WO2017045113A1 - 一种基于局部pca白化的图像表示方法和处理装置 - Google Patents

一种基于局部pca白化的图像表示方法和处理装置 Download PDF

Info

Publication number
WO2017045113A1
WO2017045113A1 PCT/CN2015/089576 CN2015089576W WO2017045113A1 WO 2017045113 A1 WO2017045113 A1 WO 2017045113A1 CN 2015089576 W CN2015089576 W CN 2015089576W WO 2017045113 A1 WO2017045113 A1 WO 2017045113A1
Authority
WO
WIPO (PCT)
Prior art keywords
image representation
feature
vlad
representation vector
projection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2015/089576
Other languages
English (en)
French (fr)
Inventor
王文敏
镇明敏
王荣刚
李革
董胜富
王振宇
李英
高文
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Peking University Shenzhen Graduate School
Original Assignee
Peking University Shenzhen Graduate School
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Peking University Shenzhen Graduate School filed Critical Peking University Shenzhen Graduate School
Priority to PCT/CN2015/089576 priority Critical patent/WO2017045113A1/zh
Priority to US15/756,193 priority patent/US10424052B2/en
Publication of WO2017045113A1 publication Critical patent/WO2017045113A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/37—Determination of transform parameters for the alignment of images, i.e. image registration using transform domain methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations

Definitions

  • the present application relates to the field of image processing, and in particular to an image representation method and a processing apparatus based on localized PCA (Principal Component Analysis) whitening.
  • PCA Principal Component Analysis
  • image representation is a very basic content. Whether for image classification, image retrieval or object recognition, an abstract representation of the image is required.
  • VLAD vector of indigenous aggregated descriptors, local feature aggregation descriptors
  • c k is the word in the dictionary.
  • a local feature is first adopted.
  • SIFT Scale-invariant feature transform
  • the present application provides an image representation method based on local PCA whitening, including:
  • Constructing a dictionary assigning each feature to a corresponding word, mapping the words and features to a high dimensional space, the dimension of the high dimensional space being higher than the dimension of the current space of the words and features;
  • the features obtained by the projection transformation are normalized to obtain the final image representation vector.
  • the present application provides an image representation processing apparatus based on local PCA whitening, including:
  • a first mapping module configured to construct a dictionary, assign each feature to a corresponding word, and map the word and the feature to a high-dimensional space, the dimension of the high-dimensional space being higher than a dimension of a current space of the word and the feature;
  • a principal component analysis module configured to perform principal component analysis in each corresponding word space to obtain a projection matrix
  • a VLAD calculation module configured to calculate a VLAD image representation vector according to the dictionary
  • a second mapping module configured to map the VLAD image representation vector into the high dimensional space
  • a projection transformation module configured to perform projection transformation on the projected VLAD image representation vector according to the projection matrix
  • the normalization processing module is configured to normalize the features obtained by the projection transformation to obtain a final image representation vector.
  • the local PCA whitening-based image representation method and processing apparatus provided by the present application, wherein the first mapping module constructs a dictionary, assigns each feature to a corresponding word, and maps words and features to a high-dimensional space; the principal component analysis module Principal component analysis is performed in each corresponding word space to obtain a projection matrix; the VLAD calculation module calculates a VLAD image representation vector according to the dictionary; and the second mapping module maps the VLAD image representation vector into the high dimensional space;
  • the projection transformation module performs projection transformation on the projected VLAD image representation vector according to the projection matrix; the normalization processing module normalizes the features obtained by the projection transformation to obtain a final image representation vector.
  • the obtained image representation vector first project it into a high-dimensional space, and then perform projection transformation on each word corresponding vector for the projection matrix calculated in advance to obtain a low-dimensional vector, so that each word corresponds to The vectors are consistent.
  • the method and processing device have better robustness and higher performance.
  • FIG. 1 is a block diagram of an image representation processing apparatus based on local PCA whitening in an embodiment of the present application
  • FIG. 2 is a schematic flow chart of an image representation method based on local PCA whitening in an embodiment of the present application
  • 3 is a schematic diagram showing the distribution of features in different word spaces generated by K-means clustering
  • Figure 4 is a comparison of different methods for different dictionary sizes on the Holidays data set
  • Figure 5 is a comparison of different methods of different dictionary sizes on the UKbench data set
  • Figure 6 is a comparison of the use of local PCA whitening and no local PCA whitening at different dictionary sizes on the Holidays data set;
  • Figure 7 is a comparison of the use of local PCA whitening and no local PCA whitening at different dictionary sizes on the UKbench data set.
  • This embodiment provides an image representation method and processing apparatus based on local PCA whitening.
  • the image representation processing apparatus based on local PCA whitening includes a first mapping module 101, a principal component analysis module 102, a VLAD calculation module 103, a second mapping module 104, a projection transformation module 105, and a normalization processing module 106.
  • the first mapping module 101 is configured to construct a dictionary, assign each feature to a corresponding word, and map the words and features to a high dimensional space, the dimension of the high dimensional space being higher than the dimension of the current space of the words and features.
  • the principal component analysis module 102 is configured to perform principal component analysis in each corresponding word space to obtain a projection matrix.
  • the VLAD calculation module 103 is configured to calculate a VLAD image representation vector from the dictionary.
  • the second mapping module 104 is for mapping the VLAD image representation vector into the high dimensional space.
