each A local feature descriptor is used for dimensionality reduction to obtain a low-dimensional local feature descriptor for each local feature descriptor, thereby reducing the dimension of the local feature descriptor in the prior art and removing the local feature descriptor in the prior art. redundant information.

附图说明Description of drawings

图1为本发明一实施例提供的获取低维局部特征描述子的流程示意图；FIG. 1 is a schematic flow diagram of obtaining low-dimensional local feature descriptors provided by an embodiment of the present invention;

图2为本发明另一实施例提供的获取低维局部特征描述子的流程示意图；FIG. 2 is a schematic flow diagram of obtaining low-dimensional local feature descriptors provided by another embodiment of the present invention;

图3为本发明一实施例提供的梯度方向直方图向量的示意图。FIG. 3 is a schematic diagram of a gradient direction histogram vector provided by an embodiment of the present invention.

具体实施方式Detailed ways

图1示出了本发明一实施例提供的获取低维局部特征描述子的方法，如图1所示，本实施例中的获取低维局部特征描述子的方法如下所述。FIG. 1 shows a method for obtaining a low-dimensional local feature descriptor provided by an embodiment of the present invention. As shown in FIG. 1 , the method for obtaining a low-dimensional local feature descriptor in this embodiment is as follows.

101、获取待处理图像的局部特征描述子。101. Acquire a local feature descriptor of an image to be processed.

举例来说，上述提及的待处理图像可以是任意一幅图像，如，该待处理图像可以是文件的照片，或者是手绘的图片，油画图像，从视频中截取的帧，地标照片、或者物品照片等，本实施例不限定上述待处理图像的类型和内容。For example, the image to be processed mentioned above can be any image, for example, the image to be processed can be a photo of a document, or a hand-painted picture, an oil painting image, a frame intercepted from a video, a photo of a landmark, or Item photos, etc., this embodiment does not limit the types and contents of the above-mentioned images to be processed.

特别地，获取待处理图像的一个或多个局部特征描述子的方式为现有的方式，例如，上述的局部特征描述子可为尺度不变描述子(Scale InvariantFeature Transform，简称：SIFT)，或者，上述的局部特征描述子可为快速鲁棒的尺度不变特征描述子(Speeded Up Robust Features，简称：SURF)，或其他局部特征描述子。In particular, the way to obtain one or more local feature descriptors of the image to be processed is an existing way, for example, the above-mentioned local feature descriptor can be a scale invariant descriptor (Scale InvariantFeature Transform, referred to as: SIFT), or , the above-mentioned local feature descriptor can be a fast and robust scale-invariant feature descriptor (Speeded Up Robust Features, referred to as: SURF), or other local feature descriptors.

应了解的是，SIFT或SURF的提取方式可为现有的提取方式，本实施例不再详述。通常，SIFT的维度可为128维，SURF的维度可为64维。It should be understood that the extraction manner of SIFT or SURF may be an existing extraction manner, which will not be described in detail in this embodiment. Generally, the dimension of SIFT can be 128 dimensions, and the dimension of SURF can be 64 dimensions.

可选地，获取待处理图像的局部特征描述子可以在上述局部特征描述子的获取方式的基础上进行特征选择等处理，选择一幅图像对应的全部局部特征描述子的一个或多个。Optionally, to obtain the local feature descriptors of the image to be processed, one or more of all the local feature descriptors corresponding to an image may be selected by performing feature selection and other processing on the basis of the above-mentioned acquisition method of the local feature descriptors.

102、将获取的局部特征描述子形成描述子集合。102. Form the acquired local feature descriptors into a descriptor set.

在本实施例中，将获取的所有局部特征描述子形成描述子集合。In this embodiment, all acquired local feature descriptors form a descriptor set.

103、根据降维矩阵，将所述描述子集合中的每一个局部特征描述子进行降维，获得与每一个局部特征描述子对应的低维局部特征描述子。103. Perform dimensionality reduction on each local feature descriptor in the descriptor set according to the dimensionality reduction matrix, to obtain a low-dimensional local feature descriptor corresponding to each local feature descriptor.

在本实施例中，步骤103中的降维矩阵可为训练预设的图像数据集得到的矩阵。In this embodiment, the dimensionality reduction matrix in step 103 may be a matrix obtained from training a preset image data set.

可选地，在步骤103之前，还可对所述描述子集合中的所有局部特征描述子进行归一化处理，进而可在步骤103中对归一化处理后的局部特征描述子进行降维处理，获得与每一个局部特征描述子对应的低维局部特征描述子。Optionally, before step 103, normalization processing can also be performed on all local feature descriptors in the descriptor set, and then dimensionality reduction can be performed on the normalized local feature descriptors in step 103 processing to obtain a low-dimensional local feature descriptor corresponding to each local feature descriptor.

所述归一化处理的步骤举例如下：The steps of the normalization processing are exemplified as follows:

A01、若所述局部特征描述子为h_t,t＝0,...,M-1，对每一维度使用L1归一化，得到h′_t,j＝h_t,j/|h_t|,j＝0,...,127；其中，|h_t|表示128维局部特征描述子向量h_t各维度绝对值的和。A01. If the local feature descriptor is h_t , t=0,...,M-1, use L1 normalization for each dimension to obtain h′_t,j =h_t,j /|h_t |,j=0,...,127; wherein, |h_t | represents the sum of the absolute values of each dimension of the 128-dimensional local feature descriptor vector h_t .

A02、对每一维度继续使用参数为0.5的power normalization归一化，得到h′_t,j←sgn(h′_t,j)|h′_t,j|^0.5；A02. Continue to use the power normalization parameter of 0.5 for each dimension to obtain h′_{t, j} ←sgn(h′_{t, j} )|h′_{t, j} |^0.5 ;

其中，|h′_t,j|表示维度h′_t,j的绝对值,$\mathrm{sgn}\left(h_{t,j}^{\&prime;}\right)=\left\{\begin{array}{cc}-1& h_{t,j}^{\&prime;}<0\\ 0& h_{t,j}^{\&prime;}=0\\ 1& h_{t,j}^{\&prime;}>0\end{array}\right..$Among them, |h′_{t, j} | represents the absolute value of dimension h′_{t, j} ,$\mathrm{sgn}\left(h_{t,j}^{\&prime;}\right)=\left\{\begin{array}{cc}-1& h_{t,j}^{\&prime;}<0\\ 0& h_{t,j}^{\&prime;}=0\\ 1& h_{t,j}^{\&prime;}>0\end{array}\right..$

应说明的是，上述方法可以是在任一设备上进行，本实施例不限制其执行主体为客户端还是服务器。It should be noted that the above method can be performed on any device, and this embodiment does not limit whether the execution subject is a client or a server.

