CN111309955B - Fusion method for image retrieval - Google Patents

Fusion method for image retrieval Download PDF

Info

Publication number
CN111309955B
CN111309955B CN202010149889.5A CN202010149889A CN111309955B CN 111309955 B CN111309955 B CN 111309955B CN 202010149889 A CN202010149889 A CN 202010149889A CN 111309955 B CN111309955 B CN 111309955B
Authority
CN
China
Prior art keywords
image
sift descriptor
sift
hausdorff distance
basic probability
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN202010149889.5A
Other languages
Chinese (zh)
Other versions
CN111309955A (en
Inventor
孙晓明
张宁
车畅
刘野
吴海滨
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin University of Science and Technology
Original Assignee
Harbin University of Science and Technology
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 Harbin University of Science and Technology filed Critical Harbin University of Science and Technology
Priority to CN202010149889.5A priority Critical patent/CN111309955B/en
Publication of CN111309955A publication Critical patent/CN111309955A/en
Application granted granted Critical
Publication of CN111309955B publication Critical patent/CN111309955B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • 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
    • G06F18/25—Fusion techniques
    • 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
    • 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]

Abstract

The invention relates to an image retrieval-oriented fusion method, which comprises the steps of fusing SIFT descriptor kernel density and SIFT descriptor histogram and comprises the following steps: firstly, obtaining a basic probability distribution function of an SIFT descriptor histogram and SIFT descriptor kernel density, and then, applying a Dempster combination rule to combine the basic probability distribution function to obtain a fusion result; the fusion method is applied to the image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance, can improve the image retrieval accuracy and provide a theoretical basis for image retrieval with complex background.