  • the projection transformation module 105 is configured to perform projection transformation on the projected VLAD image representation vector according to the projection matrix.
  • the normalization processing module 106 is configured to normalize the features obtained by the projection transformation to obtain a final image representation vector.
  • the image representation method based on local PCA whitening includes the following steps:
  • Step 1.1 The first mapping module 101 constructs a dictionary, assigns each feature to a corresponding word, and maps words and features to a high dimensional space.
  • the dimension of the high dimensional space is higher than the dimension of the current space of words and features.
  • the K-means algorithm is used to construct a dictionary, and each training feature is assigned to the nearest word, and the words and features are explicitly mapped to a high dimensional space.
  • the high dimensional space may be a 3 times dimension of a dimension of a current space of words and features.
  • the principal component analysis module 102 also performs principal component analysis in each corresponding word space to obtain a projection matrix.
  • the projection matrix is calculated by the following method:
  • D is the feature dimension.
  • the eigenvalue eigval(G i ) and the eigenvector eigvect(G i ) are obtained by eigen-decomposing the matrix G i according to the following equation, and are arranged in descending order of eigenvalues.
  • Step 1.2 The VLAD calculation module 103 calculates the original VLAD image representation vector x according to the dictionary constructed in step 1.1.
  • step 1.3 the original VLAD image representation vector x is obtained using the VLAD image representation method of the prior art.
  • Step 1.4 The second mapping module 104 maps the VLAD image representation vector into the high dimensional space.
  • mapping is performed according to the following formula:
  • ⁇ represents the index of the map.
  • the specific mapping method can be found in the following literature: A. Vedaldi and A. Zisserman, "Efficient additive kernels via explicitfeature maps,” IEEE Trans. Pattern Anal. Mach. Intell., 2012.
  • step 1.3 uses an implicitly mapped index.
  • step 1.1 when mapping words and features to a high dimensional space, the methods mentioned in the above mentioned documents may also be employed, but explicit mapping may be employed.
  • Step 1.4 Perform a projection transformation on the VLAD image representation vector obtained by the projection in step 1.4 according to the obtained projection matrix.
  • the projection transformation is performed according to the following formula to obtain the feature y,
  • Step 1.5 Normalize the features obtained by the projection transformation to obtain the final image representation vector.
  • the projected feature y is normalized by the second normal form to obtain the final image representation vector.
  • the image representation method based on local PCA whitening it can be used for image retrieval.
  • the corresponding image representation method is obtained for each picture, and then the similarity is compared according to the similarity between each picture in the picture to be retrieved and the database. The size of the arrangement can be retrieved.
  • the similarity is calculated as the cosine between the representation vectors of the two pictures. It can be seen from Fig. 3 that in the different word spaces generated by K-means clustering, the feature distribution is disorderly inconsistent, so it is necessary to whiten PCA in each word space, that is, local PCA whitening.
  • principal component analysis is performed in each corresponding word space to obtain a projection matrix, so that the method and the processing device have better robustness.
  • FIG. 4 is a comparison result of different methods of different dictionary sizes on the Holidays data set
  • FIG. 5 is a comparison result of different methods of different dictionary sizes on the UKbench data set.
  • SVLAD represents the standard VLAD method
  • HVLAD represents the VLAD method after mapping to high dimensions
  • VLAD + RPCAW is the method provided in this embodiment
  • FIG. 6 is a comparison result of using local PCA whitening (RPCAW) and not using local PCA whitening (SVLAD) in different dictionary sizes on the Holidays data set
  • FIG. 7 is a different dictionary on the UKbench data set. Comparison results using local PCA bleaching (RPCAW) and no local PCA bleaching (SVLAD) were used.
  • RPCAW local PCA bleaching
  • SVLAD local PCA bleaching
  • the local PCA whitening-based image representation method and processing device provided by the embodiments of the present application firstly project the vector into a high-dimensional space for the obtained image representation vector. Then, for the projection matrix calculated in advance, the corresponding vector of each word is subjected to projection transformation to obtain a low-dimensional vector, so that the vector corresponding to each word is consistent.
  • the method and processing device have better robustness and higher performance.
  • the program may be stored in a computer readable storage medium, and the storage medium may include: a read only memory. Random access memory, disk or optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种基于局部PCA白化的图像表示方法和处理装置，其中，第一映射模块将单词和特征映射到一个高维度空间；主成分分析模块在对应的每个单词空间中进行主成分分析，得到投影矩阵；VLAD计算模块计算VLAD图像表示向量；第二映射模块将所述VLAD图像表示向量映射到所述高维度空间中；投影变换模块将投影得到的VLAD图像表示向量进行投影变换；归一化处理模块将投影变换得到的特征进行归一化，得到最后的图像表示向量。对于得到的图像表示向量，首先将其投影到一个高维度空间中，然后对提前计算好的投影矩阵，对每个单词对应向量进行投影变换，得到一个低维度的向量，这样使得每个单词对应的向量是一致的。该方法和处理装置具有更好的鲁棒性和更高的性能。

Description

一种基于局部PCA白化的图像表示方法和处理装置 技术领域
本申请涉及图像处理领域，具体涉及一种基于局部PCA(Principal Component Analysis，主成分分析)白化的图像表示方法和处理装置。
背景技术
在计算机视觉研究中，图像表示是一个非常基本的内容。无论是对于图像分类、图像检索或是物体识别，都需要对图像做一个抽象的表示。目前，VLAD(vector of locally aggregated descriptors，局部特征聚合描述符)图像表示方法在很多研究中得到了运用。
对于原始的VLAD方法，首先在一个数据集上用K-means算法构建词典：
C＝{c1，c2，...，ck}
其中，ck为词典中的单词。对于每张图片，首先采取局部特征，通常采用SIFT(Scale-invariant feature transform，尺度不变特征变换)特征，得到该图片对应的一个特征集合：
I＝{x1，x2，...，xm}
其中，xm为特征集合中的特征。然后计算每个特征到词典中单词的距离，并将该特征赋给最近的单词。最后，对于每个单词对应的所有特征，做如下计算：
Figure PCTCN2015089576-appb-000001
其中q(x)＝ci表示特征x最近的单词是ci，vi则是第i个单词对应的向量，将所有单词对应向量连接起来就得到了最后的VLAD表示。
然而，对于VLAD图像表示方法，对于每个单词对应的特征冗余以及噪声消除问题一直尚未解决，其性能有待提高。
发明内容
根据本申请的第一方面，本申请提供了一种基于局部PCA白化的图像表示方法，包括：
构造词典，将每个特征赋给相应的单词，将单词和特征映射到一个高维度空间，所述高维度空间的维度高于单词和特征的当前空间的维度；
在对应的每个单词空间中进行主成分分析，得到投影矩阵；
根据所述词典计算VLAD图像表示向量；
将所述VLAD图像表示向量映射到所述高维度空间中；
根据所述投影矩阵将投影得到的VLAD图像表示向量进行投影变换；
将投影变换得到的特征进行归一化，得到最后的图像表示向量。
根据本申请的第二方面，本申请提供了一种基于局部PCA白化的图像表示处理装置，包括：
第一映射模块，用于构造词典，将每个特征赋给相应的单词，将单词和特征映射到一个高维度空间，所述高维度空间的维度高于单词和特征的当前空间的维度；
主成分分析模块，用于在对应的每个单词空间中进行主成分分析，得到投影矩阵；
VLAD计算模块，用于根据所述词典计算VLAD图像表示向量；
第二映射模块，用于将所述VLAD图像表示向量映射到所述高维度空间中；
投影变换模块，用于根据所述投影矩阵将投影得到的VLAD图像表示向量进行投影变换；
归一化处理模块，用于将投影变换得到的特征进行归一化，得到最后的图像表示向量。
本申请提供的基于局部PCA白化的图像表示方法和处理装置，其中，第一映射模块构造词典，将每个特征赋给相应的单词，将单词和特征映射到一个高维度空间；主成分分析模块在对应的每个单词空间中进行主成分分析，得到投影矩阵；VLAD计算模块根据所述词典计算VLAD图像表示向量；第二映射模块将所述VLAD图像表示向量映射到所述高维度空间中；投影变换模块根据所述投影矩阵将投影得到的VLAD图像表示向量进行投影变换；归一化处理模块将投影变换得到的特征进行归一化，得到最后的图像表示向量。对于得到的图像表示向量，首先将其投影到一个高维度空间中，然后对提前计算好的投影矩阵，对每个单词对应向量进行投影变换，得到一个低维度的向量，这样使得每个单词对应的向量是一致的。该方法和处理装置具有更好的鲁棒性和更高的性能。
附图说明
图1为本申请一种实施例中基于局部PCA白化的图像表示处理装置的模块示意图；
图2为本申请一种实施例中基于局部PCA白化的图像表示方法的流程示意图；
图3为在K-means聚类产生的不同单词空间中，特征的分布示意图；
图4为在Holidays数据集上不同词典大小的不同方法的比较结果；
图5为在UKbench数据集上不同词典大小的不同方法的比较结果；
图6为在Holidays数据集上不同词典大小下使用局部PCA白化和未使用局部PCA白化的比较结果；
图7为在UKbench数据集上不同词典大小下使用局部PCA白化和未使用局部PCA白化的比较结果。
具体实施方式
下面通过具体实施方式结合附图对本申请作进一步详细说明。
本实施例提供了一种基于局部PCA白化的图像表示方法和处理装置。
请参考图1，基于局部PCA白化的图像表示处理装置包括第一映射模块101、主成分分析模块102、VLAD计算模块103、第二映射模块104、投影变换模块105和归一化处理模块106。
第一映射模块101用于构造词典，将每个特征赋给相应的单词，将单词和特征映射到一个高维度空间，高维度空间的维度高于单词和特征的当前空间的维度。
主成分分析模块102用于在对应的每个单词空间中进行主成分分析，得到投影矩阵。
VLAD计算模块103用于根据词典计算VLAD图像表示向量。
第二映射模块104用于将VLAD图像表示向量映射到高维度空间中。
投影变换模块105用于根据投影矩阵将投影得到的VLAD图像表示向量进行投影变换。
归一化处理模块106用于将投影变换得到的特征进行归一化，得到最后的图像表示向量。
为了更好地对本申请进行说明，下面将基于局部PCA白化的图像表示方法及其处理装置相结合，对本申请进行说明。
请参考图2，基于局部PCA白化的图像表示方法包括下面步骤：
步骤1.1：第一映射模块101构造词典，将每个特征赋给相应的单词，将单词和特征映射到一个高维度空间。高维度空间的维度高于单词和特征的当前空间的维度。
本实施例中，利用K-means算法构造词典，将每个训练所用特征赋给距离最近的单词，将单词以及特征都显式映射到一个高维度空间。具体的，该高维度空间可以是单词和特征的当前空间的维度的3倍维度。
步骤1.1中，主成分分析模块102还在对应的每个单词空间中进行主成分分析，得到投影矩阵。
本实施例中，投影矩阵通过下面方式计算得到：
首先根据下面式子计算过渡矩阵Gi，
Figure PCTCN2015089576-appb-000002
其中ci是第i个单词，x为赋给该单词的所有特征，D为特征维度。当选择SIFT算法进行特征描述时，D通常为128。
根据下面式子对矩阵Gi做特征分解，得到特征值eigval(Gi)和特征向量eigvect(Gi)，按照特征值从大到小的顺序排列。
Figure PCTCN2015089576-appb-000003
Figure PCTCN2015089576-appb-000004
根据下面式子计算投影矩阵
Figure PCTCN2015089576-appb-000005
Figure PCTCN2015089576-appb-000006
其中，
Figure PCTCN2015089576-appb-000007
Figure PCTCN2015089576-appb-000008
ε和t为预设参数，例如设置ε＝0.00001，t属于特征维度，可以根据实际应用进行调整。
步骤1.2：VLAD计算模块103根据步骤1.1构造的词典计算原始的VLAD图像表示向量x。步骤1.3中采用现有技术中的VLAD图像表示方法得到原始的VLAD图像表示向量x。
步骤1.4：第二映射模块104将VLAD图像表示向量映射到高维度空间中。本实施例中，根据下面式子进行映射：
Figure PCTCN2015089576-appb-000009
其中，τ表示映射的索引。具体该映射方法可参见下面文献：A. Vedaldi and A.Zisserman,“Efficient additive kernels via explicitfeature maps,”IEEE Trans.Pattern Anal.Mach.Intell.,2012。
步骤1.3中，τ采用隐式映射的索引。在步骤1.1中，将单词和特征映射到高维度空间时，也可以采用上述文献提及的方法，但可以采用显式映射。
步骤1.4：根据得到的投影矩阵将步骤1.4投影得到的VLAD图像表示向量进行投影变换。
本实施例中，根据下面公式进行投影变换，以得到特征y，
Figure PCTCN2015089576-appb-000010
步骤1.5：将投影变换得到的特征进行归一化，得到最后的图像表示向量。本实施例中，将投影得到的特征y进行第二范式归一化，以得到了最后的图像表示向量。
对于基于局部PCA白化的图像表示方法，可以用于图像检索这个任务，对每张图片得到其对应的图像表示方法，然后将待检索图片和数据库中的每张图片进行相似度比较，按照相似度的大小排列可得到检索结果。其中相似度的计算为两张图片的表示向量之间的余弦值。由图3可以看出，在K-means聚类产生的不同单词空间中，其特征分布是杂乱不一致的，因此将PCA白化作用于每个单词空间，即局部PCA白化，是有必要的。本申请实施例中，在对应的每个单词空间中进行主成分分析，以得到投影矩阵，使得该方法和处理装置具有更好的鲁棒性。
请参考图4和图5，图4为在Holidays数据集上不同词典大小的不同方法的比较结果，图5为在UKbench数据集上不同词典大小的不同方法的比较结果。在图4和图5中，比较了对于不同词典大小不同方法的性能(SVLAD表示标准VLAD方法，HVLAD表示映射为高维度后的VLAD方法，VLAD+RPCAW为本实施例提供的方法)，可以看到基于局部PCA白化的图像表示方法比其他方法性能更好。
请参考图6和图7，图6为在Holidays数据集上不同词典大小下使用局部PCA白化(RPCAW)和未使用局部PCA白化(SVLAD)的比较结果，图7为在UKbench数据集上不同词典大小下使用局部PCA白化(RPCAW)和未使用局部PCA白化(SVLAD)的比较结果。在图6和图7中，比较了对于不同词典大小下使用局部PCA白化和未使用局部PCA白化的性能。可以看到将基于局部PCA白化的图像表示方法能够进一步提高性能。
本申请实施例提供的基于局部PCA白化的图像表示方法和处理装置，对于得到的图像表示向量，首先将其投影到一个高维度空间中，然 后对提前计算好的投影矩阵，对每个单词对应向量进行投影变换，得到一个低维度的向量，这样使得每个单词对应的向量是一致的。该方法和处理装置具有更好的鲁棒性和更高的性能。
本领域技术人员可以理解，上述实施方式中各种方法的全部或部分步骤可以通过程序来控制相关硬件完成，该程序可以存储于一计算机可读存储介质中，存储介质可以包括：只读存储器、随机存取存储器、磁盘或光盘等。
以上内容是结合具体的实施方式对本申请所作的进一步详细说明，不能认定本申请的具体实施只局限于这些说明。对于本申请所属技术领域的普通技术人员来说，在不脱离本申请发明构思的前提下，还可以做出若干简单推演或替换。

Claims (10)

  1. 一种基于局部PCA白化的图像表示方法，其特征在于，包括：
    构造词典，将每个特征赋给相应的单词，将单词和特征映射到一个高维度空间，所述高维度空间的维度高于单词和特征的当前空间的维度；
    在对应的每个单词空间中进行主成分分析，得到投影矩阵；
    根据所述词典计算VLAD图像表示向量；
    将所述VLAD图像表示向量映射到所述高维度空间中；
    根据所述投影矩阵将投影得到的VLAD图像表示向量进行投影变换；
    将投影变换得到的特征进行归一化，得到最后的图像表示向量。
  2. 如权利要求1所述的方法，其特征在于，利用K-means算法构造所述词典，并将每个特征赋给距离最近的单词；
    将投影变换得到的特征进行第二范式归一化，得到最后的图像表示向量。
  3. 如权利要求1或2所述的方法，其特征在于，在对应的每个单词空间中进行主成分分析，得到投影矩阵，具体为：
    根据下面式子计算过渡矩阵Gi，
    Figure PCTCN2015089576-appb-100001
    其中ci是第i个单词，x为赋给该单词的所有特征，D为特征维度；
    根据下面式子对矩阵Gi做特征分解，得到特征值eigval(Gi)和特征向量eigvect(Gi)；
    Figure PCTCN2015089576-appb-100002
    Figure PCTCN2015089576-appb-100003
    根据下面式子计算投影矩阵
    Figure PCTCN2015089576-appb-100004
    Figure PCTCN2015089576-appb-100005
    其中，
    Figure PCTCN2015089576-appb-100006
    Figure PCTCN2015089576-appb-100007
    ε和t为预设参数。
  4. 如权利要求3所述的方法，其特征在于，将所述VLAD图像表示向量映射到所述高维度空间中，具体为根据下面式子进行映射：
    Figure PCTCN2015089576-appb-100008
    其中，τ表示映射的索引。
  5. 如权利要求3所述的方法，其特征在于，根据所述投影矩阵将投影得到的VLAD图像表示向量进行投影变换，具体为：根据下面公式进行投影变换，以得到特征y，
    Figure PCTCN2015089576-appb-100009
  6. 一种基于局部PCA白化的图像表示处理装置，其特征在于，包括：
    第一映射模块，用于构造词典，将每个特征赋给相应的单词，将单词和特征映射到一个高维度空间，所述高维度空间的维度高于单词和特征的当前空间的维度；
    主成分分析模块，用于在对应的每个单词空间中进行主成分分析，得到投影矩阵；
    VLAD计算模块，用于根据所述词典计算VLAD图像表示向量；
    第二映射模块，用于将所述VLAD图像表示向量映射到所述高维度空间中；
    投影变换模块，用于根据所述投影矩阵将投影得到的VLAD图像表示向量进行投影变换；
    归一化处理模块，用于将投影变换得到的特征进行归一化，得到最后的图像表示向量。
  7. 如权利要求6所述的处理装置，其特征在于，第一映射模块用于利用K-means算法构造所述词典，并将每个特征赋给距离最近的单词；
    归一化处理模块用于将投影变换得到的特征进行第二范式归一化，得到最后的图像表示向量。
  8. 如权利要求6或7所述的处理装置，其特征在于，主成分分析模块用于在对应的每个单词空间中进行主成分分析，得到投影矩阵时：
    主成分分析模块用于根据下面式子计算过渡矩阵Gi，
    Figure PCTCN2015089576-appb-100010
    其中ci是第i个单词，x为赋给该单词的所有特征，D为特征维度；
    主成分分析模块还用于根据下面式子对矩阵Gi做特征分解，得到特征值eigval(Gi)和特征向量eigvect(Gi)；
    Figure PCTCN2015089576-appb-100011
    Figure PCTCN2015089576-appb-100012
    主成分分析模块还用于根据下面式子计算投影矩阵
    Figure PCTCN2015089576-appb-100013
    Figure PCTCN2015089576-appb-100014
    其中，
    Figure PCTCN2015089576-appb-100015
    Figure PCTCN2015089576-appb-100016
    ε和t为预设参数。
  9. 如权利要求8所述的处理装置，其特征在于，VLAD计算模块用于根据所述词典计算VLAD图像表示向量时，根据下面式子进行映射：
    Figure PCTCN2015089576-appb-100017
    其中，τ表示映射的索引。
  10. 如权利要求8所述的处理装置，其特征在于，投影变换模块用于根据所述投影矩阵将投影得到的VLAD图像表示向量进行投影变换时，根据下面公式进行投影变换，以得到特征y，
    Figure PCTCN2015089576-appb-100018
PCT/CN2015/089576 2015-09-15 2015-09-15 一种基于局部pca白化的图像表示方法和处理装置 Ceased WO2017045113A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2015/089576 WO2017045113A1 (zh) 2015-09-15 2015-09-15 一种基于局部pca白化的图像表示方法和处理装置
US15/756,193 US10424052B2 (en) 2015-09-15 2015-09-15 Image representation method and processing device based on local PCA whitening

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2015/089576 WO2017045113A1 (zh) 2015-09-15 2015-09-15 一种基于局部pca白化的图像表示方法和处理装置

Publications (1)

Publication Number Publication Date
WO2017045113A1 true WO2017045113A1 (zh) 2017-03-23

Family

ID=58288282

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/089576 Ceased WO2017045113A1 (zh) 2015-09-15 2015-09-15 一种基于局部pca白化的图像表示方法和处理装置

Country Status (2)

Country Link
US (1) US10424052B2 (zh)
WO (1) WO2017045113A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11301509B2 (en) * 2017-01-20 2022-04-12 Rakuten Group, Inc. Image search system, image search method, and program

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050058339A1 (en) * 2003-09-16 2005-03-17 Fuji Xerox Co., Ltd. Data recognition device
CN103020265A (zh) * 2012-12-25 2013-04-03 深圳先进技术研究院 图像检索的方法和系统
CN103246880A (zh) * 2013-05-15 2013-08-14 中国科学院自动化研究所 基于多阶局部显著模式特征统计的人脸识别方法
CN103838864A (zh) * 2014-03-20 2014-06-04 北京工业大学 一种视觉显著性与短语相结合的图像检索方法
CN104408479A (zh) * 2014-11-28 2015-03-11 电子科技大学 一种基于深度局部特征描述符的海量图像分类方法
CN105205497A (zh) * 2015-09-15 2015-12-30 北京大学深圳研究生院 一种基于局部pca白化的图像表示方法和处理装置

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015153480A1 (en) * 2014-03-31 2015-10-08 Ebay Inc. Image-based retrieval and searching
US20160217157A1 (en) * 2015-01-23 2016-07-28 Ebay Inc. Recognition of items depicted in images

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050058339A1 (en) * 2003-09-16 2005-03-17 Fuji Xerox Co., Ltd. Data recognition device
CN103020265A (zh) * 2012-12-25 2013-04-03 深圳先进技术研究院 图像检索的方法和系统
CN103246880A (zh) * 2013-05-15 2013-08-14 中国科学院自动化研究所 基于多阶局部显著模式特征统计的人脸识别方法
CN103838864A (zh) * 2014-03-20 2014-06-04 北京工业大学 一种视觉显著性与短语相结合的图像检索方法
CN104408479A (zh) * 2014-11-28 2015-03-11 电子科技大学 一种基于深度局部特征描述符的海量图像分类方法
CN105205497A (zh) * 2015-09-15 2015-12-30 北京大学深圳研究生院 一种基于局部pca白化的图像表示方法和处理装置

Also Published As

Publication number Publication date
US10424052B2 (en) 2019-09-24
US20180240217A1 (en) 2018-08-23

Similar Documents

Publication Publication Date Title
Arandjelovic et al. All about VLAD
EP3084682B1 (en) System and method for identifying faces in unconstrained media
Karanwal et al. Neighborhood and center difference-based-LBP for face recognition
WO2018149133A1 (zh) 基于核非负矩阵分解的字典学习和稀疏特征表示的人脸识别方法及系统
CN109255392B (zh) 基于非局部神经网络的视频分类方法、装置及设备
US8428397B1 (en) Systems and methods for large scale, high-dimensional searches
CN101615248B (zh) 年龄估计方法、设备和人脸识别系统
US20150030238A1 (en) Visual pattern recognition in an image
Zhang et al. Pairwise sparsity preserving embedding for unsupervised subspace learning and classification
CN109754018B (zh) 一种基于f范数的低秩局部保持投影的图像识别方法
CN105518668A (zh) 基于内容的图像检索
CN107480636B (zh) 基于核非负矩阵分解的人脸识别方法、系统及存储介质
CN107609596B (zh) 无参数自动加权多图正则化非负矩阵分解及图像聚类方法
CN111191719A (zh) 一种基于自表示和图谱约束的非负矩阵分解的图像聚类方法
Wang et al. Two-stage content based image retrieval using sparse representation and feature fusion
Jia et al. Heavy-tailed distances for gradient based image descriptors
CN109978064A (zh) 基于图像集的李群字典学习分类方法
Hosny et al. Face recognition using exact Gaussian-Hermit moments
CN105205497B (zh) 一种基于局部pca白化的图像表示方法和处理装置
Zhang et al. Discriminative sparse representation for face recognition
Cheng et al. A fast and robust face recognition approach combining Gabor learned dictionaries and collaborative representation
WO2017045113A1 (zh) 一种基于局部pca白化的图像表示方法和处理装置
CN105787423B (zh) 人脸图像的属性信息识别方法和装置
CN107506744B (zh) 基于局部线性表示保留鉴别嵌入的人脸识别方法
CN108229552B (zh) 一种模型处理方法、装置及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15903804

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 15756193

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 23/07/2018)

122 Ep: pct application non-entry in european phase

Ref document number: 15903804

Country of ref document: EP

Kind code of ref document: A1