本实施例中的获取低维局部特征描述子的方法，可降低现有技术中局部特征描述子的维度，并去除现有技术中局部特征描述子的冗余信息。The method for obtaining low-dimensional local feature descriptors in this embodiment can reduce the dimension of the local feature descriptors in the prior art and remove redundant information of the local feature descriptors in the prior art.

图2示出了本发明另一实施例提供的获取低维局部特征描述子的方法，如图2所示，本实施例中的获取低维局部特征描述子的方法如下所述。FIG. 2 shows a method for obtaining a low-dimensional local feature descriptor provided by another embodiment of the present invention. As shown in FIG. 2 , the method for obtaining a low-dimensional local feature descriptor in this embodiment is as follows.

201、获取待处理图像的局部特征描述子。201. Acquire a local feature descriptor of an image to be processed.

特别地，获取待处理图像的局部特征描述子的方式为如下举例说明：In particular, the way to obtain the local feature descriptor of the image to be processed is illustrated as follows:

第一步：所述待处理图像I与一组高斯滤波器卷积得到图像I在高斯尺度空间中不同尺度下的高斯模糊图像，其中σ为高斯的标准差，表达所述高斯尺度空间中每一高斯模糊图像对应的尺度。σ以2的指数幂取值，第k个尺度为σ_k，且其中σ₀为初始尺度，取值为1.6，K表示对尺度空间采样的层数，即所述高斯滤波器的个数。那么，第k个高斯模糊图像为I_k，对应的尺度为σ_k，且I_k＝I*g(σ_k),k＝0,...,K。The first step: the image to be processed I and a group of Gaussian filters Gaussian blurred images of image I at different scales in the Gaussian scale space are obtained by convolution, where σ is the standard deviation of Gaussian, expressing the scale corresponding to each Gaussian blurred image in the Gaussian scale space. σ is taken as an exponential power of 2, the kth scale is σ_k , and Where σ₀ is the initial scale, with a value of 1.6, and K represents the number of layers for sampling the scale space, that is, the number of Gaussian filters. Then, the kth Gaussian blurred image is I_k , the corresponding scale is σ_k , and I_k =I*g(σ_k ), k=0,...,K.

第二步：在所述高斯尺度空间中，每一所述高斯模糊图像再与尺度规范化的拉普拉斯滤波器卷积得到高斯拉普拉斯尺度空间响应其中$f=\left[\begin{array}{ccc}0& 1& 0\\ 1& -4& 1\\ 0& 1& 0\end{array}\right]$为拉普拉斯算子。Step 2: In the Gaussian scale space, each Gaussian blurred image is convolved with a scale-normalized Laplacian filter to obtain a Gaussian Laplacian scale-space response in$f=\left[\begin{array}{ccc}0& 1& 0\\ 1& -4& 1\\ 0& 1& 0\end{array}\right]$ is the Laplacian operator.

第三步：在所述高斯拉普拉斯尺度空间中，获取局部极大值或极小值点作为候选的兴趣点。所述兴趣点包括三个属性，即所述兴趣点在对应的高斯模糊图像中的位置坐标x,y和对应的尺度σ_k。Step 3: In the Laplacian of Gaussian scale space, obtain local maximum or minimum points as candidate interest points. The interest point includes three attributes, namely the position coordinates x, y of the interest point in the corresponding Gaussian blur image and the corresponding scale σ_k .

第四步：对所述兴趣点，获取其对应的相同尺度的高斯模糊图像I_k上以x,y为中心，以mσ为半径的圆形区域，其中m＝3.96。然后，对所述圆形区域内的像素，按以下公式计算其每个像素的梯度，包括梯度模长和梯度方向${{\mathrm{\&theta;}}_I}_k:$Step 4: For the interest point, obtain a circular area with x, y as the center and mσ as the radius on the corresponding Gaussian blurred image I_k of the same scale, where m=3.96. Then, for the pixels in the circular area, the gradient of each pixel is calculated according to the following formula, including the gradient modulus length and the gradient direction${{\mathrm{\&theta;}}_I}_k:$

${{\mathrm{<span><span>m</span>m</span>}}_{\mathrm{<span><span>I</span>I</span>}}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\sqrt{{<span><span>(</span>(</span>{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span><span><span>-</span>-</span>{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span><span><span>)</span>)</span>}^{\mathrm{<span><span>2</span>2</span>}}<span><span>+</span>+</span>{<span><span>(</span>(</span>{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span><span><span>-</span>-</span>{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span><span><span>)</span>)</span>}^{\mathrm{<span><span>2</span>2</span>}}}$

${{\mathrm{<span><span>\&theta;</span>\&theta;</span>}}_{\mathrm{<span><span>I</span>I</span>}}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span><span><span>=</span>=</span>\mathrm{<span><span>arctan</span>arctan</span>}\frac{{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span><span><span>-</span>-</span>{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}<span><span>)</span>)</span>}{{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>+</span>+</span>\mathrm{<span><span>1</span>1</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span><span><span>-</span>-</span>{\mathrm{<span><span>I</span>I</span>}}_{\mathrm{<span><span>k</span>k</span>}}<span><span>(</span>(</span>\mathrm{<span><span>x</span>x</span>}<span><span>-</span>-</span>\mathrm{<span><span>1</span>1</span>}<span><span>,</span>,</span>\mathrm{<span><span>y</span>the\; y</span>}<span><span>)</span>)</span>}$

将所述圆形区域内的每个像素的梯度方向按最近距离规则量化到圆周36等分的方向上。每个方向以梯度模长为权重做加权累计，得到一个36维的梯度方向直方图。Quantize the gradient direction of each pixel in the circular area to the direction of 36 equal divisions of the circumference according to the shortest distance rule. Each direction is weighted and accumulated with the gradient modulus length as the weight, and a 36-dimensional gradient direction histogram is obtained.

