CN107527058A - A kind of image search method based on weighting local feature Aggregation Descriptor - Google Patents
A kind of image search method based on weighting local feature Aggregation Descriptor Download PDFInfo
- Publication number
- CN107527058A CN107527058A CN201710609564.9A CN201710609564A CN107527058A CN 107527058 A CN107527058 A CN 107527058A CN 201710609564 A CN201710609564 A CN 201710609564A CN 107527058 A CN107527058 A CN 107527058A
- Authority
- CN
- China
- Prior art keywords
- local feature
- class
- image
- characteristic point
- aggregation descriptor
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
- G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Library & Information Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of image search method based on weighting local feature Aggregation Descriptor, belong to Computer Image Processing, mode identification technology.This method is in existing local feature Aggregation Descriptor (vector of locally aggregated descriptor, VLAD the method that two-stage polymerization) is used on the basis of method, add up respectively and be assigned to the residual error of characteristic point around the characteristic point and the class heart of each class heart, while the distribution situation of statistical nature point and assign different weights to characteristic point in cumulative.Contrast prior art, present invention improves original local feature to polymerize the problem of feature space division is coarse in description method, and improve the weight of the characteristic point with taste, the vector of this method generation is consistent with original method generation vector length, and retrieval precision is improved while memory space is not increased.
Description
Technical field
The present invention relates to a kind of image search method based on weighting local feature Aggregation Descriptor, belong to computer picture
Processing, mode identification technology.
Background technology
Vision is one of most basic maximally effective approach in the human cognitive world, and image is a kind of common visual information
Carrier.With the development of Internet technology, the network bandwidth that user can use is increasing, while with the application of social class
Rise, the application based on multimedia especially image occurs more and more.And among these applications, image retrieval technologies are
A maximum application of basis and usage amount the most.In conventional search techniques, search engine returns to phase according to input entry
Link is closed, and major search engine manufacturer (such as Google, Baidu) is all proposed the search skill using image as input in recent years
Art, user is facilitated to retrieve image interested in internet.
In field of image search, vision bag of words (bag of visual word, BOVW) be the most frequently used model it
One.Vision bag of words are clustered the characteristics of image of extraction, and each class heart is referred to as a vision word, all by what is obtained
Vision word collects to obtain vision code book.Each image is encoded according to code book again, obtains a high dimension vector.So,
The distance just image retrieval being converted between vector is calculated, and candidate vector is carried out into ascending order row according to the distance with query vector
Sequence, corresponding image are Query Result.
But vision bag of words need larger vision code book, memory space requirements are larger, retrieved in large-scale image
Middle practicality is poor.Local feature Aggregation Descriptor (Vector of locally aggregated descriptor, VLAD)
It is an important improvement for being directed to vision bag of words.Each image is equally expressed as a vector by the model, but using compared with
Small code book is encoded, and is greatly reduced memory space and is accelerated calculating speed.To each image, local feature polymerization description
Symbol method quantifies to the characteristic point of the width image first, that is, calculates the class heart nearest apart from each characteristic point, and by feature
Point quantifies into corresponding classification, calculate afterwards each classification be assigned to such all characteristic points vectors and the class heart difference
With：
Wherein, μiFor the vector of the class heart of i-th of class, NN () is nearest neighbor algorithm function, xjIt is characterized point j vector.
Afterwards, the local feature Aggregation Descriptor of the image is expressed as v=[v1, v2..., vK], wherein K is the number of the class heart, also referred to as
For the number of vision word.Afterwards, the retrieval between image is converted into the distance between vector to calculate.
But because the code book that local feature Aggregation Descriptor method uses is smaller, cause in this method feature when quantifying
Error is larger, and loss information is more, and feature representation is more coarse, and all characteristic points weight phase in coding in the model
Together, some statistical informations be lost.
The content of the invention
The shortcomings that the invention aims to overcome in the prior art, mainly for local feature Aggregation Descriptor model
The problem of retrieval result accuracy rate caused by characteristic quantification is coarse in retrieval and low recall rate, to local characteristic aggregation descriptor
Model is improved, and proposes a kind of image search method based on weighting local feature Aggregation Descriptor, can using this method
Realize that large-scale image is retrieved, there is the advantages that retrieval rate and recall rate are high, and the speed of service is fast and committed memory is few, tool
There is very strong robustness.
The present invention seeks to what is be achieved through the following technical solutions.
A kind of image search method based on weighting local feature Aggregation Descriptor, comprises the following steps：
The local feature region of all images in step 1, extraction image data base；
Step 2, using clustering algorithm all local feature regions extracted in step 1 are clustered, generate K class, often
Individual class heart μi(1≤i≤K) is the average that counts of the characteristic vector of such all characteristic point, and code book is [μ1, μ2..., μK],
Codebook vectors dimension is K × D, and wherein D is characterized vector dimension a little；
Step 3, for every image in database, calculate its and weight local feature Aggregation Descriptor, concretely comprise the following steps：
Step 3.1, using caused code book in step 2, for each characteristic point in present image, calculated using arest neighbors
Method NN () finds out the class heart μ nearest with iti, and its first category label is set to i；
Step 3.2, for i-th of class, it is assumed that MiTo belong to such feature point number in present image, for each class
Heart μi, find out the γ × M closest with itiIndividual characteristic point, and will be added in the second category tag set of corresponding characteristic point
Label i, wherein γ are fixed value；
Difference between the corresponding class Heart vector of step 3.3, the characteristic vector of each characteristic point is referred to as residual error, calculates step
The weighted residual for the characteristic point being assigned in rapid 3.1 in each class and, for i-th of class, belong to such for first category label
The residual errors of all characteristic points summed：
Wherein, characteristic point j characteristic vector is expressed as xj, NN () is nearest neighbor algorithm, NN (xj) represent distance xjRecently
The class heart, wjFor xjWeight, using probability against word frequency (probabilistic inverse document frequency,
IDFP) it is calculated as follows：
Wherein, K be code book in the class heart number, njFor xjFirst category label and second category tag set union
In element number, i.e. xjIt is quantized in step 3.1 in a class, if being quantized in step 3.2 in Ganlei, njFor two
The different class number sums being quantized in step；
The weighted residual for the characteristic point being assigned in step 3.4, calculation procedure 3.2 in each class and, for i-th of class,
The residual error of all characteristic points to including such label in second category tag set is summed：
Wherein, characteristic point t characteristic vector is expressed as xt, qN (μi) it is expressed as distance-like heart μiQ nearest characteristic point,
For class heart μi, q value is γ × Mi, wtFor xtWeight, wtEqually calculated using probability against word frequency (IDFP)：
Wherein, K be code book in the class heart number, ntFor xtFirst category label and second category tag set union
Element number.
Step 3.5, present image weighting local feature Aggregation Descriptor are v=[v1, v2..., vK], wherein viRepresent such as
Under：
vi=Si+ri
V vector dimensions are tieed up for K × D, and wherein D is characterized the dimension of vector；
Step 3.6, the weighting local feature Aggregation Descriptor obtained to step 3.5 are normalized, and are finally represented；
Step 4, for image to be checked, extract image local feature point according to step 1, image calculated according to step 3
Weight local feature Aggregation Descriptor；
Step 5, the weighting local feature Aggregation Descriptor for calculating image to be checked and the weighting office of all images of database
The distance of portion's characteristic aggregation descriptor, and ascending sort, corresponding image are retrieval result.
Beneficial effect
The inventive method, using the code book generated offline, using the method for two-stage polymerization, add up be assigned to each class respectively
The residual error of characteristic point around the characteristic point and the class heart of the heart, at the same the distribution situation of statistical nature point and in cumulative to characteristic point
Different weights are assigned, improve the problem of feature space division is coarse in original local feature polymerization description method, and improve
There is the weight of the characteristic point of taste, it is consistent that vector and the original method of this method generation generate vector length, is not increasing
Retrieval precision is improved while adding memory space.
Brief description of the drawings
Fig. 1 is the inventive method schematic flow sheet；
Fig. 2 is step 3.1 according to nearest neighbor algorithm, and characteristic point is distributed to the schematic diagram of the class heart of its nearest neighbours；
Fig. 3 is that step 3.2 calculates distance-like heart μiNearest γ × MiThe schematic diagram of individual characteristic point.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention,
It is not intended to limit the present invention.In addition, technical characteristic involved in each embodiment of invention described below is only
Conflict can is not formed each other to be mutually combined.
As shown in figure 1, a kind of image search method based on weighting local feature Aggregation Descriptor, comprises the following steps：
According to step 1, all image local feature points of image data base are extracted；
Specifically, SIFT, the character representation characteristic point such as SURF, RootSIFT can be used, it is preferable that the present invention uses
RootSIFT features are as local feature region.RootSIFT features can significantly improve precision and without obvious compared to other methods
Amount of calculation increase；
According to step 2, all local feature regions extracted in step 1 are clustered using clustering algorithm, generation K
Class, each class heart μi(1≤i≤K) is the average that counts of the characteristic vector of such all characteristic point, and each class heart is referred to as one and regarded
Feel word, code book can be expressed as [μ1, μ2..., μK], codebook vectors dimension is K × D, and wherein D is characterized dimension；
Specifically, the clustering methods such as K-Means, K-Means++, Hierarchical K-Means can be used to be gathered
Class, the present invention are not especially limited to this.Preferably, the present invention is clustered using K-Means++ methods, this method solve
The class heart initializes tender subject.Typically in retrieval, the more retrieval effectivenesses of vision word number are better in code book, but calculate the time
Also corresponding elongated, to balance retrieval precision and calculating the relation of time, codebook size K values are preferably 64,128,256 in the present invention
Or 512, RootSIFT are characterized as 128 dimensional vectors, therefore, codebook vectors length is 8192,16384,32768 or 65536
Dimension；
According to step 3, for every image in database, calculate it and weight local feature Aggregation Descriptor, specific steps
It is as follows：
According to step 3.1, as shown in Fig. 2 for each characteristic point in present image, using nearest neighbor algorithm find out with
Its nearest class heart μi, and its first category label is set to i.Preferably, distance calculates and uses Euclidean distance, each characteristic point
First category label there was only one；
According to step 3.2, as shown in figure 3, for i-th of class, it is assumed that MiTo belong to such characteristic point in present image
Number, for each class heart μi, find out the γ × M closest with itiIndividual characteristic point, and by the second class of corresponding characteristic point
Label i is added in distinguishing label set.
Wherein, γ values scope is [0,1], if value is 0, this algorithm deteriorates to original local feature Aggregation Descriptor
Algorithm.Preferably, γ is arranged to 0.5 in UKBench and Holidays data sets in the present invention.Preferably, distance calculates
Using Euclidean distance.As shown in Fig. 2 according to this step, each characteristic point be probably multiple class hearts nearest some characteristic points it
One, therefore each characteristic point may belong to multiple class hearts when this step calculates, i.e. second category label is not unique, therefore each
The second category label of characteristic point is stored using set.Because the code book used in local feature Aggregation Descriptor method is smaller,
And characteristic point quantify when use nearest neighbor algorithm (such as step 3.1) belongs to " Hard clustering " method, and quantization error is larger, be
Lower quantization error in the case of codebook size is not increased, this step is found out several closest characteristic points of each class heart and used
To correct the division to feature space；
According to step 3.3, the weighted residual for the characteristic point being assigned in calculation procedure 3.1 in each class and, for i-th
Class, summed to belong to the residual error of such all characteristic points to first category label：
Wherein, characteristic point j characteristic vector is expressed as xj, NN () is nearest neighbor algorithm function, NN (xj) represent distance xj
The nearest class heart.wjFor xjWeight, weight calculation can use inverse word frequency (inverse document frequency, IDF)
And its related variants are calculated, it is preferable that present invention introduces IDF important improvement method --- probability is against word frequency
(probabilistic inverse document frequency, IDFP), probability is as follows against word frequency calculation formula：
Wherein, K be code book in the class heart number, njIt is characterized point j first category label and second category tag set
And concentration element number, i.e. xjIt is quantized in step 3.1 in a class, if being quantized in step 3.2 in Ganlei,
njFor the different class number sums being quantized in two steps.IDFP is for njIt is more thanCharacteristic point weights be negative, can be with
Regard penalty term as, and in IDF strategies, weights are positive number；
According to step 3.4, the weighted residual for the characteristic point being assigned in calculation procedure 3.2 in each class and, for i-th
Class, the residual error of all characteristic points to including such label in second category tag set are summed：
Wherein, characteristic point t characteristic vector is expressed as xt, qN (μi) it is expressed as distance-like heart μiQ nearest characteristic point,
For class heart μi, q value is γ × Mi, wtFor xtWeight, wtEqually calculated using probability against word frequency (IDFP)：
Wherein, K be code book in the class heart number, ntFor xtFirst category label and second category tag set union
Element number.
According to step 3.5, present image weighting local feature Aggregation Descriptor v=[v1, v2..., vK], wherein viRepresent
It is as follows：
vi=Si+ri
The vector dimension is K × D, the dimension that wherein D is characterized, the vector length that this method obtains and original local feature
The vector length of Aggregation Descriptor is consistent；
Step 3.6, weighting local feature Aggregation Descriptor is normalized, finally represented, i.e. each image table
It is shown as a vector.Normalization process can be normalized using a variety of norms, it is preferable that the present invention uses power-law
Norm methods are normalized；It is also possible to principal component analysis (PCA) method is used to weighting local feature Aggregation Descriptor
Dimensionality reduction is carried out, obtains the compact form of descriptor, memory space is further reduced and takes, is to balance retrieval precision with calculating the time
Relation, the vector length of the compact form after dimensionality reduction is traditionally arranged to be 128 dimensions；
Step 4, for image to be checked, extract image local feature point according to step 1, query image calculated according to step 3
Weighting local feature Aggregation Descriptor；
Step 5, the weighting for the weighting local feature Aggregation Descriptor and all images of database for calculating query image are local
The distance of characteristic aggregation descriptor, and ascending sort, corresponding image are retrieval result.Preferably, distance calculates and uses Euclidean
Distance.
Claims (9)
1. a kind of image search method based on weighting local feature Aggregation Descriptor, it is characterised in that comprise the following steps：
The local feature region of all images in step 1, extraction image data base；
Step 2, using clustering algorithm all local feature regions extracted in step 1 are clustered, generate K class, each class
Center μi, 1≤i≤K, for the average that counts of the characteristic vector of such all characteristic point, code book can be expressed as [μ1, μ2...,
μK], codebook vectors dimension is K × D, and wherein D is characterized vector dimension a little；
Step 3, for every image in database, calculate its and weight local feature Aggregation Descriptor；
Step 4, for image to be checked, extract image local feature point according to step 1, the weighting of image calculated according to step 3
Local feature Aggregation Descriptor；
Step 5, the weighting of the weighting local feature Aggregation Descriptor for calculating image to be checked and all images of database are local special
The distance of Aggregation Descriptor is levied, and ascending sort, corresponding image are retrieval result.
2. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 1, its feature
It is, the weighting local feature Aggregation Descriptor is obtained by procedure below：
Step 3.1, using caused code book in step 2, for each characteristic point in present image, looked for using nearest neighbor algorithm
Go out the class heart μ nearest with iti, and its first category label is set to i；
Step 3.2, for i-th of class, make MiTo belong to such feature point number in present image, for each class heart μi, look for
Go out the γ × M closest with itiIndividual characteristic point, and label i will be added in the second category tag set of corresponding characteristic point,
Wherein γ is fixed numbers；
Difference between the corresponding class Heart vector of step 3.3, the characteristic vector of each characteristic point is referred to as residual error, calculation procedure
The weighted residual for the characteristic point being assigned in 3.1 in each class and, for i-th of class, belong to such for first category label
The residual error of all characteristic points is summed：
Wherein, characteristic point j characteristic vector is expressed as xj, NN () is nearest neighbor algorithm, NN (xj) represent distance xjNearest class
The heart, wjFor xjWeight；
The weighted residual for the characteristic point being assigned in step 3.4, calculation procedure 3.2 in each class and, for i-th of class, to
The residual error of all characteristic points comprising such label is summed in two class label set：
Wherein, characteristic point t characteristic vector is expressed as xt, qN (μi) it is expressed as distance-like heart μiQ nearest characteristic point, for
Class heart μi, q value is γ × Mi, wtFor xtWeight；
Step 3.5, present image weighting local feature Aggregation Descriptor v=[v1, v2..., vK], wherein viRepresent as follows：
vi=si+ri
V vector dimensions are K × D, and wherein D is characterized the dimension of vector；
Step 3.6, the weighting local feature Aggregation Descriptor obtained to step 3.5 are normalized, and are finally represented.
3. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 2, its feature
It is, the weight WjAnd WtCalculated using probability against word frequency, specific formula for calculation is as follows：
Wherein, njFor xjFirst category label and second category tag set and concentration element number；ntFor xtFirst
The number of class label and the element of the union of second category tag set.
4. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 2, its feature
It is：The normalization is carried out using power-law norm methods.
5. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 2, its feature
It is：Taken further to reduce memory space, principal component analysis (PCA) is used to the weighting local feature Aggregation Descriptor
Method dimensionality reduction.
6. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 1, its feature
It is：Described image local feature region is RootSIFT features.
7. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 1, its feature
It is：The γ values scope is [0,1].
8. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 1, its feature
It is：γ=0.5.
9. a kind of image search method based on weighting local feature Aggregation Descriptor according to claim 1-8, it is special
Sign is：The calculating of the distance uses Euclidean distance.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710609564.9A CN107527058B (en) | 2017-07-25 | 2017-07-25 | Image retrieval method based on weighted local feature aggregation descriptor |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710609564.9A CN107527058B (en) | 2017-07-25 | 2017-07-25 | Image retrieval method based on weighted local feature aggregation descriptor |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107527058A true CN107527058A (en) | 2017-12-29 |
| CN107527058B CN107527058B (en) | 2020-11-27 |
Family
ID=60680009
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710609564.9A Active CN107527058B (en) | 2017-07-25 | 2017-07-25 | Image retrieval method based on weighted local feature aggregation descriptor |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107527058B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108806774A (en) * | 2018-05-22 | 2018-11-13 | 长春师范大学 | Medical image search method based on geometrical constraint and spatial pixel intensity |
| CN110146080A (en) * | 2019-04-24 | 2019-08-20 | 佛山科学技术学院 | A kind of SLAM winding detection method and device based on mobile robot |
| CN110334763A (en) * | 2019-07-04 | 2019-10-15 | 北京字节跳动网络技术有限公司 | Model data file generation, image-recognizing method, device, equipment and medium |
| CN111242152A (en) * | 2018-11-29 | 2020-06-05 | 北京易讯理想科技有限公司 | Image retrieval method based on target extraction |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104951562A (en) * | 2015-07-08 | 2015-09-30 | 天云融创数据科技（北京）有限公司 | Image retrieval method based on VLAD (vector of locally aggregated descriptors) dual self-adaptation |
-
2017
- 2017-07-25 CN CN201710609564.9A patent/CN107527058B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104951562A (en) * | 2015-07-08 | 2015-09-30 | 天云融创数据科技（北京）有限公司 | Image retrieval method based on VLAD (vector of locally aggregated descriptors) dual self-adaptation |
Non-Patent Citations (3)
| Title |
|---|
| RELJA ARANDJELOVI´C等: "Three things everyone should know to improve object retrieval", 《CVPR2012》 * |
| 猴子聊人物: "词权重计算及应用", 《CSDN HTTPS://BLOG.CSDN.NET/YANGZHONGBLOG/ARTICLE/DETAILS/8723771》 * |
| 黄秀杰等: "基于最小重构误差的优化局部聚合描述符向量图像检索算法", 《计算机应用》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108806774A (en) * | 2018-05-22 | 2018-11-13 | 长春师范大学 | Medical image search method based on geometrical constraint and spatial pixel intensity |
| CN108806774B (en) * | 2018-05-22 | 2022-02-01 | 长春师范大学 | Medical image retrieval method based on geometric constraint and spatial pixel intensity |
| CN111242152A (en) * | 2018-11-29 | 2020-06-05 | 北京易讯理想科技有限公司 | Image retrieval method based on target extraction |
| CN110146080A (en) * | 2019-04-24 | 2019-08-20 | 佛山科学技术学院 | A kind of SLAM winding detection method and device based on mobile robot |
| CN110146080B (en) * | 2019-04-24 | 2024-01-19 | 佛山科学技术学院 | SLAM loop detection method and device based on mobile robot |
| CN110334763A (en) * | 2019-07-04 | 2019-10-15 | 北京字节跳动网络技术有限公司 | Model data file generation, image-recognizing method, device, equipment and medium |
| CN110334763B (en) * | 2019-07-04 | 2021-07-23 | 北京字节跳动网络技术有限公司 | Model data file generation method, model data file generation device, model data file identification device, model data file generation apparatus, model data file identification apparatus, and model data file identification medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107527058B (en) | 2020-11-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Jégou et al. | Aggregating local image descriptors into compact codes | |
| CN111198959B (en) | Two-stage image retrieval method based on convolutional neural network | |
| Maji et al. | Max-margin additive classifiers for detection | |
| WO2019134567A1 (en) | Sample set processing method and apparatus, and sample querying method and apparatus | |
| CN110222218B (en) | Image retrieval method based on multi-scale NetVLAD and depth hash | |
| CN107085607B (en) | Image feature point matching method | |
| Zheng et al. | $\mathcal {L} _p $-Norm IDF for Scalable Image Retrieval | |
| CN113918753B (en) | Image retrieval method based on artificial intelligence and related equipment | |
| Cao et al. | Deep priority hashing | |
| WO2013129580A1 (en) | Approximate nearest neighbor search device, approximate nearest neighbor search method, and program | |
| CN107527058A (en) | A kind of image search method based on weighting local feature Aggregation Descriptor | |
| CN110110800B (en) | Automatic image annotation method, device, equipment and computer readable storage medium | |
| EP3020203A1 (en) | Compact and robust signature for large scale visual search, retrieval and classification | |
| WO2023108995A1 (en) | Vector similarity calculation method and apparatus, device and storage medium | |
| CN110751027B (en) | Pedestrian re-identification method based on deep multi-instance learning | |
| CN111182364B (en) | Short video copyright detection method and system | |
| Singh et al. | Image corpus representative summarization | |
| CN105760875B (en) | The similar implementation method of differentiation binary picture feature based on random forests algorithm | |
| CN114461839A (en) | Multi-mode pre-training-based similar picture retrieval method and device and electronic equipment | |
| Zeng et al. | MCCN: Multimodal coordinated clustering network for large-scale cross-modal retrieval | |
| JP2002183171A (en) | Document data clustering system | |
| Guo | Research on sports video retrieval algorithm based on semantic feature extraction | |
| JP2017162009A (en) | Device, method, and program for generating feature quantity | |
| JP5833499B2 (en) | Retrieval device and program for retrieving content expressed by high-dimensional feature vector set with high accuracy | |
| JP6017277B2 (en) | Program, apparatus and method for calculating similarity between contents represented by set of feature vectors |
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 |