WO2020199773A1 - 图像检索方法及装置和计算机可读存储介质 - Google Patents

图像检索方法及装置和计算机可读存储介质 Download PDF

Info

Publication number
WO2020199773A1
WO2020199773A1 PCT/CN2020/075685 CN2020075685W WO2020199773A1 WO 2020199773 A1 WO2020199773 A1 WO 2020199773A1 CN 2020075685 W CN2020075685 W CN 2020075685W WO 2020199773 A1 WO2020199773 A1 WO 2020199773A1
Authority
WO
WIPO (PCT)
Prior art keywords
cluster
image
vector
local visual
sub
Prior art date
Application number
PCT/CN2020/075685
Other languages
English (en)
French (fr)
Inventor
马福强
陈丽莉
张�浩
孙建康
董泽华
吕耀宇
Original Assignee
京东方科技集团股份有限公司
北京京东方光电科技有限公司
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 京东方科技集团股份有限公司, 北京京东方光电科技有限公司 filed Critical 京东方科技集团股份有限公司
Publication of WO2020199773A1 publication Critical patent/WO2020199773A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques

Definitions

  • the present disclosure relates to the field of image processing technology, and in particular to an image retrieval method and device, electronic equipment, and computer-readable storage media.
  • image retrieval retrieves a list of images similar to the image to be recognized and corresponding content such as text, video or web page link, which has extremely promising application prospects.
  • the existing related image retrieval algorithm workflow includes: extracting the visual features of the image to be recognized; constructing a feature index or image expression; making similarity judgments based on the image features or the distance of the image expression (such as Euclidean distance, cosine distance, etc.); and Give a list of similar images.
  • an image retrieval method including:
  • a list of candidate images whose similarity with the expression vector exceeds the similarity threshold is retrieved from the candidate image library.
  • determining the expression vector of the image to be recognized based on the at least one local visual feature includes:
  • the image classification quantizer determines the cluster and subclass to which each of the at least one local visual feature belongs, and the image classification quantizer includes K1 clusters and corresponding K1 cluster centers, and the The K2 sub-categories cut out by each of the K1 clusters and the distance threshold of each sub-category; and
  • the accumulated sum vector is the expression vector of the image to be recognized, where K1 And K2 are positive integers.
  • determining the cluster and subcategory to which each of the at least one local visual feature belongs includes:
  • Determining that the cluster cluster where the cluster center closest to the local visual feature is located is the cluster cluster to which the local visual feature belongs;
  • the image retrieval method further includes:
  • the dimensionality reduction algorithm is used to perform dimensionality reduction processing on the long vector to obtain a dimensionality-reduced long vector, and the dimensionality-reduced long vector is the expression vector of the image to be recognized.
  • the image classification quantizer is trained through the following steps, including:
  • the distance between the local visual feature in the cluster and the cluster center of the cluster and the maximum distance value are obtained; based on the maximum distance value, the cluster is divided into K2 sub-categories, get the distance threshold of each sub-category,
  • the number of obtained distance thresholds is K1*K2, and the sub-categories in the same cluster do not overlap each other.
  • the cluster cluster is divided into K2 sub-categories based on the maximum distance value, and the distance boundary threshold of each sub-category is obtained, including:
  • the maximum distance value is divided into K2 segments, the local visual features corresponding to each segment constitute a sub-category, and K2 sub-categories are obtained; and the maximum distance between each segment and the cluster center of the cluster The distance is the distance threshold of the corresponding sub-category of each segment.
  • the cluster analysis algorithm includes the K-means algorithm.
  • an image retrieval device including:
  • the memory is coupled to the processor, and stores executable instructions, which when executed by the processor enable the processor to be configured as:
  • a list of candidate images whose similarity with the expression vector exceeds the similarity threshold is retrieved from the candidate image library.
  • the processor is further configured to:
  • the image classification quantizer determines the cluster and subclass to which each of the at least one local visual feature belongs, and the image classification quantizer includes K1 clusters and corresponding K1 cluster centers, and the The K2 sub-categories cut out by each of the K1 clusters and the distance threshold of each sub-category; and
  • the cumulative sum vector is the expression vector of the image to be recognized, where K1 and K2 are positive integers.
  • the processor is further configured to:
  • Determining that the cluster cluster where the cluster center closest to the local visual feature is located is the cluster cluster to which the local visual feature belongs;
  • the processor is further configured to:
  • the dimensionality reduction algorithm is used to perform dimensionality reduction processing on the long vector to obtain a dimensionality-reduced long vector, and the dimensionality-reduced long vector is the expression vector of the image to be recognized.
  • the image classification quantizer is trained through the following steps, including:
  • the number of obtained distance thresholds is K1*K2, and the sub-categories in the same cluster do not overlap each other.
  • a computer-readable storage medium having computer-executable instructions stored thereon, which is characterized in that, when the instructions are executed by a processor, the instructions in any one of claims 1 to 7 are implemented. The steps of the method.
  • Fig. 1 is a block diagram of an image retrieval method according to an embodiment of the present disclosure
  • FIG. 2 is a flowchart of obtaining an expression vector according to an embodiment of the present disclosure
  • FIG. 3 is a flowchart of obtaining a distance threshold according to an embodiment of the present disclosure
  • FIG. 4 is a flowchart of obtaining cluster cluster sub-categories according to an embodiment of the present disclosure
  • FIG. 5 is a schematic diagram showing the effect of dividing K1 clusters and K2 sub-categories according to an embodiment of the present disclosure
  • Fig. 6 is another flow chart for obtaining expression vectors according to an embodiment of the present disclosure.
  • FIGS. 7 to 10 are block diagrams of an image detection device shown in embodiments of the present disclosure.
  • FIG. 11 is a block diagram of an electronic device according to an embodiment of the present disclosure.
  • FIG. 12 is a computer-readable storage medium 1200 according to a disclosed embodiment.
  • image retrieval can retrieve a list of images similar to the image to be recognized and corresponding content such as text, video or web page links by inputting an image to be recognized, which has extremely promising application prospects.
  • the existing related image retrieval algorithm workflow includes: extracting the visual features of the image to be recognized; constructing a feature index or image expression; making similarity judgments based on the image features or the distance of the image expression (such as Euclidean distance, cosine distance, etc.); and Give a list of similar images.
  • the related technology only uses part of the visual features and distance information in the image to be recognized, and does not make full use of the visual features of the image to be recognized, which is not conducive to improving the accuracy and efficiency of retrieval of the image to be recognized.
  • the embodiments of the present disclosure provide an image retrieval method.
  • the public idea is to obtain the local visual features of the image to be recognized, and then determine the residual vector of the local visual feature, and use the residual vector to form the The expression vector of the image, and then the candidate image list can be determined based on the expression vector.
  • the residual vector can be used to make full use of the features of the image to be recognized, which is beneficial to improve the accuracy of image retrieval.
  • FIG. 1 is a block diagram of an image retrieval method shown in an embodiment of the present disclosure.
  • an image retrieval method can be applied to electronic devices, such as smart phones, tablet computers, personal computers, etc., including steps 101 to 103 ,among them:
  • step 101 at least one local visual feature of the image to be recognized is acquired.
  • the electronic device can acquire the image to be recognized input by the user, and then the electronic device can call the visual feature acquisition algorithm to extract the local visual features of the image to be recognized.
  • the visual feature acquisition algorithm may include Scale-invariant Feature Transform (SIFT) algorithm, SURF (Speed Up Robust Features, SURF) algorithm, ORB (Oriented FAST and Rotated BRIE) algorithm, etc.
  • SIFT Scale-invariant Feature Transform
  • SURF Speed Up Robust Features
  • ORB Oriented FAST and Rotated BRIE
  • obtaining local visual features includes:
  • Direction determination Based on the local gradient direction of the image to be recognized, one or more directions are assigned to each key point position. In the subsequent process, the image to be recognized is transformed based on the direction, scale and position of the key points.
  • step 102 an expression vector of the image to be recognized is determined based on the at least one local visual feature.
  • the electronic device may determine the residual vector of the local visual feature based on the local visual feature, and then the expression vector of the image to be recognized is formed by the residual vector.
  • the electronic device can call a pre-trained and stored image classification quantizer.
  • the image classification quantizer is a quantitative model constructed by clustering analysis of all visual features using, for example, the k-means algorithm. After the image classification quantizer obtains the local visual features, it can determine the cluster and sub-categories to which each local visual feature belongs.
  • the image classification quantizer includes K1 clusters and the corresponding K1 cluster centers, as well as K2 sub-categories divided from each cluster in the K1 clusters and the distance boundary threshold of each sub-category (corresponding step 201). Among them, K1 and K2 are positive integers.
  • the local visual features are described by vector representation, and the center of each cluster can also be described by vector representation. Therefore, the difference between the local visual feature and the cluster center corresponding to the cluster cluster can get the residual of the local visual feature vector.
  • the electronic device determining the cluster cluster and subcategory to which each local visual feature belongs may include: the electronic device can use a clustering analysis algorithm to obtain the distance between the local visual feature and each cluster center And the residual vector (corresponding to step 301). Then, the electronic device can determine that the cluster cluster where the cluster center closest to the local visual feature is located is the cluster cluster to which the local visual feature belongs (corresponding to step 302). After that, the electronic device may determine the sub-category to which the local visual feature belongs based on the distance and the distance boundary threshold of each sub-category in the cluster (corresponding to step 303).
  • the training step of the image classification quantizer may include: the electronic device may obtain an image training set. Then, the electronic device can obtain the local visual features of each image in the image training set (corresponding to step 401).
  • the electronic device uses a clustering analysis algorithm to perform cluster analysis on the local visual features to obtain K1 clusters and cluster centers of each cluster (corresponding to step 402).
  • the clustering analysis algorithm may include the K-means algorithm, of course, other clustering algorithms may also be selected, which is not limited here.
  • the electronic device obtains the distance between the local visual feature in the cluster and the cluster center of the cluster and the maximum distance value; based on the maximum distance value, the cluster is divided into K2 sub Class, get the distance threshold of each subclass.
  • the number of distance thresholds obtained here is K1*K2, and the sub-categories in the same cluster do not overlap each other (corresponding to step 403).
  • each distance is equally divided into K2 categories according to the maximum distance.
  • K2 is formed by taking the cluster center of the cluster as the center and the segmentation position of the maximum distance as the radius. Concentric circles, the local visual features located in each circle area are regarded as a sub-category, and the segmentation position is the distance threshold. That is, in this embodiment, the maximum distance value can be divided into K2 segments, and the local visual features in the area formed by each segment are regarded as a sub-category. The boundary between the area corresponding to each segment and the area corresponding to other segments can be used as the distance boundary threshold. Refer to Fig.
  • the cluster K11 can be further divided into K2 sub-categories.
  • K2 is equal to 3
  • the 3 sub-categories can include sub-categories K21, K22, and K23.
  • the distance thresholds of each sub-category are respectively L1, L2 and L3.
  • the electronic device can obtain and accumulate residual vectors of all local visual features belonging to the same subcategory, thereby obtaining the accumulated sum vector of K1*K2 residual vectors.
  • the accumulated sum vector is the waiting Identify the expression vector of the image (corresponding to step 202).
  • each cluster can get K2 sub-categories, and the more detailed local visual features of the image to be recognized can be obtained, that is, the obtained expression vector can reflect more accurately The more detailed local visual features of the image to be recognized help to improve the retrieval accuracy.
  • the electronic device may perform normalization processing on the accumulation sum vector, respectively, to obtain a normalized accumulation sum vector (corresponding to step 601). Then, the electronic device connects the K1*K2 normalized accumulation and vectors end to end to obtain a long vector (corresponding to step 602). After that, the electronic device uses the dimensionality reduction algorithm to perform dimensionality reduction processing on the long vector to obtain the dimensionality-reduced long vector, and the dimensionality-reduced long vector is the expression vector of the image to be recognized (corresponding to step 602). In this way, by processing the cumulative sum vector in this embodiment, the dimension of the vector can be reduced, which is beneficial to reduce the amount of data processing and improve the retrieval efficiency.
  • quantifying each local visual feature of the electronic device may include: using the k-means quantity, that is, calculating the closest one of each local visual feature to K1 cluster centers, indicating that the local visual feature belongs to the cluster cluster.
  • the electronic device also calculates the residual vector and distance value between the local visual feature and the cluster center. Then, according to the distance value, it can be determined which of the K2 areas the local visual feature is located, so as to determine the subclass to which the local visual feature belongs.
  • the above residual cumulative sum vector is the cumulative sum vector.
  • the dimensionality reduction algorithm may include a principal component analysis algorithm (Principal Components Analysis, PCA). Of course, technical personnel may also choose other dimensionality reduction algorithms, which are not limited here.
  • the training step of the image classification quantizer by the electronic device has been completed.
  • the electronic device obtains the K1 cluster centers, K2 distance thresholds obtained in the training process, and the long vector after PCA dimensionality reduction.
  • the electronic device obtains the local visual features of each test image and obtains its expression vector, and uses the PCA algorithm to reduce the dimensionality of the expression vector.
  • the electronic device calculates the similarity between the test image and the training image, using calculation methods such as cosine distance, Euclidean distance, etc., and selects the training image with the highest similarity to the test image to complete the test process.
  • step 103 a list of candidate images whose similarity with the expression vector exceeds the similarity threshold is retrieved from the candidate image library.
  • the electronic device respectively calculates the similarity between the expression vector of the image to be recognized and the expression vector of each candidate image in the candidate image library, and compares the similarity with the similarity threshold. If the similarity exceeds the similarity threshold, then Add candidate images to the candidate image list until the candidate image library is filtered out or the number of candidate images set in advance is reached.
  • the local visual features of the image to be recognized can be acquired, and then the expression vector of the image to be recognized composed of the residual vector of the local visual feature can be determined, and then the similarity with the expression vector can be retrieved A list of candidate images that exceed the similarity threshold.
  • the residual vector of the local visual feature is used in this embodiment, which is beneficial to further express the visual feature of the image to be recognized, thereby improving the accuracy of searching for candidate images.
  • FIG. 7 is a block diagram of an image detection device shown in an embodiment of the present disclosure.
  • an image retrieval device 700 includes:
  • the visual feature acquisition module 701 is configured to acquire at least one local visual feature of the image to be recognized
  • the expression vector determining module 702 is configured to determine an expression vector of the image to be recognized based on the at least one local visual feature, the expression vector being composed of a residual vector of the at least one local visual feature;
  • the image retrieval module 703 is configured to retrieve a list of candidate images whose similarity with the expression vector exceeds the similarity threshold from the candidate image library.
  • the expression vector determining module 702 includes:
  • the cluster determining unit 801 is configured to call an image classification quantizer, and the image classification quantizer determines the cluster cluster and subclass to which each of the at least one local visual feature belongs; the image classification quantizer includes K1 clusters and corresponding K1 cluster centers, as well as K2 sub-categories divided from each cluster in the K1 clusters and the distance dividing threshold of each sub-category;
  • the sum vector obtaining unit 802 is configured to obtain and accumulate residual vectors of all local visual features belonging to the same subcategory, thereby obtaining the cumulative sum vector of K1*K2 residual vectors, and the cumulative sum vector is the waiting Identify the expression vector of the image, where K1 and K2 are positive integers.
  • the cluster determination unit 801 includes:
  • the distance obtaining subunit 901 is configured to obtain the distance and residual vector between the local visual feature and each cluster center by using a cluster analysis algorithm
  • a cluster cluster obtaining subunit 902 configured to determine that the cluster cluster with the closest cluster center to the local visual feature is the cluster cluster to which the local visual feature belongs;
  • the sub-category obtaining sub-unit 903 is configured to determine the sub-category to which the local visual feature belongs based on the distance and the distance boundary threshold of each sub-category in the cluster.
  • the sum vector acquiring unit 802 further includes:
  • the sum vector obtaining subunit 1001 is configured to perform normalization processing on the accumulation sum vector respectively to obtain a normalized accumulation sum vector;
  • the long vector connection subunit 1002 is used to connect the normalized accumulation and vector end to end to obtain a long vector
  • the expression vector obtaining subunit 1003 is configured to perform dimensionality reduction processing on the long vector by using a dimensionality reduction algorithm to obtain a dimensionality-reduced long vector, and the dimensionality-reduced long vector is the expression vector of the image to be recognized.
  • the image classification quantizer is trained through the following steps, including:
  • the distance between the local visual feature in the cluster and the cluster center of the cluster and the maximum distance value are obtained; based on the maximum distance value, the cluster is divided into K2 sub-categories, get the distance threshold of each sub-category.
  • the number of the obtained distance thresholds is K1*K2, and the sub-categories in the same cluster do not overlap each other.
  • FIG. 11 is a block diagram of an electronic device shown in an embodiment of the present disclosure.
  • an electronic device 1100 includes a processor 1101 and a memory 1102 for storing executable instructions; the processor 1101 uses a communication bus 1103 It is connected to the memory 1102 and is used to read executable instructions from the memory 1102 to implement the steps of the image retrieval method shown in FIGS. 1 to 6.
  • an embodiment of the present disclosure also provides a computer-readable storage medium 1200 on which computer-executable instructions are stored.
  • the image retrieval method shown in FIGS. 1 to 6 is implemented. A step of.
  • Computer-readable media include permanent/non-permanent, volatile/non-volatile, removable/non-removable media, and information storage can be achieved by any method or technology.
  • the information can be computer-readable instructions, data structures, program modules, or other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disc (DVD) or other optical storage, Magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media can be used to store information that can be accessed by computing devices. According to the definition in this article, computer-readable media does not include transitory media, such as modulated data signals and carrier waves.
  • PRAM phase change memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • RAM random access memory
  • ROM read-only memory
  • EEPROM electrically erasable programmable read-only memory
  • flash memory or other memory technology
  • CD-ROM compact disc
  • DVD digital versatile disc
  • Magnetic cassettes magnetic tape magnetic disk storage or other magnetic storage devices or any other non

Landscapes

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

Abstract

本公开涉及一种图像检索方法及装置。图像检索方法,包括:获取待识别图像的至少一个局部视觉特征;基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,所述表达向量由所述至少一个局部视觉特征的残差向量构成;从候选图像库中检索出与所述表达向量相似度超过相似度阈值的候选图像列表。

Description

图像检索方法及装置和计算机可读存储介质
相关申请的交叉引用
本申请要求于2019年4月4日递交的中国专利申请CN201910273228.0的优先权,其全部公开内容通过引用合并于此。
技术领域
本公开涉及图像处理技术领域,尤其涉及一种图像检索方法及装置、电子设备、和计算机可读存储介质。
背景技术
在目前,图像检索通过输入待识别图像,检索出与该待识别图像相似的图像列表和相应的文字、视频或网页链接等内容,具有极为可观的应用前景。
现有相关的图像检索算法工作流程包括:提取待识别图像的视觉特征;构建特征索引或图像表达;依据图像特征或图像表达的距离(如欧式距离、余弦距离等)远近进行相似性判断;并给出相似性图像列表。
发明内容
根据本公开实施例的第一方面,提供了一种图像检索方法,包括:
获取待识别图像的至少一个局部视觉特征;
基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,所述表达向量由所述至少一个局部视觉特征的残差向量构成;以及
从候选图像库中检索出与所述表达向量相似度超过相似度阈值的候选图像列表。
在实施例中,基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,包括:
由图像分类量化器确定所述至少一个局部视觉特征中的每一个所属的聚类簇和子类,所述图像分类量化器包括K1个聚类簇和对应的K1个聚类中心,以及所述 K1个聚类簇中的每一个聚类簇切分出的K2个子类和每个子类的距离分界阈值;以及
获取属于同一个子类的所有局部视觉特征的残差向量并累加,从而得到K1*K2个残差向量的累加和向量,所述累加和向量即为所述待识别图像的表达向量,其中K1和K2为正整数。
在实施例中,确定所述至少一个局部视觉特征中的每一个所属的聚类簇和子类包括:
利用聚类分析算法获取该局部视觉特征与各聚类中心的距离和残差向量;
确定与所述局部视觉特征距离最近的聚类中心所在的聚类簇为所述局部视觉特征所属的聚类簇;以及
基于所述距离和所述聚类簇内各子类的距离分界阈值确定所述局部视觉特征所属的子类。
在实施例中,所述图像检索方法,还包括:
对所述累加和向量分别进行归一化处理,得到归一化后的累加和向量;
将归一化后的累加和向量首尾连接得到一个长向量;以及
利用降维算法对所述长向量进行降维处理,得到降维后的长向量,所述降维后的长向量即为所述待识别图像的表达向量。
在实施例中,所述图像分类量化器通过以下步骤训练,包括:
获取图像训练集中每一个图像的局部视觉特征;
利用聚类分析算法对所述局部视觉特征进行聚类分析,得到所述K1个聚类簇和各聚类簇的聚类中心;以及
针对K1个聚类簇中的每一个,获取该聚类簇内局部视觉特征与该聚类簇的聚类中心的距离以及最大距离值;基于所述最大距离值将该聚类簇切分为K2个子类,得到每个子类的距离分界阈值,
其中,得到的距离分界阈值的数量为K1*K2个,同一个聚类簇中的各子类互不重合。
在实施例中,基于所述最大距离值将该聚类簇切分为K2个子类,得到每个子类 的距离分界阈值,包括:
将所述最大距离值划分为K2个分段,每个分段对应的局部视觉特征构成一个子类,得到K2个子类;并且,每个分段与所述聚类簇的聚类中心的最大距离为所述每个分段对应子类的距离分界阈值。
在实施例中,所述聚类分析算法包括K-means算法。
根据本公开实施例的第二方面,提供了一种图像检索装置,包括:
处理器;以及
存储器,与所述处理器耦连,并存储可执行指令,所述可执行指令在由所述处理器执行时使得所述处理器被配置为:
获取待识别图像的至少一个局部视觉特征;
基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,所述表达向量由所述至少一个局部视觉特征的残差向量构成;以及
从候选图像库中检索出与所述表达向量相似度超过相似度阈值的候选图像列表。
在实施例中,所述处理器还被配置为:
由图像分类量化器确定所述至少一个局部视觉特征中的每一个所属的聚类簇和子类,所述图像分类量化器包括K1个聚类簇和对应的K1个聚类中心,以及所述K1个聚类簇中的每一个聚类簇切分出的K2个子类和每个子类的距离分界阈值;以及
获取属于同一个子类的所有局部视觉特征的残差向量,并累加,从而得到K1*K2个残差向量的累加和向量,所述累加和向量即为所述待识别图像的表达向量,其中K1和K2为正整数。
在实施例中,所述处理器还被配置为:
利用聚类分析算法获取该局部视觉特征与各聚类中心的距离和残差向量;
确定与所述局部视觉特征距离最近的聚类中心所在的聚类簇为所述局部视觉特征所属的聚类簇;以及
基于所述距离和所述聚类簇内各子类的距离分界阈值确定所述局部视觉特征所属的子类。
在实施例中,所述处理器还被配置为:
对所述累加和向量分别进行归一化处理,得到归一化后的累加和向量;
将归一化后的累加和向量首尾连接得到一个长向量;以及
利用降维算法对所述长向量进行降维处理,得到降维后的长向量,所述降维后的长向量即为所述待识别图像的表达向量。
在实施例中,所述图像分类量化器通过以下步骤训练,包括:
获取图像训练集中每个图像的局部视觉特征;
利用聚类分析算法对所述局部视觉特征进行聚类分析,得到所述K1个聚类簇和各聚类簇的聚类中心;以及
针对K1个聚类簇中的每一个,获取该聚类簇内局部视觉特征与所述聚类簇的聚类中心的距离以及最大距离值;基于所述最大距离值将该聚类簇切分为K2个子类,得到各子类的距离分界阈值,
其中,得到的距离分界阈值的数量为K1*K2个,同一个聚类簇中的各子类互不重合。
根据本公开实施例的第三方面,提供了一种计算机可读存储介质,其上存储有计算机可执行指令,其特征在于,该指令被处理器执行时实现权利要求1~7任一项所述方法的步骤。
应当理解的是,以上的一般描述和后文的细节描述仅是示例性和解释性的,并不能限制本公开。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分,示出了符合本公开的实施例,并与说明书一起用于解释本公开的原理。
图1是本公开实施例示出的一种图像检索方法的框图;
图2是本公开实施例示出的获取表达向量的流程图;
图3是本公开实施例示出的获取距离分界阈值的流程图;
图4是本公开实施例示出的获取聚类簇子类的流程图;
图5是本公开实施例示出的划分的K1个聚类簇和K2个子类的效果示意图;
图6是本公开实施例示出的另一种获取表达向量的流程图;
图7~图10是本公开实施例示出的一种图像检测装置的框图;
图11是本公开实施例示出的一种电子设备的框图;以及
图12是公开实施例示出的一种计算机可读存储介质1200。
具体实施方式
这里将详细地对示例性实施例进行说明,其示例表示在附图中。下面的描述涉及附图时,除非另有表示,不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本公开相一致的所有实施方式。相反,它们仅是与如所附权利要求书中所详述的、本公开的一些方面相一致的装置和方法的例子。
目前,图像检索通过输入待识别图像,检索出与该待识别图像相似的图像列表和相应的文字、视频或网页链接等内容,具有极为可观的应用前景。
现有相关的图像检索算法工作流程包括:提取待识别图像的视觉特征;构建特征索引或图像表达;依据图像特征或图像表达的距离(如欧式距离、余弦距离等)远近进行相似性判断;并给出相似性图像列表。
然而,相关技术仅利用了待识别图像中的部分视觉特征以及距离信息,并未充分利用待识别图像的视觉特征,不利于提升待识别图像检索的准确率和效率。
为解决上述问题,本公开实施例提供了一种图像检索方法,其公开构思在于,获取待识别图像的局部视觉特征,然后再确定出局部视觉特征的残差向量,利用残差向量形成待识别图像的表达向量,之后基于表达向量可以确定候选图像列表。本实施例中通过利用残差向量可以充分利用待识别图像的特征,有利于提升图像检索的准确率。
图1是本公开实施例示出的一种图像检索方法的框图,参见图1,一种图像检索方法,可以应用于电子设备,例如智能手机、平板电脑、个人计算机等,包括步骤101~步骤103,其中:
在步骤101中,获取待识别图像的至少一个局部视觉特征。
在本实施例中,电子设备可以获取用户输入的待识别图像,然后电子设备可以调用视觉特征获取算法提取待识别图像的局部视觉特征。
其中,视觉特征获取算法可以包括尺度不变特征变换(Scale-invariant feature transform,SIFT)算法,SURF(Speed Up Robust Features,SURF)算法,ORB(Oriented FAST and Rotated BRIE)算法等。
以SIFT算法为例,获取局部视觉特征包括:
1.尺度空间极值检测:搜索待识别图像上所有尺度上的图像位置,利用高斯微分函数来识别潜在的对于尺度和旋转不变的兴趣点。
2.关键点定位:在每个候选的兴趣点上,通过一个拟合模型来确定位置和尺度,将位置和尺度稳定程度较高的兴趣点作为关键点。
3.方向确定:基于待识别图像局部的梯度方向,分配给每个关键点位置一个或多个方向。后续过程中,基于关键点的方向、尺度和位置对待识别图像进行变换。
4.关键点描述:在每个关键点周围的邻域内,在选定的尺度上测量待识别图像的图像局部的梯度,其中梯度可以变换成表示向量,表示向量允许比较大的局部形状的变形和光照变化。换言之,上述表示向量即为待识别图像的局部视觉特征。
可理解的是,本实施例中仅描述了SIFT算法的获取视觉特征的方案,当然,技术人员可以根据具体场景选择合适的视觉特征获取算法,在能够获取到局部视觉特征的情况下,相应算法及算法方案落入本申请的保护范围。
在步骤102中,基于所述至少一个局部视觉特征确定所述待识别图像的表达向量。
在本实施例中,电子设备可以基于局部视觉特征确定出局部视觉特征的残差向量,然后由残差向量构成的待识别图像的表达向量。
参见图2,电子设备可以调用预先训练并存储的图像分类量化器。图像分类量化器是使用例如k-means算法对所有视觉特征进行聚类分析构建而成的量化模型。该图像分类量化器在获取到局部视觉特征后,可以确定出每个局部视觉特征所属的聚类簇和子类。图像分类量化器包括K1个聚类簇和对应的K1个聚类中心,以及 K1个聚类簇中由每个聚类簇切分出的K2个子类和各子类的距离分界阈值(对应步骤201)。其中,K1和K2为正整数。
其中,局部视觉特征采用向量表示来描述,各聚类簇中心也可以采用向量表示来描述,因此局部视觉特征与其所属聚类簇对应的聚类簇中心作差即可以得到局部视觉特征的残差向量。
在本实施例中,参见图3,电子设备确定每个局部视觉特征所属的聚类簇和子类,可以包括:电子设备可以利用聚类分析算法获取该局部视觉特征与各聚类中心的距离和残差向量(对应步骤301)。然后,电子设备可以确定与该局部视觉特征距离最近的聚类中心所在的聚类簇为该局部视觉特征所属的聚类簇(对应步骤302)。之后,电子设备可以基于距离和聚类簇内各子类的距离分界阈值确定所述局部视觉特征所属的子类(对应步骤303)。
可理解的是,上述图像分类量化器需要预先训练好,参见图4,图像分类量化器的训练步骤可以包括:电子设备可以获取图像训练集。然后电子设备可以获取图像训练集中的每一张图像的局部视觉特征(对应步骤401)。
然后,电子设备利用聚类分析算法对局部视觉特征进行聚类分析,得到K1个聚类簇和各聚类簇的聚类中心(对应步骤402)。其中,聚类分析算法可以包括K-means算法,当然还可以选择其他聚类算法,在此不作限定。
之后,针对每个聚类簇,电子设备获取该聚类簇内局部视觉特征与该聚类簇的聚类中心的距离以及最大距离值;基于最大距离值将该聚类簇切分为K2个子类,得到各子类的距离分界阈值。这里得到的距离分界阈值的数量为K1*K2个,同一个聚类簇中的各子类互不重合(对应步骤403)。
在一实施例中针对各聚类簇,根据最大距离将各个距离平均切分为K2个类,例如,以聚类簇的聚类中心为圆心,以最大距离的切分位置为半径形成K2个同心圆,位于各圆环区域内的局部视觉特征作为一个子类,而切分位置即是距离分界阈值。即本实施例中可以将最大距离值划分为K2个分段,每个分段形成的区域内的局部视觉特征作为一个子类。每个分段对应的区域与其他分段对应的区域之间的边界可以作为距离分界阈值。参见图5,图5中示出了K1等于4即待识别图像具有4个聚类 簇以及4个聚类中心,其中4个聚类簇分别为K11、K12、K13和K14。聚类簇K11可以继续切分为K2个子类,当K2等于3时,即切分为3个子类,其中3个子类可以包括子类K21、K22和K23。各子类的距离分界阈值分别为L1、L2和L3。
继续参见图2,电子设备可以获取属于同一个子类的所有局部视觉特征的残差向量并累加,从而得到K1*K2个残差向量的累加和向量,所述累加和向量即为所述待识别图像的表达向量(对应步骤202)。这样,本实施例中通过对聚类簇继续切分,使得每个聚类簇得到K2个子类,可以得到待识别图像的更细节的局部视觉特征,即所得到的表达向量能够更准确的反映待识别图像的更细节的局部视觉特征,有利于提高检索准确率。
在一些实施例中,参见图6,电子设备可以对累积和向量分别进行归一化处理,从而得到归一化后的累加和向量(对应步骤601)。然后,电子设备将K1*K2个归一化后的累加和向量首尾连接得到一个长向量(对应步骤602)。之后,电子设备利用降维算法对长向量进行降维处理,得到降维后的长向量,降维后的长向量即为待识别图像的表达向量(对应步骤602)。这样,本实施例中通过对累加和向量进行处理,可以降低向量的维度,有利于降低数据处理量,提升检索效率。
在本实施例中,电子设备各局部视觉特征进行量化可以包括:利用k-means量,即计算每个局部视觉特征与K1个聚类中心中最近的一个,表示该局部视觉特征属于该聚类簇。并且,电子设备还计算局部视觉特征与聚类中心的残差向量和距离值。然后,根据距离值可以确定该局部视觉特征位于K2个区域内的哪一个区域,从而确定局部视觉特征所属于的子类。
在确定了所有局部视觉特征所属于的子类后,若两个局部视觉特征同属于同一个子类,则累加它们的残差向量,重复上述步骤多次,可以得到K1*K2个残差累加和向量。在一些场景中,上述残差累加和向量即为累加和向量。
在一些实施例中,电子设备还对每个残差累加和向量分别进行归一化处理,并将K1*K2个残差累加和向量首尾连接成一个长向量,该长向量的维度为D=K1*K2*d(设置的特征维度)。然后,对长向量进行归一化处理可以得到归一化后的长向量。之后,利用降维算法对归一化后的长向量进行降维处理,得到降维后的长向量,该 降维后的长向量即为最终的累加和向量。其中,降维算法可以包括主成分分析算法(Principal Components Analysis,PCA),当然技术人员还可以选择其他降维算法,在此不作限定。
这样,电子设备对图像分类量化器的训练步骤已经完成。电子设备获取训练过程中得到的K1聚类中心、K2个距离分界阈值,以及PCA降维后的长向量。然后,电子设备获取每个测试图像的局部视觉特征,并获得其表达向量,利用PCA算法对表达向量进行降维。最后,电子设备计算测试图像和训练图像之间的相似性,计算方法如余弦距离,欧式距离等,并筛选出与测试图像相似性最高的训练图像,完成测试过程。
在步骤103中,从候选图像库中检索出与所述表达向量相似度超过相似度阈值的候选图像列表。
在本实施例中,电子设备分别计算待识别图像的表达向量和候选图像库中各候选图像的表达向量的相似度,将相似度与相似度阈值近比较,若相似度超过相似度阈值,则将候选图像加入候选图像列表,直至筛选完候选图像库或者达到预先设置的候选图像数量为止。
至此,本实施例中可以获取待识别图像的局部视觉特征,然后确定出由所述局部视觉特征的残差向量构成的所述待识别图像的表达向量,之后检索出与所述表达向量相似度超过相似度阈值的候选图像列表。这样,本实施例中利用了局部视觉特征的残差向量,有利于进一步表达出待识别图像的视觉特征,从而可以提高提高检索候选图像的准确度。
在本公开实施例提供的一种图像检索方法的基础上,本公开实施例还提供了一种图像检索装置,图7是本公开实施例示出的一种图像检测装置的框图。参见图7,一种图像检索装置700,包括:
视觉特征获取模块701,用于获取待识别图像的至少一个局部视觉特征;
表达向量确定模块702,用于基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,所述表达向量由所述至少一个局部视觉特征的残差向量构成;
图像检索模块703,用于从候选图像库中检索出与所述表达向量相似度超过相 似度阈值的候选图像列表。
在图7所示的一种图像检索装置的基础上,参见图8,所述表达向量确定模块702包括:
聚类簇确定单元801,用于调用图像分类量化器,由所述图像分类量化器确定所述至少一个局部视觉特征中的每一个所属的聚类簇和子类;所述图像分类量化器包括K1个聚类簇和对应的K1个聚类中心,以及所述K1个聚类簇中由每个聚类簇切分出的K2个子类和各子类的距离分界阈值;
和向量获取单元802,用于获取属于同一个子类的所有局部视觉特征的残差向量并累加,从而得到K1*K2个残差向量的累加和向量,所述累加和向量即为所述待识别图像的表达向量,其中K1和K2为正整数。
在图8所示的一种图像检索装置的基础上,参见图9,所述聚类簇确定单元801包括:
距离获取子单元901,用于利用聚类分析算法获取所述局部视觉特征与各聚类中心的距离和残差向量;
聚类簇获取子单元902,用于确定与所述局部视觉特征距离最近的聚类中心所在的聚类簇为所述局部视觉特征所属的聚类簇;
子类获取子单元903,用于基于所述距离和所述聚类簇内各子类的距离分界阈值确定所述局部视觉特征所属的子类。
在图8所示的一种图像检索装置的基础上,参见图10,所述和向量获取单元802还包括:
和向量获取子单元1001,用于对所述累加和向量分别进行归一化处理,得到归一化后的累加和向量;
长向量连接子单元1002,用于将归一化后的累加和向量首尾连接得到一个长向量;
表达向量获取子单元1003,用于利用降维算法对所述长向量进行降维处理,得到降维后的长向量,所述降维后的长向量即为所述待识别图像的表达向量。
在一实施例中,所述图像分类量化器通过以下步骤训练,包括:
获取图像训练集中每个图像的局部视觉特征;
利用聚类分析算法对所述局部视觉特征进行聚类分析,得到所述K1个聚类簇和各聚类簇的聚类中心;
针对K1个聚类簇中的每一个,获取该聚类簇内局部视觉特征与该聚类簇的聚类中心的距离以及最大距离值;基于所述最大距离值将该聚类簇切分为K2个子类,得到各子类的距离分界阈值。在实施例中,得到的距离分界阈值的数量为K1*K2个,同一个聚类簇中的各子类互不重合。
图11是本公开实施例示出的一种电子设备的框图,参见图11,一种电子设备1100,包括处理器1101和用于存储可执行指令的存储器1102;所述处理器1101通过通信总线1103与所述存储器1102连接,用于从所述存储器1102中读取可执行指令,以实现图1~图6所示的图像检索方法的步骤。
如图12所示,本公开实施例还提供了一种计算机可读存储介质1200,其上存储有计算机可执行指令,该指令被处理器执行时实现图1~图6所示的图像检索方法的步骤。
计算机可读介质包括永久性/非永久性、易失性/非易失性、可移动/非可移动媒体,可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括,但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带,磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质,可用于存储可以被计算设备访问的信息。按照本文中的界定,计算机可读介质不包括暂存电脑可读媒体(transitory media),如调制的数据信号和载波。
本领域技术人员在考虑说明书及实践这里公开的公开后,将容易想到本公开的其它实施方案。本公开旨在涵盖本公开的任何变型、用途或者适应性变化,这些变型、用途或者适应性变化遵循本公开的一般性原理并包括本公开未公开的本技术领 域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的,本公开的真正范围和精神由下面的权利要求指出。
应当理解的是,本公开并不局限于上面已经描述并在附图中示出的精确结构,并且可以在不脱离其范围进行各种修改和改变。本公开的范围仅由所附的权利要求来限制。

Claims (13)

  1. 一种图像检索方法,包括:
    获取待识别图像的至少一个局部视觉特征;
    基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,所述表达向量由所述至少一个局部视觉特征的残差向量构成;以及
    从候选图像库中检索出与所述表达向量相似度超过相似度阈值的候选图像列表。
  2. 根据权利要求1所述的图像检索方法,其中,基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,包括:
    由图像分类量化器确定所述至少一个局部视觉特征中的每一个所属的聚类簇和子类,所述图像分类量化器包括K1个聚类簇和对应的K1个聚类中心,以及所述K1个聚类簇中的每一个聚类簇切分出的K2个子类和每个子类的距离分界阈值;以及
    获取属于同一个子类的所有局部视觉特征的残差向量并累加,从而得到K1*K2个残差向量的累加和向量,所述累加和向量即为所述待识别图像的表达向量,其中K1和K2为正整数。
  3. 根据权利要求2所述的图像检索方法,其中,确定所述至少一个局部视觉特征中的每一个所属的聚类簇和子类包括:
    利用聚类分析算法获取该局部视觉特征与各聚类中心的距离和残差向量;
    确定与所述局部视觉特征距离最近的聚类中心所在的聚类簇为所述局部视觉特征所属的聚类簇;以及
    基于所述距离和所述聚类簇内各子类的距离分界阈值确定所述局部视觉特征所属的子类。
  4. 根据权利要求2所述的图像检索方法,还包括:
    对所述累加和向量分别进行归一化处理,得到归一化后的累加和向量;
    将归一化后的累加和向量首尾连接得到一个长向量;以及
    利用降维算法对所述长向量进行降维处理,得到降维后的长向量,所述降维后 的长向量即为所述待识别图像的表达向量。
  5. 根据权利要求2所述的图像检索方法,其中,所述图像分类量化器通过以下步骤训练,包括:
    获取图像训练集中每一个图像的局部视觉特征;
    利用聚类分析算法对所述局部视觉特征进行聚类分析,得到所述K1个聚类簇和各聚类簇的聚类中心;以及
    针对K1个聚类簇中的每一个,获取该聚类簇内局部视觉特征与该聚类簇的聚类中心的距离以及最大距离值;基于所述最大距离值将该聚类簇切分为K2个子类,得到每个子类的距离分界阈值,
    其中,得到的距离分界阈值的数量为K1*K2个,同一个聚类簇中的各子类互不重合。
  6. 根据权利要求5所述的图像检索方法,其中,基于所述最大距离值将该聚类簇切分为K2个子类,得到每个子类的距离分界阈值,包括:
    将所述最大距离值划分为K2个分段,每个分段对应的局部视觉特征构成一个子类,得到K2个子类;并且,每个分段与所述聚类簇的聚类中心的最大距离为所述每个分段对应子类的距离分界阈值。
  7. 根据权利要求5所述的图像检索方法,其中,所述聚类分析算法包括K-means算法。
  8. 一种图像检索装置,包括:
    处理器;以及
    存储器,与所述处理器耦连,并存储可执行指令,所述可执行指令在由所述处理器执行时使得所述处理器被配置为:
    获取待识别图像的至少一个局部视觉特征;
    基于所述至少一个局部视觉特征确定所述待识别图像的表达向量,所述表达向量由所述至少一个局部视觉特征的残差向量构成;以及
    从候选图像库中检索出与所述表达向量相似度超过相似度阈值的候选图像列表。
  9. 根据权利要求8所述的图像检索装置,其中,所述处理器还被配置为:
    由图像分类量化器确定所述至少一个局部视觉特征中的每一个所属的聚类簇和子类,所述图像分类量化器包括K1个聚类簇和对应的K1个聚类中心,以及所述K1个聚类簇中的每一个聚类簇切分出的K2个子类和每个子类的距离分界阈值;以及
    获取属于同一个子类的所有局部视觉特征的残差向量,并累加,从而得到K1*K2个残差向量的累加和向量,所述累加和向量即为所述待识别图像的表达向量,其中K1和K2为正整数。
  10. 根据权利要求9所述的图像检索装置,其中,所述处理器还被配置为:
    利用聚类分析算法获取该局部视觉特征与各聚类中心的距离和残差向量;
    确定与所述局部视觉特征距离最近的聚类中心所在的聚类簇为所述局部视觉特征所属的聚类簇;以及
    基于所述距离和所述聚类簇内各子类的距离分界阈值确定所述局部视觉特征所属的子类。
  11. 根据权利要求9所述的图像检索装置,其中,所述处理器还被配置为:
    对所述累加和向量分别进行归一化处理,得到归一化后的累加和向量;
    将归一化后的累加和向量首尾连接得到一个长向量;以及
    利用降维算法对所述长向量进行降维处理,得到降维后的长向量,所述降维后的长向量即为所述待识别图像的表达向量。
  12. 根据权利要求9所述的图像检索装置,其中,所述图像分类量化器通过以下步骤训练,包括:
    获取图像训练集中每个图像的局部视觉特征;
    利用聚类分析算法对所述局部视觉特征进行聚类分析,得到所述K1个聚类簇和各聚类簇的聚类中心;以及
    针对K1个聚类簇中的每一个,获取该聚类簇内局部视觉特征与所述聚类簇的聚类中心的距离以及最大距离值;基于所述最大距离值将该聚类簇切分为K2个子类,得到各子类的距离分界阈值,
    其中,得到的距离分界阈值的数量为K1*K2个,同一个聚类簇中的各子类互不 重合。
  13. 一种计算机可读存储介质,其上存储有计算机可执行指令,其特征在于,该指令被处理器执行时实现权利要求1~7任一项所述方法的步骤。
PCT/CN2020/075685 2019-04-04 2020-02-18 图像检索方法及装置和计算机可读存储介质 WO2020199773A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910273228.0A CN111783805B (zh) 2019-04-04 2019-04-04 图像检索方法及装置、电子设备、可读存储介质
CN201910273228.0 2019-04-04

Publications (1)

Publication Number Publication Date
WO2020199773A1 true WO2020199773A1 (zh) 2020-10-08

Family

ID=72664894

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/075685 WO2020199773A1 (zh) 2019-04-04 2020-02-18 图像检索方法及装置和计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN111783805B (zh)
WO (1) WO2020199773A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112926406A (zh) * 2021-02-02 2021-06-08 广东嘉铭智能科技有限公司 一种自动化免疫荧光检测方法、装置、设备及系统
CN113506284A (zh) * 2021-07-26 2021-10-15 电子科技大学 一种眼底图像微血管瘤检测装置、方法及存储介质
CN115035556A (zh) * 2021-03-03 2022-09-09 北京迈格威科技有限公司 人脸检索方法、装置、电子设备及存储介质
CN115841488A (zh) * 2023-02-21 2023-03-24 聊城市飓风工业设计有限公司 一种基于计算机视觉的pcb板的检孔方法
WO2023222091A1 (zh) * 2022-05-18 2023-11-23 华为技术有限公司 一种向量检索方法及装置

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113204665B (zh) * 2021-04-28 2023-09-22 北京百度网讯科技有限公司 图像检索方法、装置、电子设备及计算机可读存储介质
CN113536020B (zh) * 2021-07-23 2022-05-24 贝壳找房(北京)科技有限公司 数据查询的方法、存储介质和计算机程序产品
CN113792171B (zh) * 2021-11-15 2022-02-18 西安热工研究院有限公司 基于内存管理的图像检索方法、系统、设备及存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105335757A (zh) * 2015-11-03 2016-02-17 电子科技大学 一种基于局部特征聚合描述符的车型识别方法
CN106326395A (zh) * 2016-08-18 2017-01-11 北京大学 一种局部视觉特征选择方法及装置
CN108563777A (zh) * 2018-04-24 2018-09-21 京东方科技集团股份有限公司 一种获得图像表示的方法和装置

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2003256429A (ja) * 2002-03-01 2003-09-12 Kyodo Printing Co Ltd 模様検索装置及びその方法
CN103092861B (zh) * 2011-11-02 2016-01-06 阿里巴巴集团控股有限公司 一种商品代表图的选取方法和系统
GB2516037A (en) * 2013-07-08 2015-01-14 Univ Surrey Compact and robust signature for large scale visual search, retrieval and classification
CN104216949A (zh) * 2014-08-13 2014-12-17 中国科学院计算技术研究所 一种融合空间信息的图像特征聚合表示方法及系统
CN108805183B (zh) * 2018-05-28 2022-07-26 南京邮电大学 一种融合局部聚合描述符和局部线性编码的图像分类方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105335757A (zh) * 2015-11-03 2016-02-17 电子科技大学 一种基于局部特征聚合描述符的车型识别方法
CN106326395A (zh) * 2016-08-18 2017-01-11 北京大学 一种局部视觉特征选择方法及装置
CN108563777A (zh) * 2018-04-24 2018-09-21 京东方科技集团股份有限公司 一种获得图像表示的方法和装置

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112926406A (zh) * 2021-02-02 2021-06-08 广东嘉铭智能科技有限公司 一种自动化免疫荧光检测方法、装置、设备及系统
CN115035556A (zh) * 2021-03-03 2022-09-09 北京迈格威科技有限公司 人脸检索方法、装置、电子设备及存储介质
CN113506284A (zh) * 2021-07-26 2021-10-15 电子科技大学 一种眼底图像微血管瘤检测装置、方法及存储介质
CN113506284B (zh) * 2021-07-26 2023-05-09 电子科技大学 一种眼底图像微血管瘤检测装置、方法及存储介质
WO2023222091A1 (zh) * 2022-05-18 2023-11-23 华为技术有限公司 一种向量检索方法及装置
CN115841488A (zh) * 2023-02-21 2023-03-24 聊城市飓风工业设计有限公司 一种基于计算机视觉的pcb板的检孔方法

Also Published As

Publication number Publication date
CN111783805B (zh) 2024-08-23
CN111783805A (zh) 2020-10-16

Similar Documents

Publication Publication Date Title
WO2020199773A1 (zh) 图像检索方法及装置和计算机可读存储介质
CN108920580B (zh) 图像匹配方法、装置、存储介质及终端
US8805117B2 (en) Methods for improving image search in large-scale databases
Tabia et al. Covariance-based descriptors for efficient 3D shape matching, retrieval, and classification
US10474872B2 (en) Fingerprint matching using virtual minutiae
CN111914921A (zh) 一种基于多特征融合的相似性图像检索方法及系统
US20190205693A1 (en) Scale-Invariant Feature Point Extraction in Edge Map
CN104199842A (zh) 一种基于局部特征邻域信息的相似图片检索方法
CN110941989A (zh) 图像校验、视频校验方法、装置、设备及存储介质
US20210312215A1 (en) Method for book recognition and book reading device
US20180307940A1 (en) A method and a device for image matching
US9870517B2 (en) Image object retrieval
CN113343920A (zh) 人脸识别照片的分类方法、装置、电子设备和存储介质
US11449704B2 (en) Searching in multilevel clustered vector-based data
CN112445926B (zh) 一种图像检索方法以及装置
WO2019100348A1 (zh) 图像检索方法和装置以及图像库的生成方法和装置
CN110083731B (zh) 图像检索方法、装置、计算机设备及存储介质
CN109829383B (zh) 掌纹识别方法、装置和计算机设备
Zhang et al. Handwritten word image matching based on Heat Kernel Signature
JP6017277B2 (ja) 特徴ベクトルの集合で表されるコンテンツ間の類似度を算出するプログラム、装置及び方法
Vinay et al. A double filtered GIST descriptor for face recognition
CN111008294B (zh) 交通图像处理、图像检索方法及装置
CN113688263B (zh) 用于搜索图像的方法、计算设备和存储介质
CN113657378B (zh) 车辆跟踪方法、车辆跟踪系统和计算设备
CN115482398A (zh) 一种外观侵权的检测方法、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: 20782417

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20782417

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20782417

Country of ref document: EP

Kind code of ref document: A1

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 04.05.2022)

122 Ep: pct application non-entry in european phase

Ref document number: 20782417

Country of ref document: EP

Kind code of ref document: A1