第五步：选取直方图中累计最大的方向作为该兴趣点的主方向θ。同时，如果有其他方向的累计值超过主方向累计值的80％，复制扩展该兴趣点为一个新的兴趣点，并用该方向作为新兴趣点的主方向。Step 5: Select the direction with the largest cumulative value in the histogram as the main direction θ of the interest point. At the same time, if the accumulative value of other directions exceeds 80% of the accumulative value of the main direction, copy and expand this POI into a new POI, and use this direction as the main direction of the new POI.

可选地，对于所述兴趣点，按照其位置x,y，尺度σ，方向θ等属性进行重要性排序，筛选出所需要的点数M供后续的全局特征计算。Optionally, for the interest points, the importance is sorted according to their position x, y, scale σ, direction θ and other attributes, and the required number of points M is selected for subsequent global feature calculation.

第六步：对于检测到的所述兴趣点，获取的相同尺度的高斯模糊图像I_k上以x,y为中心，且坐标系旋转至与主方向θ对齐，以3σ为半径的正方形区域。然后，将所述正方形区域均匀地划分成4*4的图像块，对所述图像块中的每个像素求梯度后，将梯度方向量化到圆周8等分的方向上并计算梯度方向直方图，其累计过程采用三线性插值的方式，然后按照从左到右、从上到下的顺序拼接每个图像块的梯度方向直方图对应的8维的向量，如图3所示，获得4*4*8＝128的梯度方向直方图向量。Step 6: For the detected interest point, the acquired Gaussian blurred image I_k of the same scale is centered on x, y, and the coordinate system is rotated to align with the main direction θ, and a square area with a radius of 3σ. Then, the square area is evenly divided into 4*4 image blocks, and after calculating the gradient for each pixel in the image block, the gradient direction is quantized to the direction of 8 equal parts of the circumference and the gradient direction histogram is calculated , the accumulation process adopts the method of trilinear interpolation, and then stitches the 8-dimensional vector corresponding to the gradient direction histogram of each image block in the order from left to right and from top to bottom, as shown in Figure 3, and obtains 4* 4*8=128 gradient orientation histogram vectors.

最后，对产生的128维梯度方向直方图向量进行一次L2归一化。然后，对每一维度进行截断操作，即对每一维度的值，如果大于0.2，则截断取值为0.2。接着，再对截断后的向量进行一次L2归一化。最终产生所述局部特征描述子。Finally, an L2 normalization is performed on the generated 128-dimensional gradient orientation histogram vector. Then, a truncation operation is performed on each dimension, that is, if the value of each dimension is greater than 0.2, the truncation value is 0.2. Next, L2 normalization is performed on the truncated vector again. Finally, the local feature descriptor is generated.

若梯度向量直方图向量为h，h_i为h第i个维度的数值，i＝0,...,127，所述L2归一化的具体形式为：h′_i为h经过L2归一化后第i个维度的数值。If the gradient vector histogram vector is h, h_i is the value of the i-th dimension of h, i=0,...,127, the specific form of the L2 normalization is: h'_i is the value of the i-th dimension of h after L2 normalization.

可选地，获取待处理图像的局部特征描述子可以在上述局部特征描述子的获取方式的基础上进行特征选择等处理，选择一幅图像对应的全部局部特征描述子的一个或多个。Optionally, to obtain the local feature descriptors of the image to be processed, one or more of all the local feature descriptors corresponding to an image may be selected by performing feature selection and other processing on the basis of the above-mentioned acquisition method of the local feature descriptors.

202、将获取的局部特征描述子形成描述子集合。202. Form the acquired local feature descriptors into a descriptor set.

203、所述描述子集合中的每一局部特征描述子减去预设的均值向量，得到转换后的局部特征描述子。203. Subtract a preset mean value vector from each local feature descriptor in the descriptor set to obtain a converted local feature descriptor.

其中，预设的均值向量为训练预设的图像数据集得到的向量，且预设的均值向量的维度与所述局部特征描述子的维度相同。Wherein, the preset mean vector is a vector obtained by training a preset image data set, and the dimension of the preset mean vector is the same as that of the local feature descriptor.

204、将转换后的局部特征描述子组成数据矩阵。204. Compose the converted local feature descriptors into a data matrix.

举例来说，在每一转换后的局部特征描述子的维度为N时，将每一局部特征描述子的每一维度上的元素组成所述数据矩阵中对应一行上的数值，以获取M*N维的数据矩阵；For example, when the dimension of each converted local feature descriptor is N, the elements on each dimension of each local feature descriptor are formed into the values on the corresponding row in the data matrix to obtain M* N-dimensional data matrix;

或者，or,

在每一转换后的局部特征描述子的维度为N时，将每一局部特征描述子的每一维度上的元素组成所述数据矩阵中对应一列上的数值，以获取N*M维的数据矩阵；When the dimension of each converted local feature descriptor is N, the elements on each dimension of each local feature descriptor are composed into the values corresponding to a column in the data matrix to obtain N*M dimensional data matrix;

上述的M为所述描述子集合中转换后的局部特征描述子的个数，N等于128。The above M is the number of converted local feature descriptors in the descriptor set, and N is equal to 128.

例如，在步骤201中，每一所述局部特征描述子的维度为N＝128，且获得300个所述局部特征描述子，即M＝300，所述转换后的局部特征描述子的维度N＝128，将所述转换后的局部特征描述子的128个元素作为数据矩阵的一行，得到一个300*128维度的数据矩阵。当然，若将所述转换后的局部特征描述子的128个元素作为数据矩阵的一列，则得到128*300的数据矩阵。For example, in step 201, the dimension of each local feature descriptor is N=128, and 300 local feature descriptors are obtained, that is, M=300, and the dimension of the converted local feature descriptor is N =128, the 128 elements of the converted local feature descriptor are used as a row of the data matrix to obtain a data matrix with a dimension of 300*128. Certainly, if the 128 elements of the converted local feature descriptor are used as a column of the data matrix, a 128*300 data matrix is obtained.

205、将降维矩阵和所述数据矩阵相乘，得到结果矩阵。205. Multiply the dimensionality reduction matrix and the data matrix to obtain a result matrix.

