CN108985355A - A kind of data fusion method based on the orthogonal local sensitivity Hash of grouping - Google Patents
A kind of data fusion method based on the orthogonal local sensitivity Hash of grouping Download PDFInfo
- Publication number
- CN108985355A CN108985355A CN201810684566.9A CN201810684566A CN108985355A CN 108985355 A CN108985355 A CN 108985355A CN 201810684566 A CN201810684566 A CN 201810684566A CN 108985355 A CN108985355 A CN 108985355A
- Authority
- CN
- China
- Prior art keywords
- data
- hash
- vector
- fusion
- orthogonal
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
Abstract
A kind of data fusion method based on the orthogonal local sensitivity Hash of grouping, it concatenates the data characteristics vector of each mode to form original fusion feature vector first, then the local hash indexing method orthogonal using grouping, original fusion feature vector is mapped to hash index space under conditions of meeting and being grouped orthogonal, the quantity for finally controlling Hash projection vector realizes dimensionality reduction, completes data fusion.
Description
Technical field
The present invention relates to domain of data fusion to carry out Fusion Features using hash method come the data characteristics to different modalities
New data characteristics field is formed, it is especially a kind of based on the data fusion method for being grouped orthogonal local sensitivity Hash.
Background technique
In big data era, data source is various, self-assembling formation, magnanimity, and data are often half structure or without knot
Structure.This requires data science man and analyst can control multiplicity, multi-source data, carried out after they are combed excavate and
Analysis.In this process, data fusion just becomes an indispensable step.At the same time, with computer technology, communication skill
The development of art and microelectric technique, the various multi-source information systems towards complicated applications background largely occur, and force people right
Multiple sensors and different aforementioned sources more effectively integrate, to improve the degree of automation of information processing.Therefore, from 20 generation
It records from the seventies, multisensor (or multi-source) data fusion is just used as a new branch of science to develop rapidly.
Multisource data fusion is a kind of basic function generally existing in the mankind and other biological system.Have to human instinct
There are the information (scenery, sound, smell and tactile) for being detected the various Functional tissues (eye, ear, nose, four limbs) on body and elder generation
Test knowledge and carry out comprehensive ability, so as to around him environment and occurent event make estimation.Due to the sense of the mankind
Official has different measures characteristics, thus can measure the various physical phenomenons occurred within the scope of different spaces, and by different spies
The fusion treatment of sign is converted to the valuable explanation to environment.
Multisource data fusion is actually a kind of functional simulation to human brain integrated treatment challenge.In multisensor
In (or multi-source) system, the information that each signal source provides may have different characteristics: it is time-varying or non-time-varying, in real time
Perhaps non real-time perhaps gradual fuzzy perhaps determine accurate perhaps incomplete reliable become fastly or
It is non-reliable, it is mutually supporting or complementary, it is also possible to conflicting or conflict.The original substantially of multisource data fusion
For reason just as the process of human brain integrated treatment information, it fully utilizes multiple data resources, by various signal sources and
It the reasonable domination of its observation information and uses, by complementation of the various signal sources on room and time and redundancy according to certain
Optimality Criteria combines, and generates and the consistency of observing environment is explained and described.The target of data fusion is based on each signal
The information of source separation observation, exports more effective informations by the optimum organization to information.This is the knot of synergy
Fruit, its final purpose is the advantage to be cooperated using multiple signal sources, to improve the validity of whole system.
The multi-source data processing of single-sensor (or single source) signal processing or low level is all to human brain information process
A kind of low-level imitate, and source Data Fusion System be then by effectively utilize multi-source data obtain resource, come it is maximum
Obtain to limit the information content of detected target and environment.Between multisource data fusion and classical signals processing method there is also
Essential distinction, key are that multi-source data handled by data fusion has more complicated form, and usually different
Occur in data hierarchy, that is, data fusion has the feature of stratification.
Multisource data fusion is a kind of information for being directed to the system using multiple or multiclass sensor (information source) and carrying out
Processing method, it is associated with also referred to as multi-source, multi-source synthesis, sensor integration or Multi-sensor Fusion, but wider saying
It is multisource data fusion or Fusion, i.e. data fusion.Its definition has following several:
1) multistage various processing such as detected, be associated with, estimated and integrated with data to the information from multi-source, with
Obtain accurate state and identity estimation, and completely, timely Situation Assessment and threat estimating.