Description

Fusion method for image retrieval
The application is a divisional application of the invention patent application, namely 'image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance'.
Application date of the original case: 2017-02-13.
Original application No.: 2017100760427.
the name of the original invention is: an image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance.
Technical Field
The invention discloses an image retrieval-oriented fusion method, belongs to the technical field of image retrieval, and particularly relates to a key step in an image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance.
Background
Image retrieval methods have so far formed three important branches: text-based image retrieval, content-based image retrieval, and semantic-based image retrieval.
The image retrieval based on the text is to describe the requirements of users by using texts such as image names, image characteristics and the like, but because the text expression capability is limited and text labels have ambiguity, the retrieval result is often inconsistent with the requirements of the users;
the semantic-based image retrieval further refines the high-level semantic expression capability of the image on the basis of the visual characteristics of the image, but the retrieval process of the retrieval method is complex, and the problem that the method system is not developed completely exists;
the image retrieval based on the content is carried out by taking color, texture, shape and the like as the characteristic expression of the image and taking the characteristic expression as the basis for judging the similarity.
If the image features can be extracted accurately, content-based image retrieval would have accuracy advantages not available with the other two types of retrieval. The technical advantages are also aimed at by broad scholars, and researches on how to improve the accuracy of image feature extraction are carried out, so that the accuracy of content-based image retrieval is expected to be further improved.
Disclosure of Invention
In order to meet the technical requirements, the invention discloses an image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance, which can effectively improve the accuracy of content-based image retrieval.
The purpose of the invention is realized as follows:
the image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance comprises the following steps:
step a, extracting an image to be retrieved and SIFT characteristics of an image library;
b, generating an SIFT descriptor histogram and SIFT descriptor kernel density;
step c, fusing SIFT descriptor kernel density and SIFT descriptor histogram;
step d, improving the traditional Hausdorff distance measurement;
and e, using the improved Hausdorff distance for image matching.
The image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance comprises the following specific steps of the step a:
step a 1: constructing Gaussian difference scale function of image to be retrieved and image library
Carrying out convolution operation on Gaussian functions with different scales and the image to construct a Gaussian difference scale function D (x, y, sigma) of the two-dimensional image, wherein the method comprises the following steps:
D(x,y,σ)＝(G(x,y,kσ)-G(x,y,σ))*I(x,y)
where k is the scale scaling factor, G (x, y, σ) is a gaussian function of variable scale, I (x, y) is the image, and there are:
Figure BDA0002402054190000021
wherein, (x, y) is a scale coordinate, and the size of σ determines the degree of image smoothness;
step a 2: detecting extreme point in Gaussian difference scale space
Comparing each sampling point in the image with adjacent points of the sampling point, and when a certain sampling point is the maximum value or the minimum value in all points of the Gaussian difference scale space, considering the certain sampling point as a characteristic point of the image under the scale;
step a 3: removing the feature points with unstable edges to generate SIFT descriptors
And removing unstable characteristic points of the edge by using a Harris Corner detector, and keeping stable characteristic points to generate the SIFT descriptor.
The image retrieval method based on the combination of the vocabulary tree information fusion and the Hausdorff distance comprises the following specific steps of the step b:
step b 1: construction of an extensible lexical tree by hierarchical clustering of SIFT descriptors
Extracting SIFT descriptors of each picture to obtain a set F ═ FiThen, the set F is subjected to hierarchical clustering by adopting a K-Means clustering method, and initially, the set F is subjected to K-Means clustering at the 1 st layer, and is divided into K parts of { F }iI is more than or equal to 1 and less than or equal to k; repeating the operation until the depth reaches the preset L value to construct an expandable vocabulary tree, wherein c is B in totalLA node, wherein B is a branching factor, L is a depth, c is a total number of nodes, fiRepresenting a certain SIFT descriptor in a picture, F being a set of descriptors, FiIs a certain cluster set obtained by carrying out K-Means clustering on the set F;
step b 2: accumulating the times of the occurrence of the descriptors on each node in the extensible vocabulary tree to obtain an SIFT descriptor histogram
In constructing an expandable lexical tree, c is shared as BLEach node accumulates the occurrence frequency of SIFT descriptors on the first node to obtain SIFT descriptor histogram based on the expandable vocabulary tree, and H is used as [ H ]1,...,hi,...,hc]Is represented by the formula (I) in which hiRepresenting the times of SIFT descriptors appearing on the ith node;
step b 3: quantizing the SIFT descriptor to obtain SIFT descriptor kernel density
Quantizing all SIFT descriptors, then each SIFT descriptor fiAll correspond to a slave in the extensible vocabulary treeQuantization path from root node to leaf node, i.e. corresponding to a set of visual words
Figure BDA0002402054190000031
Each group of visual words corresponds to the kernel density f (c) thereof, and the SIFT descriptor kernel density based on the extensible vocabulary tree is obtained; wherein
Figure BDA0002402054190000032
Is a visual word, i.e. each node in the expandable vocabulary tree represents a visual word, l represents the number of layers of the node in the expandable vocabulary tree, hlIndicating the index of the node in the level tree node, L being the depth.
The image retrieval method based on the combination of the vocabulary tree information fusion and the Hausdorff distance comprises the following specific steps of step c:
step c 1: obtaining a basic probability distribution function of SIFT descriptor histogram and SIFT descriptor kernel density
For computational convenience, if the SIFT descriptor histogram is set as a, and the SIFT descriptor kernel density is set as B, then the box Ω: { A, B }, the decision box is a set describing all elements constituting the whole hypothesis space, and is represented by m () with all possible results considered by the basic probability distribution function; at this time, the process of the present invention,
the basic probability distribution function of subset A is
Figure BDA0002402054190000033
The basic probability distribution function of the subset B is
Figure BDA0002402054190000034
Wherein, M is a normalization constant,
Figure BDA0002402054190000036
m1(Ai) Denotes that the focal length is AiBasic probability assignment of (c), m2(Bj) Denotes that the focal length is BjAssigning a basic probability;
step c 2: the fusion result is obtained by applying Dempster combination rule and combining the step c1
The Dempster combination rule is:
Figure BDA0002402054190000035
substituting the results m (A) and m (B) obtained in the step c1 into m (AB);
where M is a normalization constant, and M ═ SigmaA∩B＝φ(m(A)m(B))＝1-∑A∩B≠φ(m(A)m(B))
m (a) represents the basic probability distribution function of subset a, m (B) represents the basic probability distribution function of subset B, and m (ab) represents the fused basic probability distribution function of subset a and subset B.
The image retrieval method based on the combination of the vocabulary tree information fusion and the Hausdorff distance comprises the following specific steps of the step d:
step d 1: form of differential equation for writing cost function
The differential equation form of the cost function is as follows:
Figure BDA0002402054190000041
step d 2: obtaining a general solution to a cost function
Solving the differential equation to obtain the cost function with the following expression:
Figure BDA0002402054190000042