在本实施例中，降维矩阵可为采用主成分分析方式从所述图像数据集中获取的矩阵，所述降维矩阵的维度为N*K，或者，所述降维矩阵的维度为K*N，其中，K等于32；In this embodiment, the dimensionality reduction matrix may be a matrix obtained from the image data set by principal component analysis, and the dimensionality of the dimensionality reduction matrix is N*K, or the dimensionality of the dimensionality reduction matrix is K* N, where K is equal to 32;

由上可知，降维矩阵中每一行的维度与局部特征描述子的维度相同，例如若局部特征描述子的维度为128维，则降维矩阵中每一行的维度为128维；降维矩阵中每一列的维度与低维局部特征描述子的维度相同，例如若低维局部特征描述子的维度为32维，则降维矩阵中每一列的维度为32维；It can be seen from the above that the dimension of each row in the dimensionality reduction matrix is the same as the dimensionality of the local feature descriptor. For example, if the dimensionality of the local feature descriptor is 128 dimensions, the dimensionality of each row in the dimensionality reduction matrix is 128 dimensions; in the dimensionality reduction matrix The dimension of each column is the same as that of the low-dimensional local feature descriptor. For example, if the dimension of the low-dimensional local feature descriptor is 32 dimensions, the dimension of each column in the dimensionality reduction matrix is 32 dimensions;

或者，or,

降维矩阵中每一列的维度与局部特征描述子的维度相同，例如若局部特征描述子的维度为128维，则降维矩阵中每一列的维度为128维；降维矩阵中每一行的维度与低维局部特征描述子的维度相同，例如若低维局部特征描述子的维度为32维，则降维矩阵中每一行的维度为32维。The dimension of each column in the dimensionality reduction matrix is the same as the dimensionality of the local feature descriptor. For example, if the dimensionality of the local feature descriptor is 128 dimensions, the dimensionality of each column in the dimensionality reduction matrix is 128 dimensions; the dimensionality of each row in the dimensionality reduction matrix The dimension of the low-dimensional local feature descriptor is the same as that of the low-dimensional local feature descriptor. For example, if the dimension of the low-dimensional local feature descriptor is 32 dimensions, the dimension of each row in the dimensionality reduction matrix is 32 dimensions.

因此，降维矩阵应该是一个128x32或32x128维的矩阵。Therefore, the dimensionality reduction matrix should be a 128x32 or 32x128 dimensional matrix.

应说明的是，在本步骤中降维矩阵的维度为N*K，所述数据矩阵的维度为M*N时，所述结果矩阵的维度为M*K。It should be noted that, in this step, the dimensionality of the dimensionality reduction matrix is N*K, and when the dimensionality of the data matrix is M*N, the dimensionality of the result matrix is M*K.

或者，在本步骤中所述降维矩阵的维度为K*N，所述数据矩阵的维度为N*M时，所述结果矩阵的维度为K*M。Alternatively, in this step, the dimensionality of the dimensionality reduction matrix is K*N, and when the dimensionality of the data matrix is N*M, the dimensionality of the result matrix is K*M.

特别地，数据矩阵的维度为300x128，降维矩阵的维度为128x32，得到的结果矩阵的维度为300x32，计算过程如下所示：In particular, the dimension of the data matrix is 300x128, the dimension of the dimensionality reduction matrix is 128x32, and the dimension of the obtained result matrix is 300x32. The calculation process is as follows:

上述所示的计算过程即为现有技术中的矩阵乘法运算，本实施例不再详述。The calculation process shown above is the matrix multiplication operation in the prior art, and will not be described in detail in this embodiment.

206、拆分所述结果矩阵，获得低维局部特征描述子。206. Split the result matrix to obtain low-dimensional local feature descriptors.

举例来说，若结果矩阵的维度为M*K，则提取所述结果矩阵中的每一行中的数值，将提取的每一行的数值作为一个低维局部特征描述子；For example, if the dimension of the result matrix is M*K, then extract the value in each row in the result matrix, and use the value of each row extracted as a low-dimensional local feature descriptor;

或者，若结果矩阵的维度为K*M，则提取所述结果矩阵中的每一列中的数值，将提取的每一列的数值作为一个低维局部特征描述子。Alternatively, if the dimension of the result matrix is K*M, then extract the value in each column in the result matrix, and use the extracted value in each column as a low-dimensional local feature descriptor.

M和K同上的说明。M and K are as described above.

一种优选的实现方式中，提取所述结果矩阵中的每一行中的数值，将提取的每一行的数值作为一个低维局部特征描述子，得到M个低维局部特征描述子，且每一低维局部特征描述子的维度为K；In a preferred implementation manner, the value in each row in the result matrix is extracted, and the extracted value in each row is used as a low-dimensional local feature descriptor to obtain M low-dimensional local feature descriptors, and each The dimension of the low-dimensional local feature descriptor is K;

或者，提取所述结果矩阵中的每一列中的数值，将提取的每一列的数值作为一个低维局部特征描述子，得到M个低维局部特征描述子，且每一低维局部特征描述子的维度为K。Or, extract the value in each column in the result matrix, and use the extracted value of each column as a low-dimensional local feature descriptor to obtain M low-dimensional local feature descriptors, and each low-dimensional local feature descriptor The dimension of is K.

也就是说，结果矩阵中的每一行(或每一列)对应一个低维局部特征描述子，共M个低维局部特征描述子，低维局部特征描述子的维度为K。That is to say, each row (or each column) in the result matrix corresponds to a low-dimensional local feature descriptor, and there are M low-dimensional local feature descriptors in total, and the dimension of the low-dimensional local feature descriptor is K.

例如，若结果矩阵的维度是300x32，则结果矩阵的每一行对应一个降维后的所述局部特征描述子；若结果矩阵的维度是32x300，则结果矩阵的每一列对应一个降维后的所述局部特征描述子。For example, if the dimension of the result matrix is 300x32, each row of the result matrix corresponds to a dimension-reduced local feature descriptor; if the dimension of the result matrix is 32x300, each column of the result matrix corresponds to a dimension-reduced local feature descriptor. Describe the local feature descriptor.