2) data fusion be single source is associated to the data of multi-source to information, related and combination, it is finer to obtain
Position and identity estimation, the complete and timely process of Situation Assessment.
3) several sensor observation informations chronologically obtained are subject to automatically under certain criterion using computer technology
The information process of analysis, Optimum Synthesis complete required decision and to estimate task and carry out.
4) data fusion be it is multi-level, many-sided single source detected automatically with the data of multi-source with information, be associated with,
Related, estimation and combined process.
5) data fusion is data splitting or information to estimate and predict the process of entity state.
With data warehouse, data integrated difference, the purpose of data fusion is not to concentrate in together all data
Unique truth is generated by standardization, but to generate intelligent decision-making as final goal, by the correlation in multiple data sources
Data are extracted, are merged, combing is integrated into an analysis data set.This analysis data set is a independent and flexible entity, energy
The variation with data source is reached to recombinate, adjust and update.Data fusion is better than data warehouse and data are more integrated just
It is that it can contain multi-source data.Fusion has the following performance excellent in terms of solving the problems, such as detection, tracking and identifying
Gesture:
1) survival ability is strong.Have several sensors that cannot utilize or be interfered or some target not covering model
When enclosing, information can be provided by always having a kind of sensor；
2) spatial coverage is extended.By the sensor zone of action of multiple overlapped coverages, space covering is extended
Range, a kind of sensor can detect the undetectable place of other sensors；
3) time coverage area is extended.Detection probability is improved with the synergistic effect of multiple sensors, some sensor can
To detect the target or event that other sensors cannot be taken into account；
4) confidence level is improved.One or more sensors confirm same target or event；
5) fuzziness of information is reduced.The united information of multisensor reduces uncertainty of objective；
6) detection performance is improved.To a variety of measurement effective integrations of target, the validity of detection is improved；
7) spatial resolution is improved.Multisensor can be obtained than any single-sensor higher resolution；
8) measurement space dimensionality is increased.System is not easily susceptible to enemy's action or the destruction of natural phenomena.
By taking remotely-sensed data as an example, analyzing the Methods on Multi-Sensors RS Image of areal can obtain comparing single piece of information
Source more accurate, more complete, more reliable estimation and judgement.Relative to single source Landsat images data, multi-source remote sensing image data institute
The information of offer has the following characteristics that
Redundancy: Methods on Multi-Sensors RS Image is identical to the expression, description or interpretation result of environment or target；
It is complementary: to refer to information from different freedom degree and mutually indepedent；
Cooperative: different sensors have dependence to other information when observing and handling information.
Information layered architectural characteristic: multi-source remote sensing information handled by data fusion can be on different level of information
Occur, these information levels include pixel layer, characteristic layer and decision-making level, and layered structure and parallel processing mechanism also ensure
The real-time of system.
And Data fusion technique can will be believed provided by the multiband information of single-sensor or different classes of sensor
Breath is integrated, and redundancy and contradiction that may be present between multi-sensor information are eliminated, and is subject to complementation, is improved remote sensing information and is extracted
Timeliness and reliability, improve the service efficiency of data.
At the same time, in order to which the part proposed the problems such as realizing efficient similarity calculation and similarity between data is quick
Sense Hash is a kind of valid data processing method in order to solve calculating and search in large-scale data based on similarity: will be counted
According to being compressed into compact Hash codes, certain simple distance by calculating Hash intersymbol quickly estimate initial data similarity or
Distance.
Local sensitivity hash method can also regard a kind of special dimension reduction method as, but not with general dimension reduction method
With local sensitivity hash method has the step of quantization --- each cryptographic Hash is the ratio of nonnegative integer, even 0 or 1
It is special.Quantization is advantageous in that cryptographic Hash can be easily indexed with Hash table.In addition, if each cryptographic Hash is 0 or 1
Bit, then similarity that can quickly between estimated data by the Hamming distances for calculating Hash intersymbol.In addition, using office
Portion's sensitive hash method can also realize the compression of storage in the case where not dimensionality reduction.Such as: the office of vector is directed to using one group
The 0-1 Bit String of K is grown into the real-valued vectors compression that one d is tieed up by portion's sensitive hash function, also real in storage even if K=d
Significant compression is showed.