wherein gamma is0The cost function is an initial value of the cost function, the range of the cost function is 0-1, k is a proportionality coefficient, and tau is a matching parameter;
step d 3: improved Hausdorff distance using traditional Hausdorff distance as variable of cost function
Given two finite sets X ═ X1,x2,...,xMY ═ Y1,y2,...,yNThe conventional Hausdorff distance between X and Y is defined as
Figure BDA0002402054190000043
Where d (X, Y) is the conventional Hausdorff distance, min represents the minimum, max represents the maximum, X and Y are the points in the point sets X and Y, respectively, and d (X, Y) represents the geometric distance between point X and point Y;
the improved Hausdorff distance is:
Figure BDA0002402054190000044
where | X | is the number of the finite set X, dH(X, Y) is the modified Hausdorff distance, d (X, Y) is the conventional Hausdorff distance, and γ (d (X, Y)) is a cost function with a variable d (X, Y).
The image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance comprises the following specific steps of step e:
and c, according to the fusion characteristics obtained in the step c, carrying out image similarity measurement by using an improved Hausdorff distance, and arranging the obtained similarities according to a descending order to obtain a retrieval result.
Has the advantages that:
the method adopts the following technical means that firstly, SIFT features of an image to be retrieved and an image library are extracted, then an SIFT descriptor histogram and SIFT descriptor kernel density are generated, then the SIFT descriptor kernel density and the SIFT descriptor histogram are fused, the traditional Hausdorff distance measurement is improved, and finally the improved Hausdorff distance is used for image matching; the technical means are interdependent and are all absent, and as a whole, the technical purpose that the content-based image retrieval accuracy rate cannot be effectively improved when any one means exists is achieved.
Drawings
FIG. 1 is a flow chart of the image retrieval method based on the combination of lexical tree information fusion and Housdov distance.
FIG. 2 is a graph comparing precision ratios of three methods.
Fig. 3 is a "banyan" image to be retrieved.
Fig. 4 is a "banyan" search result based on the method of the present invention.
Fig. 5 is a "banyan" search result based on the SIFT descriptor histogram method.
Fig. 6 is a "banyan" search result based on the SIFT descriptor kernel density method.
FIG. 7 is a "tiger" image to be retrieved.
FIG. 8 is the "tiger" search result based on the method of the present invention.
Fig. 9 is a "tiger" search result based on the SIFT descriptor histogram method.
Fig. 10 is a "tiger" search result based on the SIFT descriptor kernel density method.
Detailed Description
The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.
Detailed description of the preferred embodiment
The present embodiment is a theoretical embodiment of an image retrieval method based on the combination of lexical tree information fusion and hausdorff distance.
The image retrieval method based on the combination of vocabulary tree information fusion and Hausdorff distance in the embodiment has a flow chart as shown in FIG. 1, and comprises the following steps:
step a, extracting an image to be retrieved and SIFT characteristics of an image library;
b, generating an SIFT descriptor histogram and SIFT descriptor kernel density;
step c, fusing SIFT descriptor kernel density and SIFT descriptor histogram;
step d, improving the traditional Hausdorff distance measurement;
and e, using the improved Hausdorff distance for image matching.
The image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance comprises the following specific steps of the step a:
step a 1: constructing Gaussian difference scale function of image to be retrieved and image library
The method comprises the following steps of performing convolution operation on Gaussian functions with different scales and an image to construct a Gaussian difference scale function D (x, y, sigma) of a two-dimensional image, wherein the method comprises the following steps:
D(x,y,σ)＝(G(x,y,kσ)-G(x,y,σ))*I(x,y)
where k is the scale scaling factor, G (x, y, σ) is a gaussian function of variable scale, I (x, y) is the image, and there are:
Figure BDA0002402054190000061
wherein, (x, y) is a scale coordinate, and the size of σ determines the degree of image smoothness;
step a 2: detecting extreme point in Gaussian difference scale space
Comparing each sampling point in the image with adjacent points of the sampling point, and when a certain sampling point is the maximum value or the minimum value in all points of the Gaussian difference scale space, considering the certain sampling point as a characteristic point of the image under the scale;
step a 3: removing the feature points with unstable edges to generate SIFT descriptors
And removing unstable characteristic points of the edge by using a Harris Corner detector, and keeping stable characteristic points to generate the SIFT descriptor.
The image retrieval method based on the combination of the vocabulary tree information fusion and the Hausdorff distance comprises the following specific steps of the step b:
step b 1: construction of an extensible lexical tree by hierarchical clustering of SIFT descriptors
Extracting SIFT descriptor of each picture to obtain a set F ═ FiThen, the set F is subjected to hierarchical clustering by adopting a K-Means clustering method, and initially, the set F is subjected to K-Means clustering at the 1 st layer, and is divided into K parts of { F }iI is more than or equal to 1 and less than or equal to k; by analogy, the newly generated cluster is subdivided into K clusters by using K-Means, the operations are continuously repeated until the depth reaches the preset L value,constructing an expandable vocabulary tree with c ═ BLA node, wherein B is a branching factor, L is a depth, c is a total number of nodes, fiRepresenting a certain SIFT descriptor in a picture, F being a set of descriptors, FiIs a certain cluster set obtained by carrying out K-Means clustering on the set F;
step b 2: accumulating the times of the occurrence of the descriptors on each node in the extensible vocabulary tree to obtain an SIFT descriptor histogram
In constructing an expandable lexical tree, c is shared as BLEach node accumulates the occurrence frequency of SIFT descriptors on the first node to obtain SIFT descriptor histogram based on the expandable vocabulary tree, and H is used as [ H ]1,...,hi,...,hc]Is represented by the formula (I) in which hiRepresenting the times of SIFT descriptors appearing on the ith node;
step b 3: quantizing the SIFT descriptor to obtain SIFT descriptor kernel density
Quantizing all SIFT descriptors, then each SIFT descriptor fiAll corresponding to a quantization path from a root node to a leaf node in the expandable vocabulary tree, i.e. corresponding to a group of visual words
Figure BDA0002402054190000071
Each group of visual words corresponds to the kernel density f (c) thereof, and the SIFT descriptor kernel density based on the extensible vocabulary tree is obtained; wherein
Figure BDA0002402054190000076
Is a visual word, i.e. each node in the expandable vocabulary tree represents a visual word, l represents the number of layers of the node in the expandable vocabulary tree, hlIndicating the index of the node in the level tree node, L being the depth.
The image retrieval method based on the combination of the vocabulary tree information fusion and the Hausdorff distance comprises the following specific steps of step c:
step c 1: obtaining a SIFT descriptor histogram and a basic probability distribution function of SIFT descriptor kernel density
For computational convenience, if the SIFT descriptor histogram is set as a, and the SIFT descriptor kernel density is set as B, then the box Ω: { A, B }, the decision box is a set describing all elements constituting the whole hypothesis space, and is represented by m () with all possible results considered by the basic probability distribution function; at this time, the process of the present invention,
the basic probability distribution function of subset A is
Figure BDA0002402054190000072
The basic probability distribution function of the subset B is
Figure BDA0002402054190000073
Wherein, M is a normalization constant,
Figure BDA0002402054190000074
m1(Ai) Denotes that the focal length is AiBasic probability assignment of (1), m2(Bj) Denotes that the focal length is BjAssigning a basic probability;
step c 2: the fusion result is obtained by applying Dempster combination rule and combining the step c1
The Dempster combination rule is:
Figure BDA0002402054190000075
substituting the results m (A) and m (B) obtained in the step c1 into m (AB);
where M is a normalization constant, and M ═ SigmaA∩B＝φ(m(A)m(B))＝1-∑A∩B≠φ(m(A)m(B))
m (a) represents the basic probability distribution function of subset a, m (B) represents the basic probability distribution function of subset B, and m (ab) represents the fused basic probability distribution function of subset a and subset B.
The image retrieval method based on the combination of the vocabulary tree information fusion and the Hausdorff distance comprises the following specific steps of the step d:
step d 1: form of differential equation for writing cost function
The differential equation form of the cost function is as follows:
Figure BDA0002402054190000081
step d 2: obtaining a general solution to a cost function
Solving the differential equation to obtain the cost function with the following expression:
Figure BDA0002402054190000082
wherein gamma is0The cost function is an initial value of the cost function, the range of the cost function is 0-1, k is a proportionality coefficient, and tau is a matching parameter;
step d 3: improved Hausdorff distance using traditional Hausdorff distance as variable of cost function
Given two finite sets X ═ X1,x2,...,xMY ═ Y1,y2,...,yNThe conventional Hausdorff distance between X and Y is defined as
Figure BDA0002402054190000083
Where d (X, Y) is the conventional Hausdorff distance, min represents the minimum, max represents the maximum, X and Y are the points in the point sets X and Y, respectively, and d (X, Y) represents the geometric distance between point X and point Y;
the improved Hausdorff distance is:
Figure BDA0002402054190000084
where | X | is the number of the finite set X, dH(X, Y) is the modified Hausdorff distance, d (X, Y) is the conventional Hausdorff distance, and γ (d (X, Y)) is a cost function with a variable d (X, Y).
The image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance comprises the following specific steps of step e:
and c, according to the fusion characteristics obtained in the step c, carrying out image similarity measurement by using an improved Hausdorff distance, and arranging the obtained similarities according to a descending order to obtain a retrieval result.
Detailed description of the invention
The present embodiment is a theoretical embodiment of an image retrieval method based on the combination of lexical tree information fusion and hausdorff distance.
In view of the fact that most of the technicians in the field are scholars, the writing of technical documents is more customary to the writing of articles, and therefore, on the basis of no essential difference from the specific embodiment, the second specific embodiment is supplemented according to the scholars.
The image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance in the embodiment comprises the following steps:
step a: SIFT feature extraction (SIFT: scale invariant feature transform) for image to be retrieved and image library
Step a 1: constructing to-be-retrieved image and image library Gaussian difference scale function
During the extraction of SIFT descriptors, firstly constructing a Gaussian difference scale space, wherein the scale space of a two-dimensional image is
Figure BDA0002402054190000091
Where G (x, y, σ) is a gaussian function with variable scale, (x, y) is the scale coordinate, I (x, y) is the image, L (x, y, σ) is the scale space of the two-dimensional image, and the size of σ determines the degree of smoothing of the image.
For more accurate detection of image feature points, it is necessary to construct a gaussian difference scale function of a two-dimensional image, which is generated by convolving the image with gaussian functions of different scales, i.e., D (x, y, σ) — (G (x, y, k σ) -G (x, y, σ)) × I (x, y) — L (x, y, k σ) -L (x, y, σ), where D (x, y, σ) is the gaussian difference scale function of the two-dimensional image and k is a scale scaling factor
Step a 2: detecting extreme point in Gaussian difference scale space
To find the extreme point in the scale space requires that each sample point in the image is compared with its neighboring points, and when a sample point is the maximum or minimum among all points in the DoG (difference of gaussians) space, the point is considered as a feature point of the image under the scale.
Step a 3: removing the feature points with unstable edges to generate SIFT descriptors
In order to enhance the matching stable point and improve the noise capability, a Harris Commer detector is used for removing unstable characteristic points of the edge. And keeping stable feature points to generate SIFT descriptors.
Step b: generating SIFT descriptor histogram and SIFT descriptor kernel density
Step b 1: construction of an extensible lexical tree SVT by hierarchical clustering of SIFT descriptors
Extracting SIFT descriptors of each picture to obtain a set F ═ FiAnd then, performing hierarchical clustering on the set F by adopting a K-Means clustering method. Initially, K-Means clustering is performed on the set F at layer 1, and the set F is divided into K parts of { FiI is more than or equal to 1 and less than or equal to k. Similarly, the newly generated cluster is subdivided into K clusters by using K-Means, and the above operations are repeated until the depth reaches the preset L value, so that the cluster is not split, and an expandable vocabulary tree is constructed, wherein c is equal to B in totalLAnd each node is formed. Where B is the branching factor, L is the depth, c is the total number of nodes, fiRepresenting a certain SIFT descriptor in a picture, F being a set of descriptors, FiIs a certain cluster set obtained by performing K-Means clustering on the set F.
Step b 2: accumulating the times of the occurrence of the descriptors on each node in the extensible vocabulary tree to obtain an SIFT descriptor histogram
In constructing an expandable lexical tree, c is shared as BLEach node accumulates the occurrence frequency of SIFT descriptors on the first node to obtain H [ H ] for SIFT descriptor histogram based on the expandable vocabulary tree1,...,hi,...,hc]Is represented by the formula (I) in which hiThe number of times of SIFT descriptors appearing at the ith node is represented, B is a branching factor, L is depth, and c is the total number of nodes.
Step b 3: quantizing the SIFT descriptor to obtain SIFT descriptor kernel density
All the SIFT descriptors are quantized,each SIFT descriptor fiAll corresponding to a quantization path from a root node to a leaf node in the expandable vocabulary tree, i.e. corresponding to a group of visual words
Figure BDA0002402054190000101
Each group of visual words corresponds to the kernel density f (c), and the SIFT descriptor kernel density based on the extensible vocabulary tree is obtained. Wherein
Figure BDA0002402054190000102
Is a visual word, i.e. each node in the expandable vocabulary tree represents a visual word, l represents the number of layers of the node in the expandable vocabulary tree, hlIndicating the index of the node in the level tree node, L being the depth.
Step c: fusing SIFT descriptor kernel density and SIFT descriptor histogram
Step c 1: obtaining a SIFT descriptor histogram and a basic probability distribution function of SIFT descriptor kernel density
For the following computational convenience, the frame Ω is identified by setting the SIFT descriptor histogram to a and the SIFT descriptor kernel density to B: { A, B }, the discrimination box is a set of all elements describing the overall hypothetical space. All possible outcomes are considered with the basic probability distribution function (BPA), often denoted m ().
The basic probability distribution function of subset A is
Figure BDA0002402054190000103
The basic probability distribution function of the subset B is
Figure BDA0002402054190000104
Wherein, M is a normalization constant,
Figure BDA0002402054190000105
m1(Ai) Denotes that the focal length is AiBasic probability assignment of (1), m2(Bj) Denotes that the focal length is BjAssigning a basic probability;
step c 2: the fusion result is obtained by applying Dempster combination rule and combining the step c1
The Dempster combination rule is:
Figure BDA0002402054190000106
substituting the results m (A) and m (B) obtained in step c1 into m (AB).
Where M is a normalization constant, and M ═ SigmaA∩B＝φ(m(A)m(B))＝1-∑A∩B≠φ(m(A)m(B))
m (a) represents the basic probability distribution function of subset a, m (B) represents the basic probability distribution function of subset B, and m (ab) represents the fused basic probability distribution function of subset a and subset B.
Step d: improving the conventional Hausdorff distance metric
In order to improve the reliability and stability of the matching process, the invention improves the traditional Hausdorff distance measurement, namely, the traditional Hausdorff distance is used as a variable of a cost function as the improved Hausdorff distance.
Step d 1: differential equation form with cost function written out
The differential equation form of the cost function is as follows:
Figure BDA0002402054190000111
step d 2: obtaining a general solution to a cost function
Solving the differential equation to obtain the cost function with the following expression:
Figure BDA0002402054190000112
wherein gamma is0The cost function is an initial value of the cost function, the range of the cost function is 0-1, k is a proportionality coefficient, and tau is a matching parameter.
Step d 3: improved Hausdorff distance using traditional Hausdorff distance as variable of cost function
Given two finite sets X ═ X1,x2,...,xMY ═ Y1,y2,...,yNThe conventional Hausdorff distance between X and Y is defined as
Figure BDA0002402054190000113
Where d (X, Y) is the conventional Hausdorff distance, min represents the minimum, max represents the maximum, X and Y are the points in the point sets X and Y, respectively, and d (X, Y) represents the geometric distance between point X and point Y
The improved Hausdorff distance is:
Figure BDA0002402054190000114
where | X | is the number of the finite set X, dH(X, Y) is the modified Hausdorff distance, d (X, Y) is the conventional Hausdorff distance, and γ (d (X, Y)) is a cost function with a variable d (X, Y)
Step e: using improved Hausdorff distance for image matching
And c, according to the fusion characteristics obtained in the step c, carrying out image similarity measurement on the characteristics by using an improved Hausdorff distance, and arranging the obtained similarities according to a descending order to obtain a retrieval result.
Detailed description of the preferred embodiment
The present embodiment is an experimental embodiment of an image retrieval method based on combination of lexical tree information fusion and hausdov distance.
Fig. 2 shows precision rates of image retrieval based on the SIFT descriptor histogram, image retrieval based on the SIFT descriptor kernel density, and image retrieval based on the present invention.
As can be seen from fig. 2, the first four cloud, star, bird and tree in the image category are simple pictures with backgrounds, and the precision ratios of the three retrieved images are not very different; the last four items in the image category are pictures with tiger, fish, mountain and flower as complex backgrounds, the precision ratio of the three retrieval methods is greatly different, and the retrieval of the method is far larger than that of the first two.
The experimental results for the two image types are given below
In the experiment, a small self-built image database is used, wherein the database contains 8 types of images, namely flowers, birds, fish, tigers, mountains, trees, stars and clouds, the total number of the images is 800 in total, and each type of image is 100.
Experiment one: background clearness experiment of image to be retrieved
The method comprises the steps of taking a banyan image with a simple background as an image to be retrieved, randomly extracting 5 images from all banyans as query images, and finally selecting an average value of the precision ratios of the 5 images as a final result. Precision is defined as follows: the precision ratio (number of images related to the key map in the query result/number of images returned by the query) is 100%.
A banyan image with a simple background is given as an image to be retrieved, as shown in fig. 3; the retrieval result of the method of the invention is shown in fig. 4, the retrieval result based on the SIFT descriptor histogram method is shown in fig. 5, and the retrieval result based on the SIFT descriptor kernel density method is shown in fig. 6.
As can be seen from the search results of fig. 4, 5, and 6: the background of the image to be retrieved is clear, the color information of the banyan is clear, the crown of the banyan is large, most of the image is covered, and rich textural feature information is formed; the shape information between the crown and the background of the image to be retrieved and at the trunk is clearer.
Each image to be retrieved returns 30 images, wherein the images accurately retrieved by the method of the invention are respectively 23, 25 and 25, the precision ratio is respectively 76.7%, 83.3% and 83.3%, and the average precision ratio (76.7+76.7+83.3+83.3+83.3)/5 is 100% and 80.66%; the images accurately searched by the SIFT descriptor histogram method are 23, 24, 25 and 25 respectively, the precision ratio is 76.7%, 80%, 83.3% and 83.3%, and the average precision ratio is (76.7+76.7+80+83.3+83.3)/5 x 100% is 80%; the images accurately searched by the SIFT descriptor density checking method are 23, 24, 25 and 25 respectively, the precision ratio is 76.7%, 80%, 83.3% and 83.3%, and the average precision ratio is (80+76.7+76.7+83.3+83.3)/5 × 100% and 80%;
for pictures with simple backgrounds, the difference between the retrieval method disclosed by the invention and the images retrieved by the histogram retrieval based on the SIFT descriptor and the kernel density based on the SIFT descriptor is not large, and the precision rate difference is not large and reaches about 80%.
Experiment two: background complex experiment of image to be retrieved
Taking the 'tiger' image with complex background as the image to be retrieved, randomly extracting 5 images from all the 'tiger' images as query images, and finally taking the average value of the precision ratios of the selected 5 images as the final result. The precision ratio is defined as follows: the precision ratio (number of images related to the key map in the query result/number of images returned by the query) is 100%.
A tiger image with a complex background is given as an image to be retrieved, as shown in fig. 7; the retrieval result of the method according to the invention is shown in fig. 8, the retrieval result based on the SIFT descriptor histogram method is shown in fig. 9, and the retrieval result based on the SIFT descriptor kernel density method is shown in fig. 10.
As can be seen from fig. 8, a total of 30 images were returned, of which 26 were retrieved with an accuracy of 86.7%. The first image of the retrieval result is the image to be retrieved, 25 retrieved images in the remaining 29 images are all the images of the tiger class, and the shape of the tiger head, the pattern of the tiger skin, the characteristics of the background area and the like in the 25 images are very similar to the image to be retrieved.
As can be seen from fig. 9, a total of 30 images were returned, of which 12 images were retrieved with an accuracy of 40%. As can be seen from fig. 10, a total of 30 images were returned, of which 13 images were accurately retrieved with an accuracy of 43.3%. The two search results show that although 12 and 13 searched images are also the images of the tiger type, the shape of the tiger head, the patterns of the tiger skin and the background area are greatly different from the image to be searched, but the characteristic background of the searched image is single.
Retrieving the other four images to be retrieved as the images of the tiger, wherein each image to be retrieved returns 30 images, the images accurately retrieved by the method are respectively 25, 26 and 27, the precision ratio is respectively 83.3%, 86.7% and 90.0%, and the average precision ratio is (86.7+83.3+83.3+86.7+90.0)/5 x 100% is 86.0%; the images accurately searched by the SIFT descriptor histogram method are respectively 12, 13 and 13, the precision ratio is respectively 40.0%, 43.3% and 43.3%, and the average precision ratio is (40.0+40.0+40.0+43.3+43.3)/5 is 100% and is 41.32%; the images accurately searched by the SIFT descriptor check density method are respectively 12, 13 and 13, the precision ratio is respectively 40.0%, 40.0% and 43.3%, and the average precision ratio is (43.3+40.0+40.0+43.3+43.3)/5 x 100% is 41.98%;
from the search results of experiment two, it can be found that the average precision ratio of the two search results which are not fused in the search of the pictures with complicated backgrounds only reaches 41.32% and 41.98%, which is equivalent to that the pictures with complicated backgrounds cannot be searched at all. The average precision ratio of the method reaches 86%, and the precision ratio is not reduced due to the complex background, so that the retrieval result fully proves that the image retrieval method combining the expandable vocabulary tree information fusion and the Hausdorff distance can make up the defect that the original retrieval method cannot retrieve the picture with the complex background.