特别地，若对一个局部特征描述子进行降维得到低维局部特征描述子，上述步骤201至206可以用如下公式表示：In particular, if a local feature descriptor is dimensionally reduced to obtain a low-dimensional local feature descriptor, the above steps 201 to 206 can be expressed by the following formula:

${\mathrm{<span><span>x</span>x</span>}}_{\mathrm{<span><span>t</span>t</span>}}<span><span>=</span>=</span>{\mathrm{<span><span>P</span>P</span>}}^{\mathrm{<span><span>T</span>T</span>}}<span><span>(</span>(</span>{\mathrm{<span><span>h</span>h</span>}}_{\mathrm{<span><span>t</span>t</span>}}^{<span><span>\&prime;</span>\&prime;</span>}<span><span>-</span>-</span>\stackrel{<span><span>~</span>~</span>}{\mathrm{<span><span>h</span>h</span>}}<span><span>)</span>)</span>$

其中，x_t为所述低维局部特征描述子，P为所述降维矩阵，h′_t为所述局部特征描述子，为所述预设的均值向量。Wherein, x_t is the low-dimensional local feature descriptor, P is the dimensionality reduction matrix, h′_t is the local feature descriptor, is the preset mean vector.

由于上述K＝32，进而最后获取的低维局部特征描述子的维度可为32，可较好的降低了现有技术中局部特征描述子的维度，同时可去除现有技术中局部特征描述子的冗余信息，避免了噪声对局部特征描述子性能的影响。Since the above K=32, the dimension of the finally obtained low-dimensional local feature descriptor can be 32, which can better reduce the dimension of the local feature descriptor in the prior art, and can remove the local feature descriptor in the prior art redundant information, avoiding the impact of noise on the performance of local feature descriptors.

特别地，采用低维局部特征描述子聚合Fisher向量的过程具有较低的时间和空间复杂度，且能够聚合得到低维的Fisher向量，减少了对Fisher向量做压缩时所需的空间,也减少了通过无线网络传输产生的延迟，使得聚合得到的Fisher向量在图像检索和匹配中的性能有较大提高。In particular, the process of using low-dimensional local feature descriptors to aggregate Fisher vectors has low time and space complexity, and can be aggregated to obtain low-dimensional Fisher vectors, which reduces the space required for compressing Fisher vectors and reduces The delay generated by wireless network transmission is eliminated, and the performance of the aggregated Fisher vector in image retrieval and matching is greatly improved.

在另一种可选的实现方式中，前述图1中的步骤103可具体包括下述举例的图中未示出的各子步骤A1031至A1037；In another optional implementation manner, the aforementioned step 103 in FIG. 1 may specifically include the following sub-steps A1031 to A1037 not shown in the figure;

A1031、根据所述图像数据集获取所述图像数据集的样本矩阵。A1031. Obtain a sample matrix of the image data set according to the image data set.

举例来说，所述图像数据集中每一图像的若干个局部特征描述子组成所述样本矩阵中每一行的样本数值。或者，所述图像数据集中每一图像的若干个局部特征描述子组成所述样本矩阵中每一列的样本数值。For example, several local feature descriptors of each image in the image data set constitute the sample value of each row in the sample matrix. Alternatively, several local feature descriptors of each image in the image data set form the sample value of each column in the sample matrix.

特别地，图像数据集涵盖实际应用中可能出现的各类图像，包括平面物体图像，例如：名片、CD封面、DVD封面、报纸、油画、视频帧等等，也包括三维物体图像，例如：地标建筑以及各种立体实物的照片等等。图像数据集所包含的图像类型应全面，且各种类型图像的比例适当，例如：平面物体图像所占比例为80％，三维物体图像所占比例为20％。In particular, the image dataset covers all kinds of images that may appear in practical applications, including images of flat objects, such as: business cards, CD covers, DVD covers, newspapers, oil paintings, video frames, etc., as well as images of three-dimensional objects, such as: landmarks Photos of buildings and various three-dimensional objects, etc. The types of images included in the image data set should be comprehensive, and the proportion of various types of images should be appropriate, for example: the proportion of plane object images is 80%, and the proportion of three-dimensional object images is 20%.

获取图像数据集中图像的局部特征描述子，获取局部特征描述子的方式如上步骤201所述。The local feature descriptors of the images in the image data set are obtained, and the manner of obtaining the local feature descriptors is as described in step 201 above.

优选局部特征描述子，其维度为128，若得到的局部特征描述子的个数为L个，且样本矩阵的每一行对应一个局部特征描述子，则得到一个Lx128的样本矩阵；若样本矩阵的每一列对应一个局部特征描述子，则得到一个128xL的样本矩阵。The preferred local feature descriptor has a dimension of 128. If the number of local feature descriptors obtained is L, and each row of the sample matrix corresponds to a local feature descriptor, a sample matrix of Lx128 is obtained; if the sample matrix Each column corresponds to a local feature descriptor, and a 128xL sample matrix is obtained.

A1032、根据样本矩阵获得均值向量。A1032. Obtain a mean value vector according to the sample matrix.

举例来说，若样本矩阵为一个L*128的矩阵，则对所述样本矩阵每一列上的所有数值求平均值，所述均值向量的第i个维度的数值等于所述样本矩阵第i列的平均值，其中i＝1,…,N；For example, if the sample matrix is a matrix of L*128, then all values on each column of the sample matrix are averaged, and the value of the i-th dimension of the mean vector is equal to the i-th column of the sample matrix The average value of , where i=1,...,N;

或者，or,

若样本矩阵为一个128*L的矩阵，则对所述样本矩阵每一行上的所有数值求平均值，所述预设的均值向量的第i个维度的数值等于所述样本矩阵第i行的平均值，其中i＝1,…,N；If the sample matrix is a matrix of 128*L, then all values on each row of the sample matrix are averaged, and the value of the i-th dimension of the preset mean vector is equal to that of the i-th row of the sample matrix average value, where i=1,...,N;

A1033、利用均值向量对样本矩阵进行中心化，得到中心化后的样本矩阵。A1033. Use the mean vector to center the sample matrix to obtain the centered sample matrix.

