CN112507149A - Construction method of dynamic k neighbor graph and rapid image retrieval method based on dynamic k neighbor graph - Google Patents
Construction method of dynamic k neighbor graph and rapid image retrieval method based on dynamic k neighbor graph Download PDFInfo
- Publication number
- CN112507149A CN112507149A CN202011271641.2A CN202011271641A CN112507149A CN 112507149 A CN112507149 A CN 112507149A CN 202011271641 A CN202011271641 A CN 202011271641A CN 112507149 A CN112507149 A CN 112507149A
- Authority
- CN
- China
- Prior art keywords
- neighbor
- vertexes
- graph
- image
- updating
- 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.)
- Pending
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
- G06F16/51—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
Abstract
The invention relates to a construction method of a dynamic k-nearest neighbor graph and a rapid image retrieval method based on the dynamic k-nearest neighbor graph, which realize the online update of an approximate k-nearest neighbor graph and realize rapid image retrieval based on the dynamic k-nearest neighbor graph.
Description
Technical Field
The invention relates to the technical field of information retrieval, in particular to a dynamic k neighbor graph construction method and a rapid image retrieval method based on the dynamic k neighbor graph, which can be applied to scenes such as electronic commerce, search engines, installation monitoring and the like.
Background
A k-neighbor graph is a directed graph, S ═ x ∈ R for a given set of directional metricsdAnd each vertex of the directed graph represents a vector in the vector set, each vertex has k edges and points to the vertex (k neighbor) to which the k vectors nearest to the vertex belong, and the similarity is determined by the distance between the vectors, namely the Euclidean distance, the Hamming distance, the cosine distance and the like are commonly used.The k-nearest neighbor graph is an important data structure in the fields of manifold learning, computer vision, machine learning, multimedia information retrieval and the like. Based on the k-nearest neighbor graph, we can realize fast image retrieval, so we need to construct a high-quality k-nearest neighbor graph firstly.
If the k neighbor graph is constructed by adopting an exhaustive comparison mode, the time complexity is O (d.n)2) This is not feasible for large-scale data. At present, a generalized approximate k-nearest neighbor graph construction method is a nearest neighbor descent method. This method is based on the observation that "neighbors of neighbors may also be neighbors", obtaining a better k-neighbor map from a random k-neighbor map through gradual iterations.
The nearest neighbor descent method aims at constructing an approximate k-nearest neighbor graph for a fixed vector set. However, in an actual scenario, the data set sometimes changes, and for this case, a high-quality k-neighbor map cannot be constructed by using the nearest neighbor descent method.
In view of the above, the present inventors have made extensive studies to solve the problems occurring in the prior art, and as a result, have made the present invention.
Disclosure of Invention
The invention aims to provide a dynamic k neighbor graph construction method and a rapid image retrieval method based on the dynamic k neighbor graph, so as to realize the construction of a high-quality k neighbor graph for a dynamic image data set on line and realize efficient image retrieval based on the dynamic k neighbor graph.
In order to achieve the purpose, the invention adopts the technical scheme that:
a construction method of a dynamic k-nearest neighbor graph comprises the following steps:
step 2, when the number of the feature vectors in the image feature set S is less than M, calculating the distances between the feature vector q and all vectors in the image feature set S in an exhaustive comparison mode, and updating a k neighbor graph G according to the calculated distances;
and 3, when the number of the feature vectors in the image feature set S is not less than M, updating the k neighbor image G in the following mode:
step 3.1, randomly selecting a plurality of vertexes from the k neighbor graph G as starting points, and adding the vertexes into a priority queue C, wherein the vertex in the priority queue C which is closer to the characteristic vector q has higher priority;
step 3.2, taking out a head element C of the priority queue C, traversing neighbor vertexes of the vertex C in the k neighbor graph G, adding vertexes which are not visited before in the neighbor vertexes into C, and updating a k neighbor table of q;
step 3.3, repeating the step 3.2 until C is empty or the k neighbor table of q can not be updated any more;
step 3.4, updating the k neighbor graph G by using the distance information of all the vertex points and q accessed in the steps 3.2 and 3.3;
and 4, outputting the k neighbor graph.
The step 3.4 is as follows: and (3) optimizing the k neighbor graph by adopting a limited recursive neighbor propagation strategy, namely performing width-first search on all the vertexes accessed in the step 3.3 outwards, searching vertexes only expanded within a given layer number, requiring the vertexes to be within the k neighbor range of q, and updating the k neighbor graph G by utilizing the distance information calculated in the searching process.
A fast image retrieval method based on dynamic k-nearest neighbor images comprises the following steps: step 1, inputting a feature vector q of an image to be queried, and constructing a k neighbor graph of the feature vector q as an index of image retrieval;
step 1.1, when the number of feature vectors in the image feature set S is less than M, calculating the distances between the feature vector q and all vectors in the image feature set S in an exhaustive comparison mode, and updating a k neighbor graph G according to the calculated distances;
step 1.2, when the number of feature vectors in the image feature set S is not less than M, updating the k-nearest neighbor graph G according to the following mode:
step 1.21, randomly selecting a plurality of vertexes from a k-neighbor graph G as starting points, adding the vertexes into a priority queue C, wherein the vertex closer to a feature vector q in the priority queue C has higher priority, and then updating a k-neighbor table of the feature vector q, namely the k vertexes closest to the feature vector q;
step 1.22, taking out a head element C of the priority queue C, traversing neighbor vertexes of the vertex C in a k neighbor graph G, adding vertexes which are not visited before in the neighbor vertexes into C, and updating a k neighbor table of q;
step 1.23, repeat step 1.22 until C is empty or the k neighbor table of q can no longer be updated;
step 1.24, updating a k neighbor graph G by using the distance information between all the vertex points and q accessed in the step 1.23;
step 2, adopting a delay graph diversification strategy to improve the vector retrieval performance based on the k neighbor graph G, namely updating the occlusion values of the vertexes in the k neighbor table by using the distance information calculated in the step 1, marking the vertexes, of which the occlusion values exceed the average occlusion value of the k neighbor table, in the vertexes as inaccessible, and updating the k neighbor graph G; wherein, for k neighbor vertexes a and b of the vertex r, if the distance from r to b is greater than the distance from r to a, and the distance from a to b is less than the distance from r to b, then it is said that b is occluded by a in the neighbor of r, and the occlusion value of b in the k neighbor table of r is increased by one;
step 3, searching according to the k neighbor graph updated in the step 2;
step 3.1, randomly selecting a plurality of vertexes from the k neighbor graph G as starting points, and adding the vertexes into a priority queue C, wherein the vertex with the closer distance q in the priority queue C has higher priority;
step 3.2, taking out the head element C of the priority queue C, traversing the neighbor vertex of the head element C in the approximate k neighbor graph G, neglecting the vertex marked as 'inaccessible', adding the rest neighbor vertices which are not accessed into the queue C, and updating the k neighbor table of q;
step 3.3, repeating the step 3.2 until the queue C is empty or the k neighbor table of q can not be updated any more;
and 3.4, returning a k neighbor table of q, namely returning the serial numbers of the retrieved k image feature vectors most similar to q, wherein the corresponding image is the result required to be retrieved.
The step 1.24 is as follows: and (3) optimizing the K neighbor graph by adopting a limited recursive neighbor propagation strategy, namely performing width-first search on all the vertexes accessed in the step 1.23 outwards, searching vertexes only expanded within a given layer number, requiring the vertexes to be within the K neighbor range of q, and updating the K neighbor graph G by utilizing the distance information calculated in the searching process.
After the scheme is adopted, for any new image, the dynamic k-nearest neighbor image construction method firstly converts the image into a vector by a feature extraction and feature description method, and the process ensures that the distance between two vectors generated by two related images is as small as possible. And in the case of a small image data set, calculating the distance between the new vector and all the original vectors in the data set in an exhaustive comparison mode, and updating the k neighbor graph G. After the image data set has a certain scale, generating a corresponding vector q for a newly added image, starting from a plurality of random vertexes on the graph G, carrying out optimal priority search by taking q as a target, and updating the k neighbor graph G by using distance information calculated in the searching process, thereby realizing online updating of the approximate k neighbor graph and simultaneously supporting operations such as increasing, deleting, modifying and the like on the k neighbor graph. In addition, the invention also adopts a limited recursive neighbor propagation strategy to improve the quality of the k neighbor graph G.
In addition, in the image retrieval stage, the image to be retrieved is converted into a vector by the same method in the previous stage, the vector is utilized to improve the performance of the vector retrieval based on the k-nearest neighbor graph G based on the delay graph scattering strategy in the invention of the previous stage, the constructed dynamic approximate k-nearest neighbor graph carries out the optimal priority search, the graph vertex marked as inaccessible in the delay graph scattering stage is omitted in the searching process, and the retrieval efficiency is greatly improved. The dynamic k-nearest neighbor graph construction method realizes online updating of the approximate k-nearest neighbor graph, realizes rapid image retrieval based on the dynamic k-nearest neighbor graph, has good generalization due to the fact that distance measurement among vectors is not assumed, and shows the graph construction efficiency exceeding that of a nearest neighbor reduction method on most data sets while ensuring the quality of the k-nearest neighbor graph, and the retrieval efficiency on a plurality of image feature data sets is better than that of widely accepted methods such as HNSW.
Drawings
FIG. 1 is a comparison of the present invention with HNSW and recent degradation methods for retrieval performance on different image feature datasets.
Detailed Description
The invention discloses a method for constructing a dynamic k neighbor graph, which comprises the following steps:
And 2, when the scale of the image feature set S is small, namely the number of feature vectors in the image feature set S is smaller than M, calculating the distances between the vector q and all vectors in the image feature set S in an exhaustive comparison mode, and updating the k neighbor graph G according to the calculated distances. The distance here may be a euclidean distance, a hamming distance, a cosine distance, etc. between vectors.
And 3, when the image feature set S has a certain scale, namely the number of feature vectors in the image feature set S is not less than M, updating the k neighbor graph G according to the following mode:
step 3.1, randomly selecting a plurality of vertexes from the k-neighbor graph G as starting points, adding the vertexes into a priority queue C, wherein the priority of the vertexes in the priority queue C is higher as the vertexes are closer to a vector q (for convenience of description, the distance between a certain vertex and the q refers to the distance between the vector corresponding to the vertex and the q), and then updating a k-neighbor table of the vector q, namely the k vertexes closest to the q;
step 3.2, taking out a head element C of the priority queue C, traversing neighbor vertexes of the vertex C in the k neighbor graph G, adding vertexes which are not visited before in the neighbor vertexes into C, and updating a k neighbor table of q;
step 3.3, repeating the step 3.2 until C is empty or the k neighbor table of q can not be updated any more;
and 3.4, updating the k neighbor graph G by using the distance information of all the vertex points and q accessed in the step 3.3.
To improve the quality of the k-neighbor graph G, step 3.4 above may optimize the k-neighbor graph using a restricted recursive neighbor propagation strategy. That is, the breadth-first search is performed for all the vertices visited in step 3.3, the search only expands the vertices within a given number of layers, and requires these vertices to be within the K-nearest neighbor range of q, and the K-nearest neighbor graph G is updated by using the distance information calculated in the search process.
And 4, outputting the approximate k neighbor graph G.
The performance comparison of the above construction method with the nearest neighbor descent method is shown in the following table,
TABLE 1
As can be seen from table 1, the method of the present invention is superior in both the setup time and the scan rate compared to the existing nearest neighbor descent method.
Based on the same inventive concept, the invention also discloses a quick image retrieval method based on the dynamic k neighbor graph, which comprises the following steps:
when the characteristic vector q of the image to be inquired is input, an approximate k neighbor graph G is constructed by adopting the construction method of the dynamic k neighbor graph, quick retrieval is carried out based on the dynamic k neighbor graph G, the characteristic vector number closest to the q is returned, and the image corresponding to the number is the retrieval result. The method comprises the following specific steps:
step 1.1, when the number of feature vectors in the image feature set S is less than M, calculating the distances between the feature vector q and all vectors in the image feature set S in an exhaustive comparison mode, and updating a k neighbor graph G according to the calculated distances;
step 1.2, when the number of feature vectors in the image feature set S is not less than M, updating the k-nearest neighbor graph G according to the following mode:
step 1.21, randomly selecting a plurality of vertexes from a k-neighbor graph G as starting points, adding the vertexes into a priority queue C, wherein the vertex closer to a feature vector q in the priority queue C has higher priority, and then updating a k-neighbor table of the feature vector q, namely the k vertexes closest to the feature vector q;
step 1.22, taking out a head element C of the priority queue C, traversing neighbor vertexes of the vertex C in a k neighbor graph G, adding vertexes which are not visited before in the neighbor vertexes into C, and updating a k neighbor table of q;
step 1.23, repeat step 1.22 until C is empty or the k neighbor table of q can no longer be updated;
and step 1.24, updating the k neighbor graph G by using the distance information of all the vertex points and q accessed in the step 1.23.
Step 2, adopting a delay graph diversification strategy to improve the vector retrieval performance based on the k neighbor graph G, namely updating the occlusion values of the vertexes in the k neighbor table by using the distance information calculated in the step 1, marking the vertexes, of which the occlusion values exceed the average occlusion value of the k neighbor table, in the vertexes as inaccessible, and updating the k neighbor graph G; wherein, for k neighbor vertices a and b of vertex r, if the distance from r to b is greater than the distance from r to a, and the distance from a to b is less than the distance from r to b, it is said that b is occluded by a in the neighbor of r, and the occlusion value of b in the k neighbor table of r is increased by one.
Step 3, searching according to the k neighbor graph updated in the step 2;
step 3.1, randomly selecting a plurality of vertexes from the k neighbor graph G as starting points, and adding the vertexes into a priority queue C, wherein the vertex with the closer distance q in the priority queue C has higher priority;
step 3.2, taking out the head element C of the priority queue C, traversing the neighbor vertex of the head element C in the approximate k neighbor graph G, neglecting the vertex marked as 'inaccessible', adding the rest neighbor vertices which are not accessed into the queue C, and updating the k neighbor table of q;
step 3.3, repeating the step 3.2 until the queue C is empty or the k neighbor table of q can not be updated any more;
and 3.4, returning a k neighbor table of q, namely returning the serial numbers of the retrieved k image feature vectors most similar to q, wherein the corresponding image is the result required to be retrieved.
In order to improve the quality of the k-nearest neighbor graph G and further improve the retrieval accuracy, step 1.24 is: and (3) optimizing the K neighbor graph by adopting a limited recursive neighbor propagation strategy, namely performing width-first search on all the vertexes accessed in the step 1.23 outwards, searching vertexes only expanded within a given layer number, requiring the vertexes to be within the K neighbor range of q, and updating the K neighbor graph G by utilizing the distance information calculated in the searching process.
Fig. 1 is a performance comparison diagram of the present invention and a conventional retrieval method, and it can be seen from fig. 1 that the k-nearest neighbor graph construction time is less than that of the nearest neighbor reduction method under similar recall rate, and the retrieval performance of the present invention over HNSW is also present on four real-world image feature data sets (HNSW is considered to be the leading vector retrieval technology at present).
In summary, for any new image, the image is firstly converted into a vector through a method of feature extraction and feature description, and this process ensures that the distance between two vectors generated by two related images is as small as possible. And in the case of a small image data set, calculating the distance between the new vector and all the original vectors in the data set in an exhaustive comparison mode, and updating the k neighbor graph G. After the image data set has a certain scale, generating a corresponding vector q for a newly added image, starting from a plurality of random vertexes on the image G, carrying out optimal priority search by taking q as a target, and updating the k neighbor image G by using the distance information calculated in the searching process, thereby realizing the online update of the approximate k neighbor image. In addition, the invention also adopts a limited recursive neighbor propagation strategy to improve the quality of the k neighbor graph G.
In addition, in the image retrieval stage, the image to be retrieved is converted into a vector by the same method in the previous stage, the vector is utilized to improve the performance of the vector retrieval based on the k-nearest neighbor graph G based on the delay graph scattering strategy in the invention of the previous stage, the constructed dynamic approximate k-nearest neighbor graph carries out the optimal priority search, the graph vertex marked as inaccessible in the delay graph scattering stage is omitted in the searching process, and the retrieval efficiency is greatly improved. The dynamic k-nearest neighbor graph construction method realizes online updating of the approximate k-nearest neighbor graph, realizes rapid image retrieval based on the dynamic k-nearest neighbor graph, has good generalization due to the fact that distance measurement among vectors is not assumed, and shows the graph construction efficiency exceeding that of a nearest neighbor reduction method on most data sets while ensuring the quality of the k-nearest neighbor graph, and the retrieval efficiency on a plurality of image feature data sets is better than that of widely accepted methods such as HNSW.
The above description is only exemplary of the present invention and is not intended to limit the technical scope of the present invention, so that any minor modifications, equivalent changes and modifications made to the above exemplary embodiments according to the technical spirit of the present invention are within the technical scope of the present invention.
Claims (4)
1. A construction method of a dynamic k-nearest neighbor graph is characterized by comprising the following steps: the method comprises the following steps:
step 1, when an image is newly added into an image data set, extracting the characteristics of the newly added image to obtain a characteristic vector q, and adding the characteristic vector q into an image characteristic set S; the image feature set S is a vector set in a d-dimensional space, wherein vectors are obtained by an image in the image data set by a feature extraction method;
step 2, when the number of the feature vectors in the image feature set S is less than M, calculating the distances between the feature vector q and all vectors in the image feature set S in an exhaustive comparison mode, and updating a k neighbor graph G according to the calculated distances;
and 3, when the number of the feature vectors in the image feature set S is not less than M, updating the k neighbor image G in the following mode:
step 3.1, randomly selecting a plurality of vertexes from the k neighbor graph G as starting points, and adding the vertexes into a priority queue C, wherein the vertex in the priority queue C which is closer to the characteristic vector q has higher priority;
step 3.2, taking out a head element C of the priority queue C, traversing neighbor vertexes of the vertex C in the k neighbor graph G, adding vertexes which are not visited before in the neighbor vertexes into C, and updating a k neighbor table of q;
step 3.3, repeating the step 3.2 until C is empty or the k neighbor table of q can not be updated any more;
step 3.4, updating the k neighbor graph G by using the distance information of all the vertex points and q accessed in the steps 3.2 and 3.3;
and 4, outputting the k neighbor graph.
2. The method for constructing a dynamic k-nearest neighbor graph according to claim 1, wherein: the step 3.4 is as follows: and (3) optimizing the k neighbor graph by adopting a limited recursive neighbor propagation strategy, namely performing width-first search on all the vertexes accessed in the step 3.3 outwards, searching vertexes only expanded within a given layer number, requiring the vertexes to be within the k neighbor range of q, and updating the k neighbor graph G by utilizing the distance information calculated in the searching process.
3. A fast image retrieval method based on dynamic k neighbor graph is characterized in that: the method comprises the following steps:
step 1, inputting a feature vector q of an image to be queried, and constructing a k neighbor graph of the feature vector q as an index of image retrieval;
step 1.1, when the number of feature vectors in the image feature set S is less than M, calculating the distances between the feature vector q and all vectors in the image feature set S in an exhaustive comparison mode, and updating a k neighbor graph G according to the calculated distances;
step 1.2, when the number of feature vectors in the image feature set S is not less than M, updating the k-nearest neighbor graph G according to the following mode:
step 1.21, randomly selecting a plurality of vertexes from a k-neighbor graph G as starting points, adding the vertexes into a priority queue C, wherein the vertex closer to a feature vector q in the priority queue C has higher priority, and then updating a k-neighbor table of the feature vector q, namely the k vertexes closest to the feature vector q;
step 1.22, taking out a head element C of the priority queue C, traversing neighbor vertexes of the vertex C in a k neighbor graph G, adding vertexes which are not visited before in the neighbor vertexes into C, and updating a k neighbor table of q;
step 1.23, repeat step 1.22 until C is empty or the k neighbor table of q can no longer be updated;
step 1.24, updating a k neighbor graph G by using the distance information between all the vertex points and q accessed in the step 1.23;
step 2, adopting a delay graph diversification strategy to improve the vector retrieval performance based on the k neighbor graph G, namely updating the occlusion values of the vertexes in the k neighbor table by using the distance information calculated in the step 1, marking the vertexes, of which the occlusion values exceed the average occlusion value of the k neighbor table, in the vertexes as inaccessible, and updating the k neighbor graph G; wherein, for k neighbor vertexes a and b of the vertex r, if the distance from r to b is greater than the distance from r to a, and the distance from a to b is less than the distance from r to b, then it is said that b is occluded by a in the neighbor of r, and the occlusion value of b in the k neighbor table of r is increased by one;
step 3, searching according to the k neighbor graph updated in the step 2;
step 3.1, randomly selecting a plurality of vertexes from the k neighbor graph G as starting points, and adding the vertexes into a priority queue C, wherein the vertex with the closer distance q in the priority queue C has higher priority;
step 3.2, taking out the head element C of the priority queue C, traversing the neighbor vertex of the head element C in the approximate k neighbor graph G, neglecting the vertex marked as 'inaccessible', adding the rest neighbor vertices which are not accessed into the queue C, and updating the k neighbor table of q;
step 3.3, repeating the step 3.2 until the queue C is empty or the k neighbor table of q can not be updated any more;
and 3.4, returning a k neighbor table of q, namely returning the serial numbers of the retrieved k image feature vectors most similar to q, wherein the corresponding image is the result required to be retrieved.
4. The method according to claim 3, wherein the method comprises the following steps: the step 1.24 is as follows: and (3) optimizing the K neighbor graph by adopting a limited recursive neighbor propagation strategy, namely performing width-first search on all the vertexes accessed in the step 1.23 outwards, searching vertexes only expanded within a given layer number, requiring the vertexes to be within the K neighbor range of q, and updating the K neighbor graph G by utilizing the distance information calculated in the searching process.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011271641.2A CN112507149A (en) | 2020-11-13 | 2020-11-13 | Construction method of dynamic k neighbor graph and rapid image retrieval method based on dynamic k neighbor graph |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011271641.2A CN112507149A (en) | 2020-11-13 | 2020-11-13 | Construction method of dynamic k neighbor graph and rapid image retrieval method based on dynamic k neighbor graph |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112507149A true CN112507149A (en) | 2021-03-16 |
Family
ID=74957772
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011271641.2A Pending CN112507149A (en) | 2020-11-13 | 2020-11-13 | Construction method of dynamic k neighbor graph and rapid image retrieval method based on dynamic k neighbor graph |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112507149A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114840721A (en) * | 2022-07-01 | 2022-08-02 | 北京文景松科技有限公司 | Data searching method and device and electronic equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130230255A1 (en) * | 2012-03-02 | 2013-09-05 | Microsoft Corporation | Image Searching By Approximate k-NN Graph |
| CN103377237A (en) * | 2012-04-27 | 2013-10-30 | 常州市图佳网络科技有限公司 | High dimensional data neighbor search method and fast approximate image search method |
| CN103473248A (en) * | 2012-12-17 | 2013-12-25 | 东北大学 | Continuously-moving K adjacent query method in road network environment |
| CN105912611A (en) * | 2016-04-05 | 2016-08-31 | 中国科学技术大学 | CNN based quick image search method |
| CN110070121A (en) * | 2019-04-15 | 2019-07-30 | 西北工业大学 | A kind of quick approximate k nearest neighbor method based on tree strategy with balance K mean cluster |
-
2020
- 2020-11-13 CN CN202011271641.2A patent/CN112507149A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130230255A1 (en) * | 2012-03-02 | 2013-09-05 | Microsoft Corporation | Image Searching By Approximate k-NN Graph |
| CN103377237A (en) * | 2012-04-27 | 2013-10-30 | 常州市图佳网络科技有限公司 | High dimensional data neighbor search method and fast approximate image search method |
| CN103473248A (en) * | 2012-12-17 | 2013-12-25 | 东北大学 | Continuously-moving K adjacent query method in road network environment |
| CN105912611A (en) * | 2016-04-05 | 2016-08-31 | 中国科学技术大学 | CNN based quick image search method |
| CN110070121A (en) * | 2019-04-15 | 2019-07-30 | 西北工业大学 | A kind of quick approximate k nearest neighbor method based on tree strategy with balance K mean cluster |
Non-Patent Citations (1)
| Title |
|---|
| WAN-LEI ZHAO*等: "Approximatek-NN Graph Construction:a Generic Online Approach", 《HTTPS://ARXIV.ORG/PDF/1804.03032.PDF》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114840721A (en) * | 2022-07-01 | 2022-08-02 | 北京文景松科技有限公司 | Data searching method and device and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111522986B (en) | Image retrieval method, device, equipment and medium | |
| CN102422319B (en) | Image retrieval method and image registration method | |
| CN109685121B (en) | Training method of image retrieval model, image retrieval method and computer equipment | |
| CN107341178B (en) | Data retrieval method based on self-adaptive binary quantization Hash coding | |
| US20120084305A1 (en) | Compiling method, compiling apparatus, and compiling program of image database used for object recognition | |
| CN103744886B (en) | Directly extracted k nearest neighbor searching algorithm | |
| CN109871461A (en) | The large-scale image sub-block search method to be reordered based on depth Hash network and sub-block | |
| CN109829066B (en) | Local sensitive Hash image indexing method based on hierarchical structure | |
| CN111859004A (en) | Retrieval image acquisition method, device, equipment and readable storage medium | |
| CN109472282B (en) | Depth image hashing method based on few training samples | |
| CN111078952B (en) | Cross-modal variable-length hash retrieval method based on hierarchical structure | |
| CN112507149A (en) | Construction method of dynamic k neighbor graph and rapid image retrieval method based on dynamic k neighbor graph | |
| CN112434031A (en) | Uncertain high-utility mode mining method based on information entropy | |
| JP6705764B2 (en) | Generation device, generation method, and generation program | |
| JP4958228B2 (en) | Registration method of document and / or image in database and search method thereof | |
| JP4440246B2 (en) | Spatial index method | |
| JP5061147B2 (en) | Image search device | |
| CN111859192B (en) | Searching method, searching device, electronic equipment and storage medium | |
| CN115129915A (en) | Repeated image retrieval method, device, equipment and storage medium | |
| CN112685603A (en) | Efficient retrieval of top-level similarity representations | |
| JP6333306B2 (en) | SEARCH DATA MANAGEMENT DEVICE, SEARCH DATA MANAGEMENT METHOD, AND SEARCH DATA MANAGEMENT PROGRAM | |
| CN113761243A (en) | Online retrieval method and system | |
| Lin et al. | The mesh with hybrid buses: an efficient parallel architecture for digital geometry | |
| CN104615614B (en) | The method for obtaining scalable global characteristics description | |
| WO2021115154A1 (en) | Portable device positioning data processing method and apparatus, device, and storage medium |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20210316 |