Therefore, for multi-source heterogeneous data, if it is possible to it is more to realize to efficiently use the various characteristics of local sensitivity Hash
Source data fusion, will form more effective data indicates.Fused data can either achieve the purpose that dimensionality reduction, and can utilize
Local sensitivity Hash keeps the reasonable effective similar characteristic between data, and then for subsequent classification, detection and prediction etc.
Task.
Summary of the invention
Technical problem solved by the present invention is having overcome the deficiencies of the prior art and provide a kind of based on the orthogonal part of grouping
The data fusion method of sensitive hash solves under mass data environment, effectively melting between multi-source heterogeneous multi-modal data
It closes and Data Dimensionality Reduction problem.
The technical solution of the invention is as follows: a kind of based on the data fusion method for being grouped orthogonal local sensitivity Hash, packet
Include following steps:
(1) it concatenates the data characteristics vector of each mode to form original fusion feature vector；
(2) the local hash indexing method orthogonal using grouping, original fusion feature vector is orthogonal in satisfaction grouping
Under the conditions of be mapped to hash index space；
(3) quantity for controlling Hash projection vector realizes dimensionality reduction, completes data fusion.
Described concatenates the method to form original fusion feature vector for the data characteristics vector of each mode are as follows:
(1) the data characteristics vector f of n mode is set1, f2..., fn, characteristic dimension is respectively d1, d2..., dn,
In, n is integer；
(2) it enablesAnd f=[f1, f2..., fn], data characteristics vector dimension is K=N × L after fusion,
And K < d, K, N, L are positive integer.
The local hash indexing method orthogonal using grouping, original fusion feature vector is orthogonal in satisfaction grouping
Under conditions of be mapped to the method in hash index space are as follows:
(1) the random matrix H=[v of d × K is generated1, v2..., vK], in random matrix H each column be length be d with
Machine projection vector, every column element of random matrix H be all independent same distribution sample from standardized normal distributionEnable i
=0, wherein i is integer；
(2)Hi=[viN+1, viN+2..., viN+N], Hi=QiRi；
(3) Q is obtainediPreceding N column constitute new matrix wi=[wiN+1, wiN+2..., wiN+N], i=i+1, and it is transferred to step
(2), until i=L-1.
The quantity of the control Hash projection vector realizes dimensionality reduction, the method for completing data fusion are as follows: enablesAnd then obtain the data characteristics vector after data fusion
A kind of computer readable storage medium, the computer-readable recording medium storage has computer program, described
Computer program the step of the method as any such as claim 1- claim 4 is realized when being executed by processor.
It is a kind of based on the data fusion terminal device for being grouped orthogonal local sensitivity Hash, including memory, processor and
The computer program that can be run in the memory and on the processor is stored, the processor executes the meter
The step of the method as any such as claim 1- claim 4 is realized when calculation machine program.
The advantages of the present invention over the prior art are that:
(1) present invention meets under mass data environment, effective integration and data between multi-source heterogeneous multi-modal data
The problem of dimensionality reduction, for the data modality of separate sources, such as audio, image, video, by the data characteristics vector of each mode
Directly concatenation forms original fusion feature vector, guarantees the utmostly reservation to primary data information (pdi)；
(2) all feature vectors are being met grouping just using orthogonal local hash indexing method is grouped by the present invention
It is mapped to hash index space under conditions of friendship, realizes the data cross fusion between different data mode；
(3) present invention realizes Data Dimensionality Reduction by controlling the quantity of Hash projection vector, and is formed and can be used for mode knowledge
Not, machine learning it is various study and classification in fusion feature vector, it was demonstrated that the hamming of hypergeometric spy's local sensitivity Hash intersymbol away from
From still capable of providing between the unbiased esti-mator of angle vector, and variance ratio symbol accidental projection Hash is smaller, has good
Use value.
Detailed description of the invention
Fig. 1 is a kind of based on the data fusion method flow chart for being grouped orthogonal local sensitivity Hash.
Specific embodiment
Effective integration and data in order to preferably meet under mass data environment, between multi-source heterogeneous multi-modal data
Dimensionality reduction, as shown in Figure 1 to be a kind of based on the data fusion method flow chart for being grouped orthogonal local sensitivity Hash, of the invention is main
Purpose is to propose a kind of new side that data fusion and dimensionality reduction between multi-modal data are realized using local hash index
Method.This method is directed to the data modality of separate sources, such as audio, image, video, it is assumed that all data modalities are all converted to
Existing feature vector in the form of vectors.Original fusion is formed firstly, the data characteristics vector of each mode is directly concatenated
Feature vector, this mode can guarantee the utmostly reservation to primary data information (pdi).Secondly, being breathed out using orthogonal part is grouped
All feature vectors are mapped to hash index space by uncommon indexing means under conditions of meeting and being grouped orthogonal, are realized different
Data cross fusion between data modality.Data Dimensionality Reduction, and shape are realized finally by the quantity of control Hash projection vector
At the fusion feature vector that can be used in pattern-recognition, the various study of machine learning and classification.
Specifically, after the technical solution of this patent mainly passes through the projection vector grouping orthogonalization generated at random, it is right
Multi-modal data feature vector using local sensitivity hash method carry out Hash projection, last mixing together at fusion feature to
Amount.Hypergeometric spy's local sensitivity Hash uses the hash function of structuring sampling: being grouped the accidental projection vector composition of orthogonalization
Hash function.Can prove the Hamming distances of hypergeometric spy's local sensitivity Hash intersymbol still can provide between vector angle it is unbiased
Estimation, and variance ratio symbol accidental projection Hash is smaller.
Hypergeometric spy's local sensitivity Hash uses the random vector of grouping orthogonalization, the projection vector group of each orthogonalization
In each random vector certain isotropic probability distribution is still obeyed on probability.Therefore, two vector a and b are surpassed
The separated probability of hyperplane corresponding to the projection vector of bit local sensitivity hash function is proportional to angle (etc. between them
In θA, b/π).To utilize desired linear behavio(u)r, the Hamming distances of two corresponding hypergeometric spy local sensitivity Hash intersymbols are just
It provides between the unbiased esti-mator of angle vector.It specifically includes:
Firstly, for the data characteristics of the multiple modalities of multiple data sources, each data characteristics is with different dimensions
Feature vector indicates.A kind of most direct data fusion mode is that the dimension of each modal data feature vector is carried out concatenation shape
At a longer data characteristics vector.The data characteristics vector that suppose there is k mode, is denoted as f1, f2..., fk, each feature
Vector has different characteristics dimension, respectivelyThen feature vector concatenation after original fusion feature to
Amount is [f1, f2..., fk].Wherein, operator " [a, b] " indicate by vector a and vector b to be directly concatenated into length be original
The new vector of vector length sum.
Secondly, the Random Orthogonal vector for meeting data characteristics unbiased esti-mator after feature Hash generates.Pass through features described above string
Connecing the new feature vector dimension to be formed is d.So standardized normal distribution is tieed up from dMiddle independent sample obtain it is N number of with
Machine projection vector, and it is orthogonalized.Specifically, for N number of independent identically distributed random vectorWherein each vector obeys standardized normal distribution1≤N≤d is column with them
It constitutes matrix and calculates its QR decomposition, to generate N number of orthogonal vector w1, w2..., wN.If given to any two
Non-vanishing vectorWe define N number of instruction stochastic variableFor
Wherein,It indicates non-vanishing vector a utilizing vector wiProject obtained value, that is, by vector a and vector
wiCarry out inner product calculating.At this point, enablingIn unit ball, for uniform sampling from Sd-1Random vector v, haveSo to any 1≤i≤N, haveAnd such as
FruitIsotropic probability distribution is obeyed, thenObey Sd-1On be uniformly distributed.Assuming that enabling Si-1
For { w1..., wi-1Subspace, the orthogonal intersection cast shadow matrix to its orthogonal complementary subspace isSoIt enablesSo, to arbitrary 1≤i≤N, have
So, as i=1,
As 1 < i≤N, w is considerediDistribution:It is a underriding Grassmann manifold GI-1, d-i+1On uniformly point
The random matrix of cloth, thenObey Grassmann manifold GD-i+1, i-1On be uniformly distributed.BecauseWith v1, v2..., vi-1It is independent, so viWithIt is independent.Also,Obey isotropism
Distribution, andIt obeysUnit ball on be uniformly distributed, so,
Assuming that carrying out the Hash code length after local sensitivity Hash to the concatenation feature vector that length is d (namely after Hash
Feature vector dimension) be K=N × L, i≤N≤d, to arbitrary non-vanishing vector
That is, under the meaning of a constant multiplication factor C=K/ π, Hamming distances dham(h (a), h (b)) is provided
To θA, bUnbiased esti-mator.
Finally, due to can be guaranteed after being grouped orthogonalization based on above-mentioned random vector by the orthogonal mode of above-mentioned grouping
Vector projection forms after new Hash codes the unbiased esti-mator that can guarantee to original feature vector.This patent is utilizing above-mentioned grouping just
The local sensitivity hash mode of friendship realizes that multi-modal feature vector directly concatenates the further fusion of the d dimensional feature vector of formation
And compression.
Assuming that a length of K of final Hash codes, that is, the random a d for Hash projection of K that generates tie up random vector, or
Say that fused new feature vector dimension is K.Without loss of generality, it is assumed that K=N × L, wherein 1≤N≤d.That is K
Independent accidental projection vector is divided into L group, respectively by N number of accidental projection vector orthogonalization in each group.To obtain K=N
× L new projection vector w1, w2..., wK, and constitute K local sensitivity hash functionIts
In,
The fusion feature vector for being d for length after directly concatenation finally, for input, to each input vector, in utilization
It states K=N × L hash function and generates L group, every group of size is N, the Hash codes of total K bit.The Hash codes are finally to merge
Obtained data characteristics.It is special in order to efficiently control fusion since the fusion feature dimension d after directly concatenating may be very big
Dimension is levied, by controlling the size of K, i.e. guarantee K < d can effectively realize Data Dimensionality Reduction.
In order to better illustrate method described in this patent, specifically described below by pseudo-code of the algorithm once
The process of implementation.
The content that description in the present invention is not described in detail belongs to the well-known technique of those skilled in the art.
Claims (6)
1. a kind of based on the data fusion method for being grouped orthogonal local sensitivity Hash, it is characterised in that include the following steps:
(1) it concatenates the data characteristics vector of each mode to form original fusion feature vector；
(2) using orthogonal local hash indexing method is grouped, original fusion feature vector is being met into the orthogonal condition of grouping
Under be mapped to hash index space；
(3) quantity for controlling Hash projection vector realizes dimensionality reduction, completes data fusion.
2. according to claim 1 a kind of based on the data fusion method for being grouped orthogonal local sensitivity Hash, feature exists
In: it is described that the data characteristics vector of each mode is concatenated into the method to form original fusion feature vector are as follows:
(1) the data characteristics vector f of n mode is set1,f2,…,fn, characteristic dimension is respectively d1,d2,…,dn, wherein n is
Integer；
(2) it enablesAnd f=[f1,f2,…,fn], data characteristics vector dimension is K=N × L, and K < after fusion
D, K, N, L are positive integer.
3. according to claim 2 a kind of based on the data fusion method for being grouped orthogonal local sensitivity Hash, feature exists
In: original fusion feature vector is being met the orthogonal item of grouping by the local hash indexing method orthogonal using grouping
The method in hash index space is mapped under part are as follows:
(1) the random matrix H=[v of d × K is generated1,v2,…,vK], each column is the accidental projection that length is d in random matrix H
Vector, every column element of random matrix H be all independent same distribution sample from standardized normal distributionI=0 is enabled,
In, i is integer；
(2)Hi=[viN+1,viN+2,…,viN+N], Hi=QiRi；
(3) Q is obtainediPreceding N column constitute new matrix wi=[wiN+1,wiN+2,…,wiN+N], i=i+1, and step (2) are transferred to, directly
To i=L-1.
4. according to claim 3 a kind of based on the data fusion method for being grouped orthogonal local sensitivity Hash, feature exists
In: described, the quantity of control Hash projection vector realizes dimensionality reduction, the method for completing data fusion are as follows: enablesAnd then obtain the data characteristics vector after data fusion
5. a kind of computer readable storage medium, the computer-readable recording medium storage has computer program, feature
It is, the step such as any the method for claim 1- claim 4 is realized when the computer program is executed by processor
Suddenly.
6. a kind of based on the data fusion terminal device for being grouped orthogonal local sensitivity Hash, including memory, processor and deposit
Store up the computer program that can be run in the memory and on the processor, it is characterised in that: the processor is held
The step of the method as any such as claim 1- claim 4 is realized when computer program described in row.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810684566.9A CN108985355A (en) | 2018-06-28 | 2018-06-28 | A kind of data fusion method based on the orthogonal local sensitivity Hash of grouping |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810684566.9A CN108985355A (en) | 2018-06-28 | 2018-06-28 | A kind of data fusion method based on the orthogonal local sensitivity Hash of grouping |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108985355A true CN108985355A (en) | 2018-12-11 |
Family
ID=64539164
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810684566.9A Pending CN108985355A (en) | 2018-06-28 | 2018-06-28 | A kind of data fusion method based on the orthogonal local sensitivity Hash of grouping |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108985355A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110750731A (en) * | 2019-09-27 | 2020-02-04 | 成都数联铭品科技有限公司 | Duplicate removal method and system for news public sentiment |
| CN111125448A (en) * | 2019-12-23 | 2020-05-08 | 中国航空工业集团公司沈阳飞机设计研究所 | Large-scale aerial task decision method and system |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106570482A (en) * | 2016-11-03 | 2017-04-19 | 深圳先进技术研究院 | Method and device for identifying body motion |
-
2018
- 2018-06-28 CN CN201810684566.9A patent/CN108985355A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106570482A (en) * | 2016-11-03 | 2017-04-19 | 深圳先进技术研究院 | Method and device for identifying body motion |
Non-Patent Citations (1)
| Title |
|---|
| JIANQIU JI 等: "Batch-Orthogonal Locality-Sensitive Hashing for Angular Similarity", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110750731A (en) * | 2019-09-27 | 2020-02-04 | 成都数联铭品科技有限公司 | Duplicate removal method and system for news public sentiment |
| CN110750731B (en) * | 2019-09-27 | 2023-10-27 | 成都数联铭品科技有限公司 | Method and system for removing duplicate of news public opinion |
| CN111125448A (en) * | 2019-12-23 | 2020-05-08 | 中国航空工业集团公司沈阳飞机设计研究所 | Large-scale aerial task decision method and system |
| CN111125448B (en) * | 2019-12-23 | 2023-04-07 | 中国航空工业集团公司沈阳飞机设计研究所 | Large-scale aerial task decision method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| He et al. | Bounding box regression with uncertainty for accurate object detection | |
| Liu et al. | Learning efficient spatial-temporal gait features with deep learning for human identification | |
| Nelson et al. | Large-scale tests of a keyed, appearance-based 3-D object recognition system | |
| CN101558431A (en) | Face authentication device | |
| CN103150546A (en) | Video face identification method and device | |
| KR20220004009A (en) | Key point detection method, apparatus, electronic device and storage medium | |
| CN116106878A (en) | Big data analysis system and method | |
| CN103093243A (en) | High resolution panchromatic remote sensing image cloud discriminating method | |
| KR20220098312A (en) | Method, apparatus, device and recording medium for detecting related objects in an image | |
| CN108985355A (en) | A kind of data fusion method based on the orthogonal local sensitivity Hash of grouping | |
| CN113139415A (en) | Video key frame extraction method, computer device and storage medium | |
| Gianaria et al. | Robust gait identification using Kinect dynamic skeleton data | |
| Panigrahi et al. | MS-ML-SNYOLOv3: A robust lightweight modification of SqueezeNet based YOLOv3 for pedestrian detection | |
| Joodi et al. | Increasing validation accuracy of a face mask detection by new deep learning model-based classification | |
| Miao et al. | Abnormal Behavior Learning Based on Edge Computing toward a Crowd Monitoring System | |
| Li et al. | Msvit: training multiscale vision transformers for image retrieval | |
| WO2018173848A1 (en) | Object tracking system, intelligent imaging device, object feature quantity extraction device, object feature quantity extraction method, and recording medium | |
| CN114758285B (en) | Video interaction action detection method based on anchor freedom and long-term attention perception | |
| Li et al. | Detection of partially occluded pedestrians by an enhanced cascade detector | |
| Rashidan et al. | Detection of different classes moving object in public surveillance using artificial neural network (ann) | |
| Nijhawan et al. | Real-time Object Detection for Visually Impaired with Optimal Combination of Scores | |
| CN114677611A (en) | Data identification method, storage medium and device | |
| US20220335291A1 (en) | Learning apparatus, learning method, inference apparatus, inference method, and recording medium | |
| Zhang et al