举例来说，若样本矩阵为一个L*128的矩阵，则所述样本矩阵的每一行上的第i个数值减去预设的均值向量的第i个维度的数值，得到中心化后的样本矩阵,其中i＝1,…,N；For example, if the sample matrix is a matrix of L*128, the i-th value on each row of the sample matrix is subtracted from the value of the i-th dimension of the preset mean vector to obtain the centered sample matrix, where i=1,...,N;

或者，or,

若样本矩阵为一个128*L的矩阵，则所述样本矩阵的每一列上的第i个数值减去预设的均值向量的第i个维度的数值，得到中心化后的样本矩阵，其中i＝1,…,N；If the sample matrix is a 128*L matrix, then the i-th value on each column of the sample matrix is subtracted from the value of the i-th dimension of the preset mean vector to obtain a centered sample matrix, where i =1,...,N;

A1034、计算所述样本矩阵的协方差矩阵。A1034. Calculate the covariance matrix of the sample matrix.

以局部特征描述子为例，得到一个128x128的协方差矩阵。Taking the local feature descriptor as an example, a 128x128 covariance matrix is obtained.

A1035、获取所述协方差矩阵的特征值和与所述特征值对应的特征向量。A1035. Acquire eigenvalues of the covariance matrix and eigenvectors corresponding to the eigenvalues.

举例来说，可以使用现有的特征值分解方法计算协方差矩阵的特征值和与特征值对应的特征向量。For example, the existing eigenvalue decomposition method can be used to calculate the eigenvalues of the covariance matrix and the eigenvectors corresponding to the eigenvalues.

在具体应用中，特征向量的维度与局部特征描述子的维度相等，均为128维。In a specific application, the dimension of the feature vector is equal to the dimension of the local feature descriptor, both of which are 128 dimensions.

A1036、根据所述特征值的大小对所述特征向量进行由大到小排序，选取前K个所述特征向量，K＝32。A1036. According to the size of the eigenvalues, sort the eigenvectors from large to small, and select the first K eigenvectors, where K=32.

A1037、所述前K个特征向量组成所述降维矩阵；A1037. The first K eigenvectors form the dimensionality reduction matrix;

举例来说，前K个特征向量中所有特征向量的所有维度的元素组成所述降维矩阵中行/列的数值；For example, elements of all dimensions of all eigenvectors in the first K eigenvectors form the values of rows/columns in the dimensionality reduction matrix;

也就是说，前K个特征向量构成所述降维矩阵，每一个特征向量的N个元素对应所述降维矩阵的一列或一行。That is to say, the first K eigenvectors constitute the dimensionality reduction matrix, and the N elements of each eigenvector correspond to one column or row of the dimensionality reduction matrix.

特别地，若每一个特征向量的128个元素对应所述降维矩阵的一列，则得到一个128x32的降维矩阵；若每一个特征向量的128个元素对应所述降维矩阵的一行，则得到一个32x128的降维矩阵。In particular, if the 128 elements of each eigenvector correspond to a column of the dimensionality reduction matrix, a 128x32 dimensionality reduction matrix is obtained; if the 128 elements of each eigenvector correspond to a row of the dimensionality reduction matrix, then it is obtained A 32x128 dimensionality reduction matrix.

可选地，在本实施例中，步骤A1031的样本矩阵中每一行的样本数值有N个；步骤A1032中协方差矩阵的维度为N*N；步骤1033中所述特征向量的维度为N；Optionally, in this embodiment, there are N sample values in each row of the sample matrix in step A1031; the dimension of the covariance matrix in step A1032 is N*N; the dimension of the feature vector in step 1033 is N;

其中，N等于128，K等于32；Among them, N is equal to 128, and K is equal to 32;

上述任一实施例中图像数据集至少包括平面物体图像和三维物体图像。优选地，若图像数据集仅包括平面物体图像和三维物体图像，则平面物体图像所占的比例可为80％，三维物体图像所占的比例可为20％。In any of the above embodiments, the image data set includes at least a planar object image and a three-dimensional object image. Preferably, if the image data set only includes plane object images and three-dimensional object images, the proportion of plane object images may be 80%, and the proportion of three-dimensional object images may be 20%.

另外，为较好的说明本实施例中获取低维局部特征描述子的过程，本发明实施例给一个均值向量的具体数值，如下表一所示,其中，预设的均值向量的各个维度的数值按照从左到右的顺序依次写入表一，表一中第一行的第一个数值为预设的均值向量的第一个元素。In addition, in order to better illustrate the process of obtaining low-dimensional local feature descriptors in this embodiment, the embodiment of the present invention gives a specific value of the mean value vector, as shown in Table 1 below, wherein the preset values of each dimension of the mean value vector Values are written in Table 1 in order from left to right, and the first value in the first row in Table 1 is the first element of the preset mean vector.

也就是说，表一中的数值为所述均值向量对应的数值；That is to say, the values in Table 1 are the values corresponding to the mean vector;

其中，表一中数值为所述均值向量的各维度的数值，所述均值向量各维度的数值按照从左到右的顺序依次排列，表一中第一行的第一个数值为预设的均值向量的第一个元素；Wherein, the value in Table 1 is the value of each dimension of the mean value vector, and the values of each dimension of the mean value vector are arranged in order from left to right, and the first value in the first row in Table 1 is the preset the first element of the mean vector;

表一：Table I:

0.0780.0780.0490.0490.0350.0350.0430.0430.0670.0670.0550.0550.050.050.0580.0580.1160.1160.0690.0690.0420.0420.0450.0450.0620.0620.0520.0520.0540.0540.0790.0790.1180.1180.0770.0770.050.050.0490.0490.060.060.0450.0450.0440.0440.0720.0720.0810.0810.0580.0580.0470.0470.0520.0520.0630.0630.0410.0410.0360.0360.0510.0510.0960.0960.0560.0560.0370.0370.0520.0520.0830.0830.0620.0620.050.050.0640.0640.1560.1560.0840.0840.0420.0420.0510.0510.0750.0750.060.060.0530.0530.090.090.1550.1550.0870.0870.050.050.0580.0580.0720.0720.0530.0530.0460.0460.0890.0890.1010.1010.0640.0640.0480.0480.0590.0590.0760.0760.0510.0510.0390.0390.060.060.0960.0960.0630.0630.050.050.0630.0630.0830.0830.0520.0520.0370.0370.0560.0560.1560.1560.090.090.0530.0530.060.060.0750.0750.0510.0510.0420.0420.0850.0850.1550.1550.0880.0880.0460.0460.0530.0530.0730.0730.0570.0570.050.050.0870.0870.1010.1010.0590.0590.0390.0390.0510.0510.0760.0760.0580.0580.0480.0480.0640.0640.0780.0780.0580.0580.050.050.0560.0560.0670.0670.0420.0420.0340.0340.0490.0490.1160.1160.0780.0780.0540.0540.0520.0520.0620.0620.0440.0440.0420.0420.0690.0690.1180.1180.0710.0710.0440.0440.0450.0450.060.060.0490.0490.050.050.0770.0770.0810.0810.0510.0510.0360.0360.0420.0420.0630.0630.0520.0520.0470.0470.0590.059