Claims (1)

1. The fusion method for image retrieval is characterized by comprising fusion of SIFT descriptor kernel density and SIFT descriptor histogram and comprises the following steps:
step c 1: obtaining a basic probability distribution function of SIFT descriptor histogram and SIFT descriptor kernel density
For computational convenience, if the SIFT descriptor histogram is set as a, and the SIFT descriptor kernel density is set as B, then the box Ω: { A, B }, the decision box is a set describing all elements constituting the whole hypothesis space, and is represented by m () with all possible results considered by the basic probability distribution function; at this time, the process of the present invention,
the basic probability distribution function of subset A is
Figure FDA0003650845960000011
The basic probability distribution function of the subset B is
Figure FDA0003650845960000012
Wherein, M is a normalization constant,
Figure FDA0003650845960000013
m1(Ai) Denotes that the focal length is AiBasic probability assignment of (1), m2(Bj) Denotes that the focal length is BjAssigning a basic probability;
step c 2: the fusion result is obtained by applying Dempster combination rule and combining the step c1
The Dempster combination rule is:
Figure FDA0003650845960000014
substituting the results m (A) and m (B) obtained in the step c1 into m (AB);
where M is a normalization constant, and M ═ SigmaA∩B＝φ(m(A)m(B))＝1-∑A∩B≠φ(m(A)m(B))
m (a) represents the basic probability distribution function of subset a, m (B) represents the basic probability distribution function of subset B, and m (ab) represents the fused basic probability distribution function of subset a and subset B;
the fusion method facing the image retrieval is used for an image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance, and the image retrieval method based on combination of the vocabulary tree information fusion and the Hausdorff distance comprises the following steps:
step a, extracting an image to be retrieved and SIFT characteristics of an image library; the method comprises the following specific steps:
step a 1: constructing a Gaussian difference scale function of an image to be retrieved and an image library;
step a 2: detecting extreme points in a Gaussian difference scale space;
step a 3: removing feature points with unstable edges and generating SIFT descriptors;
b, generating an SIFT descriptor histogram and SIFT descriptor kernel density; the method comprises the following specific steps:
step b 1: constructing an extensible vocabulary tree through hierarchical clustering of SIFT descriptors;
step b 2: accumulating the occurrence times of the descriptors on each node in the extensible vocabulary tree to obtain an SIFT descriptor histogram;
step b 3: quantizing the SIFT descriptors to obtain SIFT descriptor kernel density;
step c, fusing SIFT descriptor kernel density and SIFT descriptor histogram; the method comprises the following specific steps:
step c 1: obtaining a SIFT descriptor histogram and a basic probability distribution function of SIFT descriptor kernel density;
step c 2: a fusion result is obtained by applying Dempster combination rule and combining the step c 1;
step d, improving the traditional Hausdorff distance measurement; the method comprises the following specific steps:
step d 1: writing a differential equation form of the cost function;
the differential equation form of the cost function is as follows:
Figure FDA0003650845960000021
step d 2: obtaining a general solution of the cost function;
solving the differential equation to obtain the cost function with the following expression:
Figure FDA0003650845960000022
wherein, γ0The cost function is an initial value of the cost function, the range of the cost function is 0-1, k is a proportionality coefficient, and tau is a matching parameter;
step d 3: the traditional Hausdorff distance is used as a variable of the cost function, and the Hausdorff distance is improved;
given two haveLimited set X ═ X1,x2,...,xMY ═ Y1,y2,...,yNThe conventional Hausdorff distance between X and Y is defined as
Figure FDA0003650845960000023
Where d (X, Y) is the conventional Hausdorff distance, min represents the minimum, max represents the maximum, X and Y are the points in the point sets X and Y, respectively, and d (X, Y) represents the geometric distance between point X and point Y;
the improved Hausdorff distance is:
Figure FDA0003650845960000024
where | X | is the number of the finite set X, dH(X, Y) is the modified Hausdorff distance, d (X, Y) is the conventional Hausdorff distance, and γ (d (X, Y)) is a cost function with a variable d (X, Y);
step e, using the improved Hausdorff distance for image matching; the method comprises the following specific steps:
and c, according to the fusion characteristics obtained in the step c, carrying out image similarity measurement by using an improved Hausdorff distance, and arranging the obtained similarities according to a descending order to obtain a retrieval result.
CN202010149889.5A 2017-02-13 2017-02-13 Fusion method for image retrieval Expired - Fee Related CN111309955B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010149889.5A CN111309955B (en) 2017-02-13 2017-02-13 Fusion method for image retrieval

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710076042.7A CN106844733B (en) 2017-02-13 2017-02-13 Image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance
CN202010149889.5A CN111309955B (en) 2017-02-13 2017-02-13 Fusion method for image retrieval

Related Parent Applications (1)

Application Number Title Priority Date Filing Date
CN201710076042.7A Division CN106844733B (en) 2017-02-13 2017-02-13 Image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance

Publications (2)

Publication Number Publication Date
CN111309955A CN111309955A (en) 2020-06-19
CN111309955B true CN111309955B (en) 2022-06-24

Family

ID=59128893

Family Applications (5)

Application Number Title Priority Date Filing Date
CN202010149899.9A Expired - Fee Related CN111368126B (en) 2017-02-13 2017-02-13 Image retrieval-oriented generation method
CN202010149894.6A Expired - Fee Related CN111309956B (en) 2017-02-13 2017-02-13 Image retrieval-oriented extraction method
CN202010149889.5A Expired - Fee Related CN111309955B (en) 2017-02-13 2017-02-13 Fusion method for image retrieval
CN201710076042.7A Expired - Fee Related CN106844733B (en) 2017-02-13 2017-02-13 Image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance
CN202010149888.0A Expired - Fee Related CN111368125B (en) 2017-02-13 2017-02-13 Distance measurement method for image retrieval

Family Applications Before (2)

Application Number Title Priority Date Filing Date
CN202010149899.9A Expired - Fee Related CN111368126B (en) 2017-02-13 2017-02-13 Image retrieval-oriented generation method
CN202010149894.6A Expired - Fee Related CN111309956B (en) 2017-02-13 2017-02-13 Image retrieval-oriented extraction method

Family Applications After (2)

Application Number Title Priority Date Filing Date
CN201710076042.7A Expired - Fee Related CN106844733B (en) 2017-02-13 2017-02-13 Image retrieval method based on combination of vocabulary tree information fusion and Hausdorff distance
CN202010149888.0A Expired - Fee Related CN111368125B (en) 2017-02-13 2017-02-13 Distance measurement method for image retrieval

Country Status (1)

Country Link
CN (5) CN111368126B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108009154B (en) * 2017-12-20 2021-01-05 哈尔滨理工大学 Image Chinese description method based on deep learning model
CN109978829B (en) * 2019-02-26 2021-09-28 深圳市华汉伟业科技有限公司 Detection method and system for object to be detected
CN111797268B (en) * 2020-07-17 2023-12-26 中国海洋大学 RGB-D image retrieval method
CN111931791B (en) * 2020-08-11 2022-10-11 重庆邮电大学 Method for realizing image turnover invariance

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020111A (en) * 2012-10-29 2013-04-03 苏州大学 Image retrieval method based on vocabulary tree level semantic model
CN103164856A (en) * 2013-03-07 2013-06-19 南京工业大学 Video copy and paste blind detection method based on dense scale-invariant feature transform stream
CN103489176A (en) * 2012-06-13 2014-01-01 中国科学院电子学研究所 Method for extracting TPs from SAR image of serious geometric distortion
CN104036524A (en) * 2014-06-18 2014-09-10 哈尔滨工程大学 Fast target tracking method with improved SIFT algorithm
CN104487916A (en) * 2012-07-26 2015-04-01 高通股份有限公司 Interactions of tangible and augmented reality objects
CN105183746A (en) * 2015-07-08 2015-12-23 西安交通大学 Method for realizing image retrieval by mining distinguishing features from multiple relevant pictures
CN106294577A (en) * 2016-07-27 2017-01-04 北京小米移动软件有限公司 Figure chip detection method and device

Family Cites Families (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5999653A (en) * 1996-01-19 1999-12-07 Xerox Corporation Fast techniques for searching images using the Hausdorff distance
EP1394727B1 (en) * 2002-08-30 2011-10-12 MVTec Software GmbH Hierarchical component based object recognition
US7912291B2 (en) * 2003-11-10 2011-03-22 Ricoh Co., Ltd Features for retrieval and similarity matching of documents from the JPEG 2000-compressed domain
US7542606B2 (en) * 2004-07-29 2009-06-02 Sony Corporation Use of Hausdorff distances in the earth mover linear program
US20080159622A1 (en) * 2006-12-08 2008-07-03 The Nexus Holdings Group, Llc Target object recognition in images and video
CN100550037C (en) * 2007-11-23 2009-10-14 重庆大学 Utilize and improve Hausdorff apart from the method for extracting the identification human ear characteristic
CN100592297C (en) * 2008-02-22 2010-02-24 南京大学 Multiple meaning digital picture search method based on representation conversion
CN101493891B (en) * 2009-02-27 2011-08-31 天津大学 Characteristic extracting and describing method with mirror plate overturning invariability based on SIFT
WO2011005865A2 (en) * 2009-07-07 2011-01-13 The Johns Hopkins University A system and method for automated disease assessment in capsule endoscopy
US8787682B2 (en) * 2011-03-22 2014-07-22 Nec Laboratories America, Inc. Fast image classification by vocabulary tree based image retrieval
US8811726B2 (en) * 2011-06-02 2014-08-19 Kriegman-Belhumeur Vision Technologies, Llc Method and system for localizing parts of an object in an image for computer vision applications
US20130046793A1 (en) * 2011-08-19 2013-02-21 Qualcomm Incorporated Fast matching of image features using multi-dimensional tree data structures
CN102542058B (en) * 2011-12-29 2013-04-03 天津大学 Hierarchical landmark identification method integrating global visual characteristics and local visual characteristics
CN102662955A (en) * 2012-03-05 2012-09-12 南京航空航天大学 Image retrieval method based on fractal image coding
US8768049B2 (en) * 2012-07-13 2014-07-01 Seiko Epson Corporation Small vein image recognition and authorization using constrained geometrical matching and weighted voting under generic tree model
US9177404B2 (en) * 2012-10-31 2015-11-03 Qualcomm Incorporated Systems and methods of merging multiple maps for computer vision based tracking
US8891908B2 (en) * 2012-11-14 2014-11-18 Nec Laboratories America, Inc. Semantic-aware co-indexing for near-duplicate image retrieval
CN102945289B (en) * 2012-11-30 2016-01-06 苏州搜客信息技术有限公司 Based on the image search method of CGCI-SIFT local feature
CN103336971B (en) * 2013-07-08 2016-08-10 浙江工商大学 Target matching method between multiple-camera based on multiple features fusion and incremental learning
CN103605765B (en) * 2013-11-26 2016-11-16 电子科技大学 A kind of based on the massive image retrieval system clustering compact feature
CN103729654A (en) * 2014-01-22 2014-04-16 青岛新比特电子科技有限公司 Image matching retrieval system on account of improving Scale Invariant Feature Transform (SIFT) algorithm
CN104008174B (en) * 2014-06-04 2017-06-06 北京工业大学 A kind of secret protection index generation method of massive image retrieval
CN104915949B (en) * 2015-04-08 2017-09-29 华中科技大学 A kind of image matching method of combination point feature and line feature
CN105022835B (en) * 2015-08-14 2018-01-12 武汉大学 A kind of intelligent perception big data public safety recognition methods and system
CN105138672B (en) * 2015-09-07 2018-08-21 北京工业大学 A kind of image search method of multiple features fusion
CN105550381B (en) * 2016-03-17 2019-04-05 北京工业大学 A kind of efficient image search method based on improvement SIFT feature
CN106339486A (en) * 2016-08-30 2017-01-18 西安电子科技大学 Image retrieval method based on incremental learning of large vocabulary tree

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103489176A (en) * 2012-06-13 2014-01-01 中国科学院电子学研究所 Method for extracting TPs from SAR image of serious geometric distortion
CN104487916A (en) * 2012-07-26 2015-04-01 高通股份有限公司 Interactions of tangible and augmented reality objects
CN103020111A (en) * 2012-10-29 2013-04-03 苏州大学 Image retrieval method based on vocabulary tree level semantic model
CN103164856A (en) * 2013-03-07 2013-06-19 南京工业大学 Video copy and paste blind detection method based on dense scale-invariant feature transform stream
CN104036524A (en) * 2014-06-18 2014-09-10 哈尔滨工程大学 Fast target tracking method with improved SIFT algorithm
CN105183746A (en) * 2015-07-08 2015-12-23 西安交通大学 Method for realizing image retrieval by mining distinguishing features from multiple relevant pictures
CN106294577A (en) * 2016-07-27 2017-01-04 北京小米移动软件有限公司 Figure chip detection method and device

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
Chandrika P 等.Multi modal semantic indexing for image retrieval.《Proceedings of the ACM International Conference on Image and Video Retrieval》.2010, *
Z. Wang 等.An Effective Web Image Searching Engine Based on SIFT Feature Matching.《2009 2nd International Congress on Image and Signal Processing》.2009, *
吴海滨 等.视频监控图像的运动模糊方向估计.《液晶与显示》.2014,第29卷(第4期), *
张乐锋.单一深度图像人体部位识别.《中国优秀硕士学位论文全文数据库 信息科技辑》.2016,(第2016（04）期), *

Also Published As

Publication number Publication date
CN111368125A (en) 2020-07-03
CN111368125B (en) 2022-06-10
CN111368126B (en) 2022-06-07
CN111309956B (en) 2022-06-24
CN106844733A (en) 2017-06-13
CN111368126A (en) 2020-07-03
CN111309955A (en) 2020-06-19
CN111309956A (en) 2020-06-19
CN106844733B (en) 2020-04-03

Similar Documents

Publication Publication Date Title
CN106682233B (en) Hash image retrieval method based on deep learning and local feature fusion
CN104850633B (en) A kind of three-dimensional model searching system and method based on the segmentation of cartographical sketching component
CN111309955B (en) Fusion method for image retrieval
Fakhari et al. Combination of classification and regression in decision tree for multi-labeling image annotation and retrieval
Niu et al. Knowledge-based topic model for unsupervised object discovery and localization
CN106874397B (en) Automatic semantic annotation method for Internet of things equipment
Wang et al. Image retrieval based on exponent moments descriptor and localized angular phase histogram
Kuric et al. ANNOR: Efficient image annotation based on combining local and global features
Pengcheng et al. Fast Chinese calligraphic character recognition with large-scale data
CN105740360B (en) Method for identifying and searching classical titles in artwork images
Ghosh et al. Efficient indexing for query by string text retrieval
Richter et al. Leveraging community metadata for multimodal image ranking
Pérez-Pimentel et al. A genetic algorithm applied to content-based image retrieval for natural scenes classification
Shi et al. Efficient Image Retrieval via Feature Fusion and Adaptive Weighting
Elhady et al. Weighted feature voting technique for content-based image retrieval
Janarthanam et al. Active Salient Component Classifier System on Local Features for Image Retrieval
Kumari et al. A Study and usage of Visual Features in Content Based Image Retrieval Systems.
Derakhshan et al. A Review of Methods of Instance-based Automatic Image Annotation
Amirshahi Presenting a method based on automatic image annotation techniques for the semantic recovery of images using artificial neural networks
EP4127965A1 (en) Computer-implemented method for analogue retrieval of documents
Nhi et al. A SEMANTIC-BASED IMAGE RETRIEVAL SYSTEM USING A HYBRID METHOD K-MEANS AND K-NEAREST-NEIGHBOR
DENIZIAK et al. IMPROVED QUERY BY APPROXIMATE SHAPES IMAGE RETRIEVAL METHOD.
Ghosh et al. A tutorial review of automatic image tagging technique using text mining
Ananthan et al. User Interactive Image Segmentation For Efficient Image Database Indexing
JP5444106B2 (en) Tag assignment device, conversion rule generation device, and tag assignment program

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20220624

CF01 Termination of patent right due to non-payment of annual fee