WO2018166273A1 - 高维图像特征匹配方法和装置 - Google Patents

高维图像特征匹配方法和装置 Download PDF

Info

Publication number
WO2018166273A1
WO2018166273A1 PCT/CN2017/119489 CN2017119489W WO2018166273A1 WO 2018166273 A1 WO2018166273 A1 WO 2018166273A1 CN 2017119489 W CN2017119489 W CN 2017119489W WO 2018166273 A1 WO2018166273 A1 WO 2018166273A1
Authority
WO
WIPO (PCT)
Prior art keywords
dimensional image
low
image
database
features
Prior art date
Application number
PCT/CN2017/119489
Other languages
English (en)
French (fr)
Inventor
安山
陈宇
Original Assignee
北京京东尚科信息技术有限公司
北京京东世纪贸易有限公司
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 北京京东尚科信息技术有限公司, 北京京东世纪贸易有限公司 filed Critical 北京京东尚科信息技术有限公司
Priority to US16/494,632 priority Critical patent/US11210555B2/en
Publication of WO2018166273A1 publication Critical patent/WO2018166273A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/55—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/51—Indexing; Data structures therefor; Storage structures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06F18/211—Selection of the most significant subset of features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/231—Hierarchical techniques, i.e. dividing or merging pattern sets so as to obtain a dendrogram
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G06V10/7625—Hierarchical techniques, i.e. dividing or merging patterns to obtain a tree-like representation; Dendograms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G06V10/763—Non-hierarchical techniques, e.g. based on statistics of modelling distributions

Definitions

  • the present disclosure relates to the field of image retrieval, and in particular, to a high-dimensional image feature matching method and apparatus.
  • the feature matching problem is usually defined as Approximate Nearest Neighbor, which is the feature in the database feature that searches for the closest Euclidean distance to the retrieved feature.
  • High-dimensional features refer to image features over 1000 dimensions, and the matching of high-dimensional features is difficult due to the existence of the condition of dimensionality.
  • the description method of the image includes two types of local features and global features.
  • the local features of an image are usually hundreds of features with lower dimensions, and the global features of the image are usually a single feature with a higher dimension.
  • global features can be matched using an exhaustive search, which can ensure an exact match of nearest neighbors at Euclidean distance.
  • Feature matching can also be performed by learning the binary codes, which convert high-dimensional features into binary codes with low dimensions but retaining similarity by using bilinear mapping.
  • the inventors have realized that the efficiency of matching global features is extremely low by using exhaustive search.
  • Feature learning is performed by learning binary codes. Due to bilinear mapping of high-dimensional features, mapping errors may occur, resulting in matching errors.
  • One technical problem to be solved by the present disclosure is to provide a high-dimensional image feature matching method and apparatus capable of improving the accuracy of feature matching.
  • a high-dimensional image feature matching method including: extracting a high-dimensional image feature of an image to be retrieved; and dividing a high-dimensional image feature of the image to be retrieved into a plurality of low-dimensional image features; Each low-dimensional image feature of the retrieved image is compared with each cluster center of the low-dimensional image feature of the image in the database; and the similarity of the low-dimensional image feature of the image to be retrieved and the image in the database is determined according to the comparison result, thereby searching in the database A feature that matches the high dimensional image features of the image to be retrieved.
  • the cluster center of each layer of the low-dimensional image feature of the image in the database is determined by: extracting high-dimensional image features of the image in the database; and dividing the high-dimensional image feature of the image in the database into multiple low-dimensional images Feature; hierarchical clustering calculation of low-dimensional image features of images in the database to determine clustering centers of layers of low-dimensional image features.
  • the cluster center of each layer of each low dimensional image feature of the image in the database is associated with the inverted file.
  • performing hierarchical clustering calculation on low-dimensional image features of the image in the database to determine each layer of the clustering center of the low-dimensional image feature includes: performing unsupervised training on the set of low-dimensional image features of the image in the database, wherein The clustering bifurcation factor is defined as k; a low-dimensional image feature data point is randomly selected in the initial training set as the initial clustering center of the kmeans clustering algorithm; k initial clustering centers are determined according to the probability ratio of the point to the nearest center
  • the kmeans clustering algorithm is used to divide the set of low-dimensional image features of the image in the database into k sets, which is the first layer of the kmeans tree.
  • the kmeans clustering algorithm is used to iterate successively for each set of the current layer in the kmeans tree. Divided until the leaf layer of the kmeans tree is reached; the cluster center of each set of each layer of the low-dimensional image features of the image in the database is calculated from the cluster mean.
  • comparing each low-dimensional image feature of the image to be retrieved with each layer cluster center of the low-dimensional image feature of the image in the database includes: selecting each low-dimensional image feature of the image to be retrieved from the image in the database The first layer of the kmeans tree of the low-dimensional image feature starts, and compared with the k cluster centers of each layer, the cluster center closest to each low-dimensional image feature of the image to be retrieved is determined; the cluster with the closest recursion and distance The center's sub-cluster centers are compared until the leaf layer of the kmeans tree is reached.
  • the method further comprises: using a TF-IDF word frequency-reverse file frequency algorithm to score the low-dimensional image feature similarity between the image to be retrieved and the image in the database; and using the fast selection algorithm combined with the insertion ordering method to obtain the N highest scores
  • the image, the high-dimensional image features of the N images with the highest scores are taken as features matching the high-dimensional image features of the image to be retrieved.
  • a high-dimensional image feature matching apparatus including: a first high-dimensional image feature extraction unit, configured to extract a high-dimensional image feature of an image to be retrieved; and a first low-dimensional image feature segmentation a unit, configured to divide a high-dimensional image feature of the image to be retrieved into a plurality of low-dimensional image features; and a feature comparing unit, configured to display each low-dimensional image feature of the image to be retrieved and each layer of the low-dimensional image feature of the image in the database
  • the clustering center performs comparison; the feature matching determining unit is configured to determine, according to the comparison result, a similarity degree of the low-dimensional image feature of the image to be retrieved and the image in the database, thereby searching for a feature matching the high-dimensional image feature of the image to be retrieved in the database.
  • the cluster center of each layer of the low-dimensional image feature of the image in the database is determined by the following unit: a second high-dimensional image feature extraction unit for extracting high-dimensional image features of the image in the database; and a second low-dimensional image
  • the feature dividing unit is configured to divide the high-dimensional image feature of the image in the database into a plurality of low-dimensional image features
  • the cluster center determining unit is configured to perform hierarchical clustering calculation on the low-dimensional image features of the image in the database to determine the low dimension The clustering center of each layer of image features.
  • the cluster center of each layer of each low dimensional image feature of the image in the database is associated with the inverted file.
  • the cluster center determining unit is configured to perform unsupervised training on the set of low-dimensional image features of the image in the database, wherein the branching factor defining the cluster is k; randomly selecting a low-dimensional image in the initial training set
  • the feature data points are used as the initial clustering center of the kmeans clustering algorithm; k initial cluster centers are determined according to the probability ratio of the points to the nearest center;
  • the kmeans clustering algorithm is used to divide the set of low-dimensional image features of the images in the database into k
  • the set is used as the first layer of the kmeans tree; the kmeans clustering algorithm is used to iteratively divide each set of the current layer in the kmeans tree until the leaf layer of the kmeans tree is reached; the low dimension of the image in the database is calculated according to the cluster mean The clustering center of each set of each layer of the image's kmeans tree.
  • the feature matching determining unit is configured to compare each low-dimensional image feature of the image to be retrieved from the first layer of the kmeans tree of the low-dimensional image feature of the image in the database, and compare with each of the k cluster centers And determining a cluster center closest to each low-dimensional image feature of the image to be retrieved; recursively comparing with the sub-cluster center of the nearest cluster center until reaching the leaf layer of the kmeans tree.
  • the feature matching determining unit is further configured to use the TF-IDF word frequency-reverse file frequency algorithm to score the low-dimensional image feature similarity between the image to be retrieved and the image in the database; and use the fast selection algorithm combined with the insertion ordering method to obtain the highest score.
  • N images, the high-dimensional image features of the N images with the highest scores are taken as features matching the high-dimensional image features of the image to be retrieved.
  • a high dimensional image feature matching apparatus comprising: a memory; and a processor coupled to the memory, the processor being configured to perform the method described above based on instructions stored in the memory.
  • a computer readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the steps of the method described above.
  • the embodiment of the present disclosure divides the high-dimensional image features of the image to be retrieved into a plurality of low-dimensional image features, and then clusters each low-dimensional image feature with each layer of the low-dimensional image features of the image in the database.
  • the center compares and determines the similarity degree of the low-dimensional image feature of the image to be retrieved and the image in the database, so that the feature matching the high-dimensional image feature of the image to be retrieved is retrieved in the database, and the accuracy of the feature matching can be improved.
  • FIG. 1 is a schematic flow chart of some embodiments of a high-dimensional image feature matching method according to the present disclosure.
  • FIG. 2 is a schematic flow chart of some embodiments of an offline training process of the high-dimensional image feature matching method of the present disclosure.
  • Figure 3 is a schematic diagram of an inverted file.
  • FIG. 4 is a schematic flow chart of some embodiments of an online training process of the high-dimensional image feature matching method according to the present disclosure.
  • FIG. 5 is a schematic flow chart of still another embodiment of a high-dimensional image feature matching method according to the present disclosure.
  • FIG. 6 is a schematic structural diagram of some embodiments of a high-dimensional image feature matching apparatus according to the present disclosure.
  • FIG. 7 is a schematic structural diagram of still another embodiment of a high-dimensional image feature matching apparatus according to the present disclosure.
  • FIG. 8 is a schematic structural diagram of still another embodiment of a high-dimensional image feature matching apparatus according to the present disclosure.
  • FIG. 9 is a schematic structural diagram of still another embodiment of a high-dimensional image feature matching apparatus according to the present disclosure.
  • FIG. 1 is a schematic flow chart of some embodiments of a high-dimensional image feature matching method according to the present disclosure. The method includes the following steps 110-140.
  • high dimensional image features of the image to be retrieved are extracted.
  • the high-dimensional image feature refers to image features exceeding 1000 dimensions
  • the deep neural network can be used to extract high-dimensional image features of the image to be retrieved.
  • the depth neural network can obtain 4096 dimensions for each output layer.
  • the linear combination of the characteristics of the latter two output layers can be used to describe the image, that is, an image can be represented by one 8192-dimensional feature.
  • the high dimensional image features of the image to be retrieved are divided into a plurality of low dimensional image features.
  • the division formula is as follows:
  • each low dimensional image feature of the image to be retrieved is compared to a respective cluster center of the low dimensional image feature of the image in the database.
  • the high-dimensional image features of the images in the database may be extracted in advance, and the high-dimensional image features of the images in the database are divided into a plurality of low-dimensional image features, and then the low-dimensional image features of the images in the database are hierarchically clustered and determined to be low.
  • Each layer of the image feature is clustered, and finally each low-dimensional image feature of the image to be retrieved is compared with each cluster center.
  • a low-dimensional image feature similarity of the image to be retrieved and the image in the database is determined based on the comparison result, thereby retrieving features in the database that match the high-dimensional image features of the image to be retrieved. For example, the similarity of image features is first scored, and the high-dimensional image features of the N images with the highest score are used as features matching the high-dimensional image features of the image to be retrieved.
  • the high-dimensional image features of the image to be retrieved are divided into a plurality of low-dimensional image features, and then each low-dimensional image feature is compared with each cluster center of the low-dimensional image features of the image in the database to determine The low-dimensional image feature similarity between the image to be retrieved and the image in the database, so that the feature matching the high-dimensional image feature of the image to be retrieved is retrieved in the database, and the accuracy of the feature matching can be improved.
  • the high-dimensional image feature matching method of the present disclosure includes an offline training process and an online retrieval process.
  • FIG. 2 is a schematic flowchart of some embodiments of an offline training process of the high-dimensional image feature matching method according to the present disclosure.
  • high dimensional image features of the image in the database are extracted.
  • a deep neural network can be used to extract high-dimensional image features of images in a database.
  • the high dimensional image features of the images in the database are divided into a plurality of low dimensional image features.
  • the high-dimensional image feature division manner of the database image and the image to be retrieved may be the same.
  • hierarchical clustering calculations are performed on low dimensional image features of the images in the database to determine clustering centers of the layers of the low dimensional image features.
  • unsupervised training can be performed on the set of low-dimensional image features of the image in the database, and the branching factor of the cluster is defined as k, and k is a natural number.
  • a low-dimensional image feature data point is randomly selected in the initial training set as the initial cluster center of the kmeans clustering algorithm, and k initial cluster centers are determined according to the probability ratio of the point to the nearest center.
  • kmeans clustering algorithm to divide the set of low-dimensional image features of the image in the database into k sets, as the first layer of the kmeans tree, and then use the kmeans clustering algorithm to sequentially perform each set of the current layer in the kmeans tree. Iteratively divides until the leaf layer of the kmeans tree is reached, and the cluster center of each set of each layer of the low-dimensional image features of the image in the database is calculated from the cluster mean.
  • the offline training process may further include a step 240 of associating the cluster center of each layer of each low-dimensional image feature of the image in the database with the inverted file.
  • each cluster center can be defined as a visual word, and each visual word is associated with an inverted file, wherein the inverted file counts the frequency of visual vocabulary appearing in each document, wherein The inverted file is shown in Figure 3. 1, 2...9 represent visual words.
  • each visual word w j can be measured using an inverse document frequency.
  • the calculation formula for the inverted file frequency is as follows:
  • idf(i) is the inverted file frequency
  • N is the number of training images
  • n j is the number of times the visual word w j appears in the training image.
  • the intermediate node with the smallest L2 distance is selected in each layer, and the image feature vector is from the first layer of the kmeans tree to the leaf layer. Perform traversal.
  • the term frequency of each word in the computed image I t is as follows:
  • tf(i, It) is the word frequency of each vocabulary in the image I t ;
  • n iIt represents the number of times the vocabulary appears in the image, and
  • n It represents the total number of words in the image.
  • the i-th item of the vector v t is:
  • step 410 high dimensional image features of the image to be retrieved are extracted.
  • the high dimensional image features of the image to be retrieved are divided into a plurality of low dimensional image features.
  • each low dimensional image feature is compared to each layer cluster center to achieve feature quantization.
  • Each low-dimensional image feature of the image to be retrieved is compared with the first layer of the kmeans tree of the low-dimensional image feature of the image in the database, and compared with each of the k cluster centers, determining each low of the image to be retrieved
  • the dimension image feature is closest to the cluster center, and the recursion is compared with the sub-cluster center of the nearest cluster center until the leaf layer of the kmeans tree is reached. This process quantifies each low-dimensional image feature.
  • the computational cost required for quantization is calculated as kL times. When k is not large, the quantization process is very fast.
  • the TF-IDF (Term Frequency Inverse Document Frequency) is used to score the low-dimensional image feature similarity of the image to be retrieved from the image in the database.
  • a weight can be set for each layer of the kmeans tree, and then a corresponding score is added to the low-dimensional image features that pass through the same layer. Since the amount of information contained in different layers is different, the weights are also different. When the two low-dimensional image features are close to the leaf layer, the two features are more similar, and the weight of the layer is also larger, and the weight is smaller when it is close to the first layer.
  • the weight of the layer number i is set according to the information entropy as:
  • N is the number of images in the database
  • N i is the number of images of at least one low-dimensional image feature passing through the number of layers i in the database.
  • n i are the number of low-dimensional image features passing through the number i of layers in the image to be retrieved and the database image, respectively.
  • the similarity score between the two image features is:
  • step 450 the fast selection algorithm is combined with the insertion ranking method to obtain the N images with the highest score, and the high-dimensional image features of the N images with the highest score are used as the features matching the high-dimensional image features of the image to be retrieved.
  • a fast selection algorithm can be used to find the Nth largest similarity score, and an unsorted top N large similarity score can be obtained.
  • N Usually N ⁇ 10, so the N similarity scores are sorted using the insertion ordering method to obtain the search results.
  • the fast selection algorithm has a complexity of O(n), and the insertion sorting consumption time is negligible, thereby greatly improving the retrieval efficiency.
  • the offline image and the online search are respectively divided into high-dimensional image features of the image in the database and the image to be retrieved into a plurality of low-dimensional image features, and then low-dimensional image features and low-dimensional images of the images in the database are respectively
  • the clustering centers of the image features are compared to determine the similarity of the low-dimensional image features of the image to be retrieved and the image in the database, so that the features matching the high-dimensional image features of the image to be retrieved are retrieved in the database, and the feature matching can be improved.
  • the accuracy rate in addition, because the fast selection algorithm and the insertion ordering method are used to obtain the N images with the highest score, the high-dimensional image features of the N images with the highest score are used as the features matching the high-dimensional image features of the image to be retrieved, It also improves the efficiency of high-dimensional image feature matching.
  • FIG. 5 is a schematic flow chart of still another embodiment of a high-dimensional image feature matching method according to the present disclosure.
  • the computer used in the experiment is Intel Xeon E5-2630, clocked at 2.30GHz and 64G memory.
  • the method includes the following steps 510-570.
  • high dimensional image features of the image in the database are extracted. For example, using a deep neural network, 200,000 high-dimensional image features are extracted for 200,000 women's images.
  • step 512 high dimensional image features of the image to be retrieved are extracted.
  • each output layer of the deep neural network can acquire features of 4096 dimensions and a linear combination of 8192 features, and divides the 8192-dimensional features into 128 64-dimensional sub-features.
  • the high dimensional image features of the image to be retrieved are divided into a plurality of low dimensional image features.
  • Each 8192-dimensional feature of the image to be retrieved is also divided into 128 64-dimensional sub-features.
  • step 530 hierarchical clustering calculations are performed on low dimensional image features of the images in the database to determine clustering centers of the layers of the low dimensional image features.
  • the initial clustering center of the kmeans clustering algorithm can be selected using the kmeans++ method.
  • the kmeans++ algorithm selects the initial center based on the probability ratio of the point to the nearest center, as follows:
  • a point of fixed grid distance may also be selected as the initial cluster center, or a random initialization method, a genetic algorithm initialization method, or the like may be selected to select the initial cluster center.
  • the kmeans clustering algorithm is an unsupervised clustering algorithm.
  • the steps of the algorithm are as follows:
  • cluster mean n i is the number of vectors belonging to cluster i
  • p x is a vector belonging to cluster i.
  • x j is the data vector
  • S i is the cluster where x j is located
  • u i is the average of the points in the cluster S i .
  • the clustering centers of the respective layers of each low-dimensional image feature of the image in the database are associated with the inverted file so that the clustering center of the low-dimensional image features can be quickly found during online retrieval.
  • each low dimensional image feature is compared to a respective cluster center to achieve feature quantization.
  • the low-dimensional image feature similarity of the image to be retrieved from the image in the database is scored using the TF-IDF.
  • step 570 the N images with the highest score are obtained by using the quick selection algorithm in combination with the insertion ordering method, and the high-dimensional image features of the N images with the highest score are used as features matching the high-dimensional image features of the image to be retrieved.
  • the result returned by the search is a correct match
  • the result is recorded as accurate, and the surface retrieval accuracy rate of the experimental result is as high as 99.9470%.
  • the retrieval time is less than 0.015 seconds, wherein the feature quantization time of step 550 is about 0.001 seconds, the scoring time of step 560 is about 0.010 seconds, and the first eight scoring image times of step 570 are about 0.003 seconds.
  • a deep neural network is used to extract 1 million 8192 dimensional image features for 1 million women's clothing images.
  • the 8192-dimensional feature is divided into 128 64-dimensional sub-features, and hierarchical clustering is performed using the method in the above embodiment.
  • the search session uses all 1 million high-dimensional image features for retrieval, and each 8192-dimensional feature is divided into 128 64-dimensional sub-features.
  • the retrieval accuracy rate is as high as 96.2927%.
  • the retrieval time is less than 0.025 seconds, wherein the feature quantization time of step 550 is about 0.002 seconds, the scoring time of step 560 is about 0.008 seconds, and the first 8 scoring image times of step 570 is about 0.012 seconds.
  • the high-dimensional image features are divided into a plurality of low-dimensional image features, so that they can be retrieved by the hierarchical kmeans method, and the initial cluster center is selected by using the kmeans++ algorithm, and the initial cluster center is randomly selected.
  • the clustering effect and clustering efficiency are improved.
  • the fast selection algorithm combined with the insertion ordering method is used to obtain the first N score corresponding images, and the algorithm speed is improved.
  • FIG. 6 is a schematic structural diagram of some embodiments of a high-dimensional image feature matching apparatus according to the present disclosure.
  • the apparatus includes a first high-dimensional image feature extraction unit 610, a first low-dimensional image feature division unit 620, a feature comparison unit 630, and a feature matching determination unit 640.
  • the first high-dimensional image feature extraction unit 610 is configured to extract high-dimensional image features of the image to be retrieved.
  • a deep neural network can be used to extract high-dimensional image features of the image to be retrieved.
  • the first low-dimensional image feature dividing unit 620 is configured to divide the high-dimensional image feature of the image to be retrieved into a plurality of low-dimensional image features.
  • the feature comparison unit 630 is configured to compare each low-dimensional image feature of the image to be retrieved with each layer cluster center of the low-dimensional image feature of the image in the database.
  • the high-dimensional image features of the images in the database may be extracted in advance, and the high-dimensional image features of the images in the database are divided into a plurality of low-dimensional image features, and then the low-dimensional image features of the images in the database are hierarchically clustered and determined to be low.
  • Each layer of the image feature is clustered and then each low-dimensional image feature of the image to be retrieved is compared to the cluster center of each layer.
  • the feature matching determining unit 640 is configured to determine, according to the comparison result, a low-dimensional image feature similarity of the image to be retrieved and the image in the database, thereby retrieving a feature in the database that matches the high-dimensional image feature of the image to be retrieved. For example, the similarity of image features is first scored, and the high-dimensional image features of the N images with the highest score are used as features matching the high-dimensional image features of the image to be retrieved.
  • the high-dimensional image features of the image to be retrieved are divided into a plurality of low-dimensional image features, and then each low-dimensional image feature is compared with each cluster center of the low-dimensional image features of the image in the database to determine The low-dimensional image feature similarity between the image to be retrieved and the image in the database, so that the feature matching the high-dimensional image feature of the image to be retrieved is retrieved in the database, the accuracy of the feature matching can be improved, and the feature matching of the high-dimensional image is improved. s efficiency.
  • FIG. 7 is a schematic structural diagram of still another embodiment of a high-dimensional image feature matching apparatus according to the present disclosure.
  • the apparatus includes a first high-dimensional image feature extraction unit 740, a first low-dimensional image feature division unit 750, a feature comparison unit 760, and a feature matching determination unit 770.
  • the cluster center of each layer of the low-dimensional image feature of the image in the database is determined by the second high-dimensional image feature extracting unit 710, the second low-dimensional image feature dividing unit 720, and the cluster center determining unit 730.
  • the second high dimensional image feature extraction unit 710 is configured to extract high dimensional image features of the image in the database.
  • the second low-dimensional image feature dividing unit 720 is configured to divide the high-dimensional image features of the image in the database into a plurality of low-dimensional image features.
  • the high-dimensional image feature division manner of the database image and the image to be retrieved may be the same.
  • the cluster center determining unit 730 is configured to perform hierarchical clustering calculation on the low-dimensional image features of the image in the database to determine the cluster center of each layer of the low-dimensional image feature.
  • unsupervised training can be performed on the set of low-dimensional image features of the image in the database, and the branching factor of the cluster is defined as k, and k is a natural number.
  • a low-dimensional image feature data point is randomly selected in the initial training set as the initial cluster center of the kmeans clustering algorithm, and k initial cluster centers are determined according to the probability ratio of the point to the nearest center.
  • kmeans clustering algorithm to divide the set of low-dimensional image features of the image in the database into k sets, as the first layer of the kmeans tree, and then use the kmeans clustering algorithm to sequentially perform each set of the current layer in the kmeans tree. Iteratively divides until the leaf layer of the kmeans tree is reached, and the cluster center of each set of each layer of the low-dimensional image features of the image in the database is calculated from the cluster mean.
  • the first high-dimensional image feature extraction unit 740 is configured to extract high-dimensional image features of the image to be retrieved.
  • the first low-dimensional image feature dividing unit 750 is configured to divide the high-dimensional image feature of the image to be retrieved into a plurality of low-dimensional image features.
  • the feature comparison unit 760 is configured to compare the respective low-dimensional image features with the respective cluster centers to achieve feature quantization.
  • Each low-dimensional image feature of the image to be retrieved is compared with the first layer of the kmeans tree of the low-dimensional image feature of the image in the database, and compared with each of the k cluster centers, determining each low of the image to be retrieved.
  • the dimension image feature is closest to the cluster center, and the recursion is compared with the sub-cluster center of the nearest cluster center until the leaf layer of the kmeans tree is reached. This process quantifies each low-dimensional image feature.
  • the computational cost required for quantization is calculated as kL times. When k is not large, the quantization process is very fast.
  • the feature matching determining unit 770 is configured to compare each low-dimensional image feature of the image to be retrieved from the first layer of the kmeans tree of the low-dimensional image feature of the image in the database, and compare with each of the k cluster centers to determine and Each low-dimensional image feature of the image to be retrieved is closest to the cluster center; the recursion is compared with the sub-cluster center of the nearest cluster center until the leaf layer of the kmeans tree is reached.
  • the feature matching determining unit 770 is further configured to acquire the N images with the highest score by using the quick selection algorithm in combination with the insertion sorting method, and use the high-dimensional image features of the N images with the highest score as the features matching the high-dimensional image features of the image to be retrieved.
  • a fast selection algorithm can be used to find the Nth largest similarity score, and an unsorted top N large similarity score can be obtained.
  • N Usually N ⁇ 10, so the N similarity scores are sorted using the insertion ordering method to obtain the search results.
  • the fast selection algorithm has a complexity of O(n), and the insertion sorting consumption time is negligible, thereby greatly improving the retrieval efficiency.
  • the cluster center of each layer of each low-dimensional image feature of the image in the database may also be associated with the inverted file.
  • the offline image and the online search are respectively divided into high-dimensional image features of the image in the database and the image to be retrieved into a plurality of low-dimensional image features, and then low-dimensional image features and low-dimensional images of the images in the database are respectively
  • the clustering centers of the image features are compared to determine the similarity of the low-dimensional image features of the image to be retrieved and the image in the database, so that the features matching the high-dimensional image features of the image to be retrieved are retrieved in the database, and the feature matching can be improved.
  • the accuracy rate in addition, because the fast selection algorithm and the insertion ordering method are used to obtain the N images with the highest score, the high-dimensional image features of the N images with the highest score are used as the features matching the high-dimensional image features of the image to be retrieved, It also improves the efficiency of high-dimensional image feature matching.
  • FIG. 8 is a schematic structural diagram of still another embodiment of a high-dimensional image feature matching apparatus according to the present disclosure.
  • the device includes a memory 810 and a processor 820.
  • Memory 810 can be a magnetic disk, flash memory, or any other non-volatile storage medium.
  • the memory is used to store the instructions in the embodiment corresponding to Figures 1-4.
  • the processor 820 is coupled to the memory 810 and can be implemented as one or more integrated circuits, such as a microprocessor or a microcontroller.
  • the processor 820 is configured to execute instructions stored in the memory, can improve the accuracy of feature matching, and also improve the efficiency of high-dimensional image feature matching.
  • the high-dimensional image feature matching apparatus 900 includes a memory 910 and a processor 920.
  • Processor 920 is coupled to memory 910 via BUS bus 930.
  • the high dimensional image feature matching device 900 can also be coupled to the external storage device 950 via the storage interface 940 for invoking external data, and can also be connected to the network or another computer system (not shown) via the network interface 960. It will not be described in detail here.
  • the accuracy of feature matching can be improved, and the efficiency of feature matching of the high-dimensional image is also improved.
  • a computer readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the steps of the method of the embodiments of FIGS. 1, 2, 4, and 5.
  • a processor may implement the steps of the method of the embodiments of FIGS. 1, 2, 4, and 5.
  • the present disclosure may be provided as a method, apparatus, or computer program product. Accordingly, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment, or a combination of software and hardware aspects. Moreover, the present disclosure may take the form of a computer program product embodied on one or more computer-usable non-transitory storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. .
  • the computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device.
  • the apparatus implements the functions specified in one or more blocks of a flow or a flow and/or block diagram of the flowchart.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device such that a series of operational steps are performed on a computer or other programmable device to produce computer-implemented processing for execution on a computer or other programmable device.
  • the instructions provide steps for implementing the functions specified in one or more of the flow or in a block or blocks of a flow diagram.
  • the methods and apparatus of the present disclosure may be implemented in a number of ways.
  • the methods and apparatus of the present disclosure may be implemented in software, hardware, firmware or any combination of software, hardware, firmware.
  • the above-described sequence of steps for the method is for illustrative purposes only, and the steps of the method of the present disclosure are not limited to the order specifically described above unless otherwise specifically stated.
  • the present disclosure may also be embodied as programs recorded in a recording medium, the programs including machine readable instructions for implementing a method in accordance with the present disclosure.
  • the present disclosure also covers a recording medium storing a program for executing the method according to the present disclosure.

Landscapes

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

Abstract

本公开公开了一种高维图像特征匹配方法和装置，涉及图像检索领域。其中的方法包括：提取待检索图像的高维图像特征；将待检索图像的高维图像特征划分为多个低维图像特征；将待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较；根据比较结果确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征。本公开能够提高特征匹配的准确率，并且还能够提高高维图像特征匹配的效率。

Description

高维图像特征匹配方法和装置
本申请是以CN申请号为201710158472.3，申请日为2017年3月17的申请为基础，并主张其优先权，该CN申请的公开内容在此作为整体引入本申请中。
技术领域
本公开涉及图像检索领域，尤其涉及一种高维图像特征匹配方法和装置。
背景技术
特征匹配问题通常定义为近似最近邻搜索问题(Approximate Nearest Neighbor)，即在数据库特征中搜索与检索特征欧式距离最近的特征。高维特征指超过1000维的图像特征，由于维度灾难(Curse of dimensionality)的存在，高维特征的匹配很困难。
图像的描述方法包括局部特征和全局特征两类。图像的局部特征通常是数百个维数较低的特征，图像全局特征通常是维数较高的单个特征。目前，可以利用穷举搜索(exaustive search)实现全局特征的匹配，穷举搜索能够保证欧氏距离下最近邻的精确匹配。还可以通过学习二值码(learning binary codes)来进行特征匹配，该方法通过使用双线性映射，将高维特征转换为维度低但保留相似性的二值码。
发明内容
发明人认识到，利用穷举搜索实现全局特征的匹配效率极其低下，通过学习二值码来进行特征匹配，由于对高维特征进行了双线性映射，可能存在映射误差，导致匹配错误。
本公开要解决的一个技术问题是：提供一种高维图像特征匹配方法和装置能够提高特征匹配的准确率。
根据本公开的一些实施例，提出一种高维图像特征匹配方法，包括：提取待检索图像的高维图像特征；将待检索图像的高维图像特征划分为多个低维图像特征；将待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较；根据比较结果确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征。
可选地，数据库中图像的低维图像特征的各层聚类中心通过以下方法进行确定： 提取数据库中图像的高维图像特征；将数据库中图像的高维图像特征划分为多个低维图像特征；对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心。
可选地，数据库中图像的各低维图像特征的各层聚类中心与倒排文件关联。
可选地，对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心包括：对数据库中图像的低维图像特征组成的集合进行无监督训练，其中，定义聚类的分叉因子为k；在初始训练集中随机选择一个低维图像特征数据点作为kmeans聚类算法的初始聚类中心；根据点到最近中心的概率比值确定k个初始聚类中心；利用kmeans聚类算法将数据库中图像的低维图像特征组成的集合划分为k个集合，作为kmeans树的第一层；利用kmeans聚类算法依次对kmeans树中当前层的每个集合进行迭代划分，直到达到kmeans树的叶子层；根据聚类均值计算数据库中图像的低维图像特征的每一层的每个集合的聚类中心。
可选地，将待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较包括：将待检索图像的每个低维图像特征，从数据库中图像的低维图像特征的kmeans树的第一层开始，与每一层k个聚类中心比较，确定与待检索图像的每个低维图像特征距离最近的聚类中心；递归与距离最近的聚类中心的子聚类中心进行比较，直到达到kmeans树的叶子层。
可选地，该方法还包括：利用TF-IDF词频-逆向文件频率算法对待检索图像与数据库中图像的低维图像特征相似度进行评分；利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征。
根据本公开的另一些实施例，还提出一种高维图像特征匹配装置，包括：第一高维图像特征提取单元，用于提取待检索图像的高维图像特征；第一低维图像特征划分单元，用于将待检索图像的高维图像特征划分为多个低维图像特征；特征比较单元，用于将待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较；特征匹配确定单元，用于根据比较结果确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征。
可选地，数据库中图像的低维图像特征的各层聚类中心通过以下单元进行确定：第二高维图像特征提取单元，用于提取数据库中图像的高维图像特征；第二低维图像特征划分单元，用于将数据库中图像的高维图像特征划分为多个低维图像特征；聚类 中心确定单元，用于对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心。
可选地，数据库中图像的各低维图像特征的各层聚类中心与倒排文件关联。
可选地，聚类中心确定单元用于对数据库中图像的低维图像特征组成的集合进行无监督训练，其中，定义聚类的分叉因子为k；在初始训练集中随机选择一个低维图像特征数据点作为kmeans聚类算法的初始聚类中心；根据点到最近中心的概率比值确定k个初始聚类中心；利用kmeans聚类算法将数据库中图像的低维图像特征组成的集合划分为k个集合，作为kmeans树的第一层；利用kmeans聚类算法依次对kmeans树中当前层的每个集合进行迭代划分，直到达到kmeans树的叶子层；根据聚类均值计算数据库中图像的低维图像特征的kmeans树的每一层的每个集合的聚类中心。
可选地，特征匹配确定单元用于将待检索图像的每个低维图像特征，从数据库中图像的低维图像特征的kmeans树的第一层开始，与每一层k个聚类中心比较，确定与待检索图像的每个低维图像特征距离最近的聚类中心；递归与距离最近的聚类中心的子聚类中心进行比较，直到达到kmeans树的叶子层。
可选地，特征匹配确定单元还用于利用TF-IDF词频-逆向文件频率算法对待检索图像与数据库中图像的低维图像特征相似度进行评分；利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征。
根据本公开的另一些实施例，还提出一种高维图像特征匹配装置，包括：存储器；以及耦接至存储器的处理器，处理器被配置为基于存储在存储器的指令执行上述的方法。
根据本公开的另一些实施例，还提出一种计算机可读存储介质，其上存储有计算机程序指令，该指令被处理器执行时实现上述的方法的步骤。
与相关技术相比，本公开实施例通过将待检索图像的高维图像特征划分为多个低维图像特征，然后将各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较，确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征，能够提高特征匹配的准确率。
通过以下参照附图对本公开的示例性实施例的详细描述，本公开的其它特征及其优点将会变得清楚。
附图说明
构成说明书的一部分的附图描述了本公开的实施例，并且连同说明书一起用于解释本公开的原理。
参照附图，根据下面的详细描述，可以更加清楚地理解本公开，其中：
图1为本公开高维图像特征匹配方法的一些实施例的流程示意图。
图2为本公开高维图像特征匹配方法的离线训练过程的一些实施例的流程示意图。
图3为倒排文件示意图。
图4为本公开高维图像特征匹配方法的在线训练过程的一些实施例的流程示意图。
图5为本公开高维图像特征匹配方法的另一些实施例的流程示意图。
图6为本公开高维图像特征匹配装置的一些实施例的结构示意图。
图7为本公开高维图像特征匹配装置的另一些实施例的结构示意图。
图8为本公开高维图像特征匹配装置的再一些实施例的结构示意图。
图9为本公开高维图像特征匹配装置的又一些实施例的结构示意图。
具体实施方式
现在将参照附图来详细描述本公开的各种示例性实施例。应注意到：除非另外具体说明，否则在这些实施例中阐述的部件和步骤的相对布置、数字表达式和数值不限制本公开的范围。
同时，应当明白，为了便于描述，附图中所示出的各个部分的尺寸并不是按照实际的比例关系绘制的。
以下对至少一个示例性实施例的描述实际上仅仅是说明性的，决不作为对本公开及其应用或使用的任何限制。
对于相关领域普通技术人员已知的技术、方法和设备可能不作详细讨论，但在适当情况下，所述技术、方法和设备应当被视为授权说明书的一部分。
在这里示出和讨论的所有示例中，任何具体值应被解释为仅仅是示例性的，而不是作为限制。因此，示例性实施例的其它示例可以具有不同的值。
应注意到：相似的标号和字母在下面的附图中表示类似项，因此，一旦某一项在一个附图中被定义，则在随后的附图中不需要对其进行进一步讨论。
为使本公开的目的、技术方案和优点更加清楚明白，以下结合具体实施例，并参照附图，对本公开进一步详细说明。
图1为本公开高维图像特征匹配方法的一些实施例的流程示意图。该方法包括以下步骤110-140。
在步骤110，提取待检索图像的高维图像特征。其中，高维图像特征指超过1000维的图像特征，可以使用深度神经网络提取待检索图像的高维图像特征。深度神经网络每个输出层可获取的特征达到4096维，可以使用后两个输出层的特征线性组合为8192维特征来描述图像，即一幅图像可以由1个8192维特征表示。
在步骤120，将待检索图像的高维图像特征划分为多个低维图像特征。其中，将高维图像特征x划分为m个不同的低维图像特征u j，1≤j≤m，低维图像特征u j的维度为D *＝D/m，其中D是m的倍数，划分公式如下：
x,···x D*|···|x D-D*+1,···x D→u 1(x),···u m(x)  (1)
在步骤130，将待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较。其中，可以预先提取数据库中图像的高维图像特征，将数据库中图像的高维图像特征划分为多个低维图像特征，然后对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心，最后将待检索图像的每个低维图像特征与各层聚类中心进行比较。
在步骤140，根据比较结果确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征。例如，先对图像特征的相似度进行评分，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征。
在该实施例中，将待检索图像的高维图像特征划分为多个低维图像特征，然后将各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较，确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征，能够提高特征匹配的准确率。
本公开高维图像特征匹配方法包括离线训练过程和在线检索过程，其中，图2为本公开高维图像特征匹配方法的离线训练过程的一些实施例的流程示意图。
在步骤210，提取数据库中图像的高维图像特征。其中，可以使用深度神经网络提取数据库中图像的高维图像特征。
在步骤220，将数据库中图像的高维图像特征划分为多个低维图像特征。其中， 数据库图像和待检索图像的高维图像特征划分方式可以相同。
在步骤230，对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心。其中，可以对数据库中图像的低维图像特征组成的集合进行无监督训练，定义聚类的分叉因子为k，k为自然数。在初始训练集中随机选择一个低维图像特征数据点作为kmeans聚类算法的初始聚类中心，根据点到最近中心的概率比值确定k个初始聚类中心。然后利用kmeans聚类算法将数据库中图像的低维图像特征组成的集合划分为k个集合，作为kmeans树的第一层，然后利用kmeans聚类算法依次对kmeans树中当前层的每个集合进行迭代划分，直到达到kmeans树的叶子层，根据聚类均值计算数据库中图像的低维图像特征的每一层的每个集合的聚类中心。
为了在在线检索时能够快速查找到低维图像特征的聚类中心，该离线训练过程还可以包括步骤240，即将数据库中图像的各低维图像特征的各层聚类中心关联倒排文件。其中，可以将每个聚类中心定义为一个视觉词(visual word)，将每个视觉词都关联一个倒排文件，其中，倒排文件统计了视觉词汇在每篇文档中出现的频率，其中，倒排文件如图3所示，1、2…9表示视觉词。
在具体应用中，可以使用倒排文件频率(inverse document frequency)衡量每个视觉词w j。倒排文件频率计算公式如下：
Figure PCTCN2017119489-appb-000001
其中，idf(i)为倒排文件频率，N是训练图像的数量，n j是视觉词w j在训练图像中出现的次数。
为了将图像I t转换为词袋(bag-of-words)矢量v t∈R W，在每一层选择L2距离最小的中间结点，将图像特征矢量从kmeans树的第一层向叶子层进行遍历。计算图像I t中每个词汇的词频(term frequency)如下所示：
Figure PCTCN2017119489-appb-000002
其中，tf(i,It)为图像I t中每个词汇的词频；n iIt表示词汇在图像中出现的次数，n It表示图像中词汇的总数。
矢量v t的第i项为：
Figure PCTCN2017119489-appb-000003
在建立bag-of-words模型的同时维护一个倒排文件，为词汇库中的每个词汇w j存 储包含该词汇的图像I t的列表，这样在检索数据库时只需比较与待检索图像有相同词汇的图像。
在线训练过程如图4所示，在步骤410，提取待检索图像的高维图像特征。
在步骤420，将待检索图像的高维图像特征划分为多个低维图像特征。
在步骤430，将各个低维图像特征与各层聚类中心进行比较实现特征量化。将待检索图像的每个低维图像特征，从数据库中图像的低维图像特征的kmeans树的第一层开始，与每一层k个聚类中心比较，确定与待检索图像的每个低维图像特征距离最近的聚类中心，递归与距离最近的聚类中心的子聚类中心进行比较，直到达到kmeans树的叶子层，该过程实现了每个低维图像特征的量化。
如果kmeans树包含L层，则对于一个低维图像特征而言，量化共需要的计算成本为kL次距离计算，当k不是很大时，量化过程非常迅速。
在步骤440，利用TF-IDF(Term Frequency Inverse Document Frequency，词频-逆向文件频率)对待检索图像与数据库中图像的低维图像特征相似度进行评分。
在一些实施例中，可以给kmeans树的每一层设定一个权值，然后对经过同一层的低维图像特征加上相应的分数。由于不同层所含信息量不同，所以权值也有所不同。当两个低维图像特征接近叶子层时这两个特征更为相似，该层的权值也较大，当接近于第一层时权值较小。将层数i的权值根据信息熵设定为：
Figure PCTCN2017119489-appb-000004
d i＝m i×w i  (6)
其中，N为数据库中图像数，N i为数据库中至少有一个低维图像特征通过层数i的图像数。根据上述权值定义待检索矢量和数据库矢量：
q i＝n i×w i  (7)
其中m i、n i分别为待检索图像与数据库图像中通过层数i的低维图像特征个数。两幅图像特征间的相似性评分为：
Figure PCTCN2017119489-appb-000005
其中，可以使用L1范数进行计算：
Figure PCTCN2017119489-appb-000006
在步骤450，利用快速选择算法结合插入排序法获取评分最高的N个图像，将评 分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征。
计算得到待检索图像与数据库图像的相似度评分之后，需要获取前N个评分对应的图像作为检索结果。常用方法是对评分进行排序，获得前N个的图像。这种方法在数据库图像较多时效率低下，即使使用快速排序方法，时间复杂度也过高，为O(nlogn)。如检索100万的图像数据库，排序所用时间超过35秒。
而在该实施例中可以采用快速选择算法，寻找第N大的相似度评分，同时可获得未排序的前N大的相似度评分。通常N<10，因此使用插入排序法对这N个相似度评分进行排序，获得检索结果。快速选择算法复杂度为O(n)，插入排序消耗时间可忽略不计，由此可大幅度提高检索效率。
在该实施例中，通过离线训练和在线检索，分别将数据库中图像和待检索图像的高维图像特征划分为多个低维图像特征，然后将各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较，确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征，能够提高特征匹配的准确率，另外，由于利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征，因此还提高了高维图像特征匹配的效率。
图5为本公开高维图像特征匹配方法的另一些实施例的流程示意图。其中，实验所用计算机配置为Intel Xeon E5-2630，主频2.30GHz，64G内存。该方法包括以下步骤510-570。
在步骤510，提取数据库中图像的高维图像特征。例如，使用深度神经网络，对20万女装图像提取20万高维图像特征。
在步骤512，提取待检索图像的高维图像特征。
在步骤520，将数据库中图像的高维图像特征划分为多个低维图像特征。例如，深度神经网络每个输出层可获取的特征达到4096维，线性组合为8192维特征，将这8192维特征划分为128个64维子特征。
在步骤522，将待检索图像的高维图像特征划分为多个低维图像特征。即将待检索图像的每个8192维特征同样划分为128个64维子特征。
在步骤530，对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心。
在一些实施例中，可以使用kmeans++方法选择kmeans聚类算法的初始聚类中 心。kmeans++算法根据点到最近中心的概率比值选择初始中心，步骤如下：
(1)从数据点X中随机选择一个初始中心c1＝x。
(2)设置D(x)作为从一个数据点x到最近中心的最短欧氏距离。
(3)选择下一个中心ci，其中ci＝x'∈X，其概率值为D(x')2/ΣD(x)2。
(4)重复(2)和(3)直到选择所有的k个中心。
在一些实施例中，还可以选择固定栅格距离的点作为初始聚类中心，或随机初始化方法、遗传算法初始化方法等选择初始聚类中心。
kmeans聚类算法是一种无监督聚类算法。该算法的步骤如下：
(1)从n个数据矢量中任意选择k个矢量作为初始聚类中心。
(2)对于其他矢量，根据它们到这些聚类中心的欧氏距离，分别将它们分配给与其最近的聚类。
(3)根据聚类均值计算新的聚类中心，聚类均值
Figure PCTCN2017119489-appb-000007
n i为属于聚类i的矢量个数，p x为属于聚类i的矢量。
(4)如此循环，直到目标函数值满足终止条件，最终将数据分为k类。其中，可以采用误差平方和准则函数作为目标函数：
Figure PCTCN2017119489-appb-000008
其中，x j为数据矢量，S i为x j所处的聚类，u i为聚类S i中点的平均值。
在步骤540，将数据库中图像的各低维图像特征的各层聚类中心关联倒排文件，以便在在线检索时能够快速查找到低维图像特征的聚类中心。
在步骤550，将各个低维图像特征与各层聚类中心进行比较实现特征量化。
在步骤560，利用TF-IDF对待检索图像与数据库中图像的低维图像特征相似度进行评分。
在步骤570，利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征。
在该实施例中，若检索返回的第一个结果为正确匹配，则记为结果准确，实验结果表面检索准确率高达99.9470％。检索时间小于0.015秒，其中步骤550的特征量化时间约为0.001秒，步骤560的评分时间约为0.010秒，步骤570的获取前8个评分图像时间约为0.003秒。
在另一些实施例中，使用深度神经网络，对100万女装图像提取100万8192维 图像特征。将这8192维特征划分为128个64维子特征，使用上述实施例中的方法进行分层聚类。检索环节，使用所有的100万高维图像特征进行检索，每个8192维特征划分为128个64维子特征。检索准确率高达96.2927％。检索时间小于0.025秒，其中步骤550的特征量化时间约为0.002秒，步骤560的评分时间约为0.008秒，步骤570的获取前8个评分图像时间约为0.012秒。
通过上述各实施例中，将高维图像特征划分为多个低维图像特征，使之能够通过分层kmeans方法进行检索，由于使用kmeans++算法选择初始聚类中心，与随机选择初始聚类中心相比，提高了聚类效果和聚类效率，另外，使用快速选择算法结合插入排序法获取前N个评分对应图像，提高算法速度。
图6为本公开高维图像特征匹配装置的一些实施例的结构示意图。该装置包括第一高维图像特征提取单元610、第一低维图像特征划分单元620、特征比较单元630、特征匹配确定单元640。
第一高维图像特征提取单元610用于提取待检索图像的高维图像特征。其中，可以使用深度神经网络提取待检索图像的高维图像特征。
第一低维图像特征划分单元620用于将待检索图像的高维图像特征划分为多个低维图像特征。
特征比较单元630用于将待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较。其中，可以预先提取数据库中图像的高维图像特征，将数据库中图像的高维图像特征划分为多个低维图像特征，然后对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心，然后将待检索图像的每个低维图像特征与各层聚类中心进行比较。
特征匹配确定单元640用于根据比较结果确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征。例如，先对图像特征的相似度进行评分，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征。
在该实施例中，将待检索图像的高维图像特征划分为多个低维图像特征，然后将各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较，确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征，能够提高特征匹配的准确率，并且提高了高维图像特征匹配的效率。
图7为本公开高维图像特征匹配装置的另一些实施例的结构示意图。该装置包括第一高维图像特征提取单元740、第一低维图像特征划分单元750、特征比较单元760、特征匹配确定单元770。其中，数据库中图像的低维图像特征的各层聚类中心通过第二高维图像特征提取单元710、第二低维图像特征划分单元720和聚类中心确定单元730进行确定。
第二高维图像特征提取单元710用于提取数据库中图像的高维图像特征。
第二低维图像特征划分单元720用于将数据库中图像的高维图像特征划分为多个低维图像特征。其中，数据库图像和待检索图像的高维图像特征划分方式可以相同。
聚类中心确定单元730用于对数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心。其中，可以对数据库中图像的低维图像特征组成的集合进行无监督训练，定义聚类的分叉因子为k，k为自然数。在初始训练集中随机选择一个低维图像特征数据点作为kmeans聚类算法的初始聚类中心，根据点到最近中心的概率比值确定k个初始聚类中心。然后利用kmeans聚类算法将数据库中图像的低维图像特征组成的集合划分为k个集合，作为kmeans树的第一层，然后利用kmeans聚类算法依次对kmeans树中当前层的每个集合进行迭代划分，直到达到kmeans树的叶子层，根据聚类均值计算数据库中图像的低维图像特征的每一层的每个集合的聚类中心。
第一高维图像特征提取单元740用于提取待检索图像的高维图像特征。
第一低维图像特征划分单元750用于将待检索图像的高维图像特征划分为多个低维图像特征。
特征比较单元760用于将各个低维图像特征与各层聚类中心进行比较实现特征量化。将待检索图像的每个低维图像特征，从数据库中图像的低维图像特征的kmeans树的第一层开始，与每一层k个聚类中心比较，确定与待检索图像的每个低维图像特征距离最近的聚类中心，递归与距离最近的聚类中心的子聚类中心进行比较，直到达到kmeans树的叶子层，该过程实现了每个低维图像特征的量化。
如果kmeans树包含L层，则对于一个低维图像特征而言，量化共需要的计算成本为kL次距离计算，当k不是很大时，量化过程非常迅速。
特征匹配确定单元770用于将待检索图像的每个低维图像特征，从数据库中图像的低维图像特征的kmeans树的第一层开始，与每一层k个聚类中心比较，确定与待检索图像的每个低维图像特征距离最近的聚类中心；递归与距离最近的聚类中心的子 聚类中心进行比较，直到达到kmeans树的叶子层。
特征匹配确定单元770还用于利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征。
计算得到待检索图像与数据库图像的相似度评分之后，需要获取前N个评分对应的图像作为检索结果。常用方法是对评分进行排序，获得前N个的图像。这种方法在数据库图像较多时效率低下，即使使用快速排序方法，时间复杂度也过高，为O(nlogn)。如检索100万的图像数据库，排序所用时间超过35秒。
而在该实施例中可以采用快速选择算法，寻找第N大的相似度评分，同时可获得未排序的前N大的相似度评分。通常N<10，因此使用插入排序法对这N个相似度评分进行排序，获得检索结果。快速选择算法复杂度为O(n)，插入排序消耗时间可忽略不计，由此可大幅度提高检索效率。
为了在在线检索时能够快速查找到低维图像特征的聚类中心，还可以将数据库中图像的各低维图像特征的各层聚类中心关联倒排文件。
在该实施例中，通过离线训练和在线检索，分别将数据库中图像和待检索图像的高维图像特征划分为多个低维图像特征，然后将各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较，确定待检索图像与数据库中图像的低维图像特征相似度，从而在数据库中检索出与待检索图像的高维图像特征匹配的特征，能够提高特征匹配的准确率，另外，由于利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与待检索图像的高维图像特征匹配的特征，因此还提高了高维图像特征匹配的效率。
图8为本公开高维图像特征匹配装置的再一些实施例的结构示意图。该装置包括存储器810和处理器820。
存储器810可以是磁盘、闪存或其它任何非易失性存储介质。存储器用于存储图1-4所对应实施例中的指令。
处理器820耦接至存储器810，可以作为一个或多个集成电路来实施，例如微处理器或微控制器。该处理器820用于执行存储器中存储的指令，能够提高特征匹配的准确率，并且还提高了高维图像特征匹配的效率。
在一些实施例中，还可以如图9所示，高维图像特征匹配装置900包括存储器910和处理器920。处理器920通过BUS总线930耦合至存储器910。该高维图像特征匹 配装置900还可以通过存储接口940连接至外部存储装置950以便调用外部数据，还可以通过网络接口960连接至网络或者另外一台计算机系统(未标出)。此处不再进行详细介绍。
在该实施例中，通过存储器存储数据指令，再通过处理器处理上述指令，能够提高特征匹配的准确率，并且还提高了高维图像特征匹配的效率。
在另一些实施例中，一种计算机可读存储介质，其上存储有计算机程序指令，该指令被处理器执行时实现图1、2、4、5所对应实施例中的方法的步骤。本领域内的技术人员应明白，本公开的实施例可提供为方法、装置、或计算机程序产品。因此，本公开可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本公开可采用在一个或多个其中包含有计算机可用程序代码的计算机可用非瞬时性存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本公开是参照根据本公开实施例的方法、设备(系统)和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
至此，已经详细描述了本公开。为了避免遮蔽本公开的构思，没有描述本领域所公知的一些细节。本领域技术人员根据上面的描述，完全可以明白如何实施这里公开的技术方案。
可能以许多方式来实现本公开的方法以及装置。例如，可通过软件、硬件、固件或者软件、硬件、固件的任何组合来实现本公开的方法以及装置。用于所述方法的步骤的上述顺序仅是为了进行说明，本公开的方法的步骤不限于以上具体描述的顺序，除非以其它方式特别说明。此外，在一些实施例中，还可将本公开实施为记录在记录介质中的程序，这些程序包括用于实现根据本公开的方法的机器可读指令。因而，本公开还覆盖存储用于执行根据本公开的方法的程序的记录介质。
虽然已经通过示例对本公开的一些特定实施例进行了详细说明，但是本领域的技术人员应该理解，以上示例仅是为了进行说明，而不是为了限制本公开的范围。本领域的技术人员应该理解，可在不脱离本公开的范围和精神的情况下，对以上实施例进行修改。本公开的范围由所附权利要求来限定。

Claims (14)

  1. 一种高维图像特征匹配方法，包括：
    提取待检索图像的高维图像特征；
    将所述待检索图像的高维图像特征划分为多个低维图像特征；
    将所述待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较；
    根据比较结果确定所述待检索图像与所述数据库中图像的低维图像特征相似度，从而在所述数据库中检索出与所述待检索图像的高维图像特征匹配的特征。
  2. 根据权利要求1所述的方法，其中，所述数据库中图像的低维图像特征的各层聚类中心通过以下方法进行确定：
    提取所述数据库中图像的高维图像特征；
    将所述数据库中图像的高维图像特征划分为多个低维图像特征；
    对所述数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心。
  3. 根据权利要求2所述的方法，其中，所述数据库中图像的各低维图像特征的各层聚类中心与倒排文件关联。
  4. 根据权利要求2所述的方法，其中，对所述数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心包括：
    对所述数据库中图像的低维图像特征组成的集合进行无监督训练，其中，定义聚类的分叉因子为k；
    在初始训练集中随机选择一个低维图像特征数据点作为kmeans聚类算法的初始聚类中心；
    根据点到最近中心的概率比值确定k个初始聚类中心；
    利用kmeans聚类算法将所述数据库中图像的低维图像特征组成的集合划分为k个集合，作为kmeans树的第一层；
    利用kmeans聚类算法依次对kmeans树中当前层的每个集合进行迭代划分，直 到达到kmeans树的叶子层；
    根据聚类均值计算所述数据库中图像的低维图像特征的每一层的每个集合的聚类中心。
  5. 根据权利要求4所述的方法，其中，将所述待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较包括：
    将所述待检索图像的每个低维图像特征，从所述数据库中图像的低维图像特征的kmeans树的第一层开始，与每一层k个聚类中心比较，确定与所述待检索图像的每个低维图像特征距离最近的聚类中心；
    递归与距离最近的聚类中心的子聚类中心进行比较，直到达到kmeans树的叶子层。
  6. 根据权利要求1-5任一所述的方法，还包括：
    利用TF-IDF词频-逆向文件频率算法对所述待检索图像与所述数据库中图像的低维图像特征相似度进行评分；
    利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与所述待检索图像的高维图像特征匹配的特征。
  7. 一种高维图像特征匹配装置，包括：
    第一高维图像特征提取单元，用于提取待检索图像的高维图像特征；
    第一低维图像特征划分单元，用于将所述待检索图像的高维图像特征划分为多个低维图像特征；
    特征比较单元，用于将所述待检索图像的各低维图像特征与数据库中图像的低维图像特征的各层聚类中心进行比较；
    特征匹配确定单元，用于根据比较结果确定所述待检索图像与所述数据库中图像的低维图像特征相似度，从而在所述数据库中检索出与所述待检索图像的高维图像特征匹配的特征。
  8. 根据权利要求7所述的装置，其中，所述数据库中图像的低维图像特征的各层聚类中心通过以下单元进行确定：
    第二高维图像特征提取单元，用于提取所述数据库中图像的高维图像特征；
    第二低维图像特征划分单元，用于将所述数据库中图像的高维图像特征划分为多个低维图像特征；
    聚类中心确定单元，用于对所述数据库中图像的低维图像特征进行分层聚类计算确定低维图像特征的各层聚类中心。
  9. 根据权利要求8所述的装置，其中，所述数据库中图像的各低维图像特征的各层聚类中心与倒排文件关联。
  10. 根据权利要求8所述的装置，其中，所述聚类中心确定单元用于对所述数据库中图像的低维图像特征组成的集合进行无监督训练，其中，定义聚类的分叉因子为k；在初始训练集中随机选择一个低维图像特征数据点作为kmeans聚类算法的初始聚类中心；根据点到最近中心的概率比值确定k个初始聚类中心；利用kmeans聚类算法将所述数据库中图像的低维图像特征组成的集合划分为k个集合，作为kmeans树的第一层；利用kmeans聚类算法依次对kmeans树中当前层的每个集合进行迭代划分，直到达到kmeans树的叶子层；根据聚类均值计算所述数据库中图像的低维图像特征的kmeans树的每一层的每个集合的聚类中心。
  11. 根据权利要求10的装置，其中，所述特征匹配确定单元用于将所述待检索图像的每个低维图像特征，从所述数据库中图像的低维图像特征的kmeans树的第一层开始，与每一层k个聚类中心比较，确定与所述待检索图像的每个低维图像特征距离最近的聚类中心；递归与距离最近的聚类中心的子聚类中心进行比较，直到达到kmeans树的叶子层。
  12. 根据权利要求7-11任一所述的装置，其中，所述特征匹配确定单元还用于利用TF-IDF词频-逆向文件频率算法对所述待检索图像与所述数据库中图像的低维图像特征相似度进行评分；利用快速选择算法结合插入排序法获取评分最高的N个图像，将评分最高的N个图像的高维图像特征作为与所述待检索图像的高维图像特征匹配的特征。
  13. 一种高维图像特征匹配装置，包括：
    存储器；以及
    耦接至所述存储器的处理器，所述处理器被配置为基于存储在所述存储器的指令执行如权利要求1至6任一项所述的方法。
  14. 一种计算机可读存储介质，其上存储有计算机程序指令，该指令被处理器执行时实现权利要求1至6任一项所述的方法的步骤。
PCT/CN2017/119489 2017-03-17 2017-12-28 高维图像特征匹配方法和装置 WO2018166273A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/494,632 US11210555B2 (en) 2017-03-17 2017-12-28 High-dimensional image feature matching method and device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710158472.3A CN108629345B (zh) 2017-03-17 2017-03-17 高维图像特征匹配方法和装置
CN201710158472.3 2017-03-17

Publications (1)

Publication Number Publication Date
WO2018166273A1 true WO2018166273A1 (zh) 2018-09-20

Family

ID=63522754

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/119489 WO2018166273A1 (zh) 2017-03-17 2017-12-28 高维图像特征匹配方法和装置

Country Status (3)

Country Link
US (1) US11210555B2 (zh)
CN (1) CN108629345B (zh)
WO (1) WO2018166273A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110472079A (zh) * 2019-07-08 2019-11-19 浙江省北大信息技术高等研究院 目标图像的检索方法、装置、设备及存储介质
CN111581413A (zh) * 2020-04-03 2020-08-25 北京联合大学 一种面向高维图像数据检索的数据过滤方法及系统
CN111651624A (zh) * 2020-06-11 2020-09-11 浙江大华技术股份有限公司 一种图像检索方法及装置

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10885627B2 (en) * 2018-04-03 2021-01-05 Nec Corporation Unsupervised neighbor-preserving embedding for image stream visualization and anomaly detection
CN109376940B (zh) * 2018-11-02 2021-08-17 中国水利水电科学研究院 获取降雨过程中的降雨时空分布规律的方法和装置
CN113297331B (zh) * 2020-09-27 2022-09-09 阿里云计算有限公司 数据存储方法及装置、数据查询方法及装置
CN117390013A (zh) * 2023-09-12 2024-01-12 博瀚智能(深圳)有限公司 数据存储方法、检索方法、系统、设备及存储介质
CN116912925A (zh) * 2023-09-14 2023-10-20 齐鲁空天信息研究院 人脸识别方法、装置、电子设备及介质

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101038622A (zh) * 2007-04-19 2007-09-19 上海交通大学 基于几何保存的人脸子空间识别方法
JP2008102856A (ja) * 2006-10-20 2008-05-01 Matsushita Electric Ind Co Ltd パターン識別装置およびパターン識別方法
CN101211355A (zh) * 2006-12-30 2008-07-02 中国科学院计算技术研究所 一种基于聚类的图像查询方法
WO2009058915A1 (en) * 2007-10-29 2009-05-07 The Trustees Of The University Of Pennsylvania Computer assisted diagnosis (cad) of cancer using multi-functional, multi-modal in-vivo magnetic resonance spectroscopy (mrs) and imaging (mri)
CN102004917A (zh) * 2010-12-17 2011-04-06 南方医科大学 一种图像边缘近邻描述特征算子的提取方法
CN102073748A (zh) * 2011-03-08 2011-05-25 武汉大学 一种基于视觉关键词的遥感影像语义检索方法
CN104008174A (zh) * 2014-06-04 2014-08-27 北京工业大学 一种海量图像检索的隐私保护索引生成方法
CN104239859A (zh) * 2014-09-05 2014-12-24 西安电子科技大学 基于结构化因子分析的人脸识别方法
CN104699781A (zh) * 2015-03-12 2015-06-10 西安电子科技大学 基于双层锚图散列的sar图像检索方法
CN104765764A (zh) * 2015-02-06 2015-07-08 南京理工大学 一种基于大规模图像检索方法

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101989326B (zh) * 2009-07-31 2015-04-01 三星电子株式会社 人体姿态识别方法和装置
CN102123172B (zh) * 2011-02-25 2014-09-10 南京邮电大学 一种基于神经网络聚类优化的Web服务发现的实现方法
CN102760128A (zh) * 2011-04-26 2012-10-31 华东师范大学 一种基于智能客服机器人交互的电信领域套餐推荐方法
JP5923744B2 (ja) * 2012-05-24 2016-05-25 パナソニックＩｐマネジメント株式会社 画像検索システム、画像検索方法及び検索装置
CN103034963B (zh) * 2012-11-28 2017-10-27 东南大学 一种基于相关性的服务选择系统及选择方法
CN104778284B (zh) * 2015-05-11 2017-11-21 苏州大学 一种空间图像查询方法和系统
CN106326288B (zh) * 2015-06-30 2019-12-03 阿里巴巴集团控股有限公司 图像搜索方法及装置
CN105095435A (zh) 2015-07-23 2015-11-25 北京京东尚科信息技术有限公司 一种图像高维特征的相似比较方法及装置
CN105912611B (zh) * 2016-04-05 2019-04-26 中国科学技术大学 一种基于cnn的快速图像检索方法

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2008102856A (ja) * 2006-10-20 2008-05-01 Matsushita Electric Ind Co Ltd パターン識別装置およびパターン識別方法
CN101211355A (zh) * 2006-12-30 2008-07-02 中国科学院计算技术研究所 一种基于聚类的图像查询方法
CN101038622A (zh) * 2007-04-19 2007-09-19 上海交通大学 基于几何保存的人脸子空间识别方法
WO2009058915A1 (en) * 2007-10-29 2009-05-07 The Trustees Of The University Of Pennsylvania Computer assisted diagnosis (cad) of cancer using multi-functional, multi-modal in-vivo magnetic resonance spectroscopy (mrs) and imaging (mri)
CN102004917A (zh) * 2010-12-17 2011-04-06 南方医科大学 一种图像边缘近邻描述特征算子的提取方法
CN102073748A (zh) * 2011-03-08 2011-05-25 武汉大学 一种基于视觉关键词的遥感影像语义检索方法
CN104008174A (zh) * 2014-06-04 2014-08-27 北京工业大学 一种海量图像检索的隐私保护索引生成方法
CN104239859A (zh) * 2014-09-05 2014-12-24 西安电子科技大学 基于结构化因子分析的人脸识别方法
CN104765764A (zh) * 2015-02-06 2015-07-08 南京理工大学 一种基于大规模图像检索方法
CN104699781A (zh) * 2015-03-12 2015-06-10 西安电子科技大学 基于双层锚图散列的sar图像检索方法

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110472079A (zh) * 2019-07-08 2019-11-19 浙江省北大信息技术高等研究院 目标图像的检索方法、装置、设备及存储介质
CN110472079B (zh) * 2019-07-08 2022-04-05 杭州未名信科科技有限公司 目标图像的检索方法、装置、设备及存储介质
CN111581413A (zh) * 2020-04-03 2020-08-25 北京联合大学 一种面向高维图像数据检索的数据过滤方法及系统
CN111581413B (zh) * 2020-04-03 2023-02-28 北京联合大学 一种面向高维图像数据检索的数据过滤方法及系统
CN111651624A (zh) * 2020-06-11 2020-09-11 浙江大华技术股份有限公司 一种图像检索方法及装置
CN111651624B (zh) * 2020-06-11 2023-09-19 浙江大华技术股份有限公司 一种图像检索方法及装置

Also Published As

Publication number Publication date
US20200026958A1 (en) 2020-01-23
CN108629345A (zh) 2018-10-09
CN108629345B (zh) 2021-07-30
US11210555B2 (en) 2021-12-28

Similar Documents

Publication Publication Date Title
WO2018166273A1 (zh) 高维图像特征匹配方法和装置
US11048966B2 (en) Method and device for comparing similarities of high dimensional features of images
McCann et al. Local naive bayes nearest neighbor for image classification
Cheng et al. Fast and accurate image matching with cascade hashing for 3d reconstruction
CN110851645B (zh) 一种基于深度度量学习下相似性保持的图像检索方法
CN108038122B (zh) 一种商标图像检索的方法
Xu Multiple-instance learning based decision neural networks for image retrieval and classification
CN104281572B (zh) 一种基于互信息的目标匹配方法及其系统
Nayini et al. A novel threshold-based clustering method to solve K-means weaknesses
JP5833499B2 (ja) 高次元の特徴ベクトル集合で表現されるコンテンツを高精度で検索する検索装置及びプログラム
Johns et al. Pairwise probabilistic voting: Fast place recognition without RANSAC
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
Pei-Xia et al. Learning discriminative CNN features and similarity metrics for image retrieval
WO2015109781A1 (zh) 基于期望最大确定统计模型参数的方法和装置
CN110209895B (zh) 向量检索方法、装置和设备
JP6601965B2 (ja) 探索木を用いて量子化するプログラム、装置及び方法
Neumann et al. Propagation kernels for partially labeled graphs
Karampatziakis et al. Scalable multilabel prediction via randomized methods
Ha et al. Text-to-image retrieval based on incremental association via multimodal hypernetworks
Chen et al. Generating vocabulary for global feature representation towards commerce image retrieval
Altintakan et al. An improved BOW approach using fuzzy feature encoding and visual-word weighting
Iscen et al. Scaling group testing similarity search
Sinuraya et al. Accuracy Analysis on Images Retrieval System using Radial Basis Function Algorithm and Coefficient Correlation
Wang et al. Similarity search for image retrieval via local-constrained linear coding
Feng et al. Efficient indexing for mobile image retrieval

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: 17901037

Country of ref document: EP

Kind code of ref document: A1

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 (EPO FORM 1205A DATED 17.12.2019)

122 Ep: pct application non-entry in european phase

Ref document number: 17901037

Country of ref document: EP

Kind code of ref document: A1