本发明实施例还给出一个降维矩阵的具体数值，如表二所示，其中，降维矩阵的每一行为32个数值，按照一行一行的顺序写入表二，每一行的数值按照从左到右的顺序依次写入表二，表二中第一行的第一个数值为降维矩阵第一行的第一个元素。The embodiment of the present invention also provides specific values of a dimensionality reduction matrix, as shown in Table 2, wherein, each row of the dimensionality reduction matrix contains 32 values, which are written into Table 2 in the order of one row and one row, and the values of each row are in order from The order from left to right is written in Table 2. The first value in the first row in Table 2 is the first element in the first row of the dimensionality reduction matrix.

也就是说，表二中的元素组成所述降维矩阵，或者，表二中的元素组成所述降维矩阵的转置矩阵；That is to say, the elements in Table 2 form the dimensionality reduction matrix, or, the elements in Table 2 form the transposition matrix of the dimensionality reduction matrix;

其中，所述降维矩阵的每一行为32个数值，表二中数值为所述降维矩阵中一行一行的数值，每一行的数值按照从左到右的顺序依次排列，表二中第一行的第一个数值为所述降维矩阵中第一行的第一个元素；Wherein, each row of the dimensionality reduction matrix has 32 values, and the values in Table 2 are the values of each row in the dimensionality reduction matrix, and the values of each row are arranged in order from left to right, and the first in Table 2 is The first value of the row is the first element of the first row in the dimensionality reduction matrix;

表二：Table II:

上述降维矩阵对任一待处理图像的局部特征描述子进行降维处理，可去除了局部特征描述子中冗余的信息，避免了噪声对局部特征描述子性能的影响，使得由降维后的局部特征描述子聚合得到的Fisher向量在图像检索和匹配中的性能也有提高。The above dimensionality reduction matrix performs dimensionality reduction processing on the local feature descriptor of any image to be processed, which can remove redundant information in the local feature descriptor and avoid the influence of noise on the performance of the local feature descriptor, so that after dimensionality reduction The performance of Fisher vectors obtained by aggregating local feature descriptors in image retrieval and matching is also improved.

上述实施例的低维局部特征描述子聚合成Fisher向量时可具有较低的时间和空间复杂度，使得聚合得到的Fisher向量的维度也会相对较低，减少了对Fisher向量做压缩时所需的空间。The low-dimensional local feature descriptors of the above-mentioned embodiments can have lower time and space complexity when aggregated into Fisher vectors, so that the dimensions of the aggregated Fisher vectors will be relatively low, reducing the time required for compressing the Fisher vectors. Space.

进一步地，上述方法可以在任一终端上实现，尤其可在移动终端上实现。根据现有技术中的无线网络带宽，采用上述实施例中获取的低维局部特征描述子聚合得到的Fisher向量可实现较快传输，提高图像检索或图像分类的响应时间；此外，采用低维局部特征描述子聚合Fisher向量，还可提高Fisher向量的判别力和鲁棒性。Further, the above method can be implemented on any terminal, especially on a mobile terminal. According to the wireless network bandwidth in the prior art, using the Fisher vectors obtained by the aggregation of the low-dimensional local feature descriptors obtained in the above embodiments can achieve faster transmission and improve the response time of image retrieval or image classification; in addition, using low-dimensional local feature descriptors Feature descriptors aggregate Fisher vectors, which can also improve the discriminative power and robustness of Fisher vectors.

本领域普通技术人员可以理解：实现上述各方法实施例的全部或部分步骤可以通过程序指令相关的硬件来完成。前述的程序可以存储于一计算机可读取存储介质中。该程序在执行时，执行包括上述各方法实施例的步骤；而前述的存储介质包括：ROM、RAM、磁碟或者光盘等各种可以存储程序代码的介质。Those of ordinary skill in the art can understand that all or part of the steps for implementing the above method embodiments can be completed by program instructions and related hardware. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above-mentioned method embodiments; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other various media that can store program codes.

最后应说明的是：以上各实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述各实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分或者全部技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: It is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the various embodiments of the present invention. scope.

Claims (10)

1. A method for obtaining a low-dimensional local feature descriptor, comprising:

acquiring a local feature descriptor of an image to be processed;

forming the obtained local feature descriptors into a descriptor set;

reducing the dimension of each local feature descriptor in the descriptor set according to the dimension reduction matrix to obtain a low-dimensional local feature descriptor corresponding to each local feature descriptor; the dimension reduction matrix is obtained by training a preset image data set.

2. The method according to claim 1, wherein performing dimension reduction on each local feature descriptor in the descriptor set according to a dimension reduction matrix to obtain a low-dimensional local feature descriptor corresponding to each local feature descriptor comprises:

subtracting a preset mean vector from each local feature descriptor in the descriptor set to obtain a converted local feature descriptor;

the converted local feature description is sub-grouped into a data matrix;

multiplying the dimensionality reduction matrix and the data matrix to obtain a result matrix;

splitting the result matrix to obtain a low-dimensional local feature descriptor;

the preset mean vector is obtained by training a preset image data set, and the dimensionality of the preset mean vector is the same as that of the local feature descriptor.

3. The method of claim 2, wherein the sub-assembling the converted local feature descriptions into a data matrix comprises:

when the dimensionality of each converted local feature descriptor is N, forming elements on each dimensionality of each local feature descriptor into numerical values on a corresponding row in the data matrix to obtain an M-N dimensional data matrix;

or,

when the dimensionality of each converted local feature descriptor is N, forming elements on each dimensionality of each local feature descriptor into numerical values on a corresponding column in the data matrix to obtain a data matrix with the dimension of N x M;

wherein M is the number of the converted local feature descriptors, M is a natural number, and N is equal to 128.

4. The method according to claim 3, wherein the dimensionality reduction matrix is a matrix obtained from the image dataset by principal component analysis, and the dimensionality of the dimensionality reduction matrix is N x K, or the dimensionality of the dimensionality reduction matrix is K x N;

when the dimensionality of the dimensionality reduction matrix is N x K and the dimensionality of the data matrix is M x N, the dimensionality of the result matrix is M x K; or,

when the dimensionality of the dimensionality reduction matrix is K x N and the dimensionality of the data matrix is N x M, the dimensionality of the result matrix is K x M;

where K is equal to 32.

5. The method of claim 4, wherein the splitting the result matrix to obtain the low-dimensional local feature descriptors comprises:

if the dimension of the result matrix is M x K, extracting the numerical value in each row in the result matrix, and taking the extracted numerical value of each row as a low-dimensional local feature descriptor;

or,

if the dimension of the result matrix is K x M, extracting a numerical value in each column in the result matrix, and taking the extracted numerical value in each column as a low-dimensional local feature descriptor;

wherein M is the number of the converted local feature descriptors, M is a natural number, and K is equal to 32.

6. The method of claim 2, wherein the splitting the result matrix to obtain the low-dimensional local feature descriptors comprises:

extracting a numerical value in each row in the result matrix, and taking the extracted numerical value in each row as a low-dimensional local feature descriptor to obtain M low-dimensional local feature descriptors, wherein the dimension of each low-dimensional local feature descriptor is K;

wherein M is the number of the converted local feature descriptors, M is a natural number, and K is equal to 32;

or,

extracting a numerical value in each column in the result matrix, and taking the extracted numerical value in each column as a low-dimensional local feature descriptor to obtain M low-dimensional local feature descriptors, wherein the dimension of each low-dimensional local feature descriptor is K;

wherein M is the number of the converted local feature descriptors, M is a natural number, and K is equal to 32.

7. The method according to claim 1, wherein before the local feature descriptors in the descriptor set are dimension-reduced according to a dimension-reducing matrix to obtain low-dimensional local feature descriptors, the method further comprises:

obtaining a sample matrix of the image dataset;

obtaining a mean vector according to the sample matrix;

centralizing the sample matrix by using the mean vector to obtain a centralized sample matrix;

calculating a covariance matrix of the centered sample matrix;

acquiring an eigenvalue of the covariance matrix and an eigenvector corresponding to the eigenvalue;

sorting the eigenvectors from big to small according to the magnitude of the eigenvalue, and selecting the first K eigenvectors;

forming the first K eigenvectors into the dimensionality reduction matrix;

where K is equal to 32.

8. The method of claim 7, wherein each local feature descriptor of each image in the image dataset corresponds to a row of values in the sample matrix, each image in the image dataset corresponds to a number of rows of sample values in the sample matrix, and there are N sample values in each row of the sample matrix;

the obtaining of the mean vector according to the sample matrix includes:

averaging all values on each column of the sample matrix, the value of the ith dimension of the mean vector being equal to the average value of the ith column of the sample matrix, where i is 1, …, N;

the centralizing the sample matrix by using the mean vector to obtain a centralized sample matrix includes:

subtracting the value of the ith dimension of the mean vector from the ith numerical value of each row of the sample matrix to obtain a centralized sample matrix, wherein i is 1, …, N;

the dimension of the covariance matrix is N x N;

the dimension of the feature vector is N;

elements of all dimensions of all the feature vectors in the first K feature vectors form numerical values of rows/columns in the dimension reduction matrix;

or,

each local feature descriptor of each image in the image data set corresponds to a column of values in the sample matrix, each image in the image data set corresponds to a plurality of columns of sample values in the sample matrix, and N sample values are provided for each column in the sample matrix;

the obtaining of the mean vector according to the sample matrix includes:

averaging all values in each row of the sample matrix, wherein the value in the ith dimension of the mean vector is equal to the average value in the ith row of the sample matrix, and i is 1, …, N;

the centralizing the sample matrix by using the mean vector to obtain a centralized sample matrix includes:

subtracting the value of the ith dimension of the mean vector from the ith value on each column of the sample matrix to obtain a centralized sample matrix, wherein i is 1, …, N;

the dimension of the covariance matrix is N x N;

the dimension of the feature vector is N;

elements of all dimensions of all the feature vectors in the first K feature vectors form numerical values of columns/rows in the dimension reduction matrix;

where N equals 128 and K equals 32.

9. The method of claim 8, wherein the image dataset comprises:

planar object images and three-dimensional object images.

10. The method according to any one of claims 2 to 9,

the numerical value in the table I is the numerical value corresponding to the mean value vector;

the numerical values in the first table are numerical values of all dimensions of the mean vector, the numerical values of all dimensions of the mean vector are sequentially arranged from left to right, and the first numerical value in the first row in the first table is a first element of a preset mean vector;

table one:

0.0780.0490.0350.0430.0670.0550.050.0580.1160.0690.0420.0450.0620.0520.0540.0790.1180.0770.050.0490.060.0450.0440.0720.0810.0580.0470.0520.0630.0410.0360.0510.0960.0560.0370.0520.0830.0620.050.0640.1560.0840.0420.0510.0750.060.0530.090.1550.0870.050.0580.0720.0530.0460.0890.1010.0640.0480.0590.0760.0510.0390.060.0960.0630.050.0630.0830.0520.0370.0560.1560.090.0530.060.0750.0510.0420.0850.1550.0880.0460.0530.0730.0570.050.0870.1010.0590.0390.0510.0760.0580.0480.0640.0780.0580.050.0560.0670.0420.0340.0490.1160.0780.0540.0520.0620.0440.0420.0690.1180.0710.0440.0450.060.0490.050.0770.0810.0510.0360.0420.0630.0520.0470.059

elements in the second table form the dimension reduction matrix, or elements in the second table form a transpose matrix of the dimension reduction matrix;

the numerical values in the second table are numerical values in a row and a column in the dimension reduction matrix, the numerical values in each row are sequentially arranged from left to right, and the first numerical value in the first row in the second table is the first element in the first row in the dimension reduction matrix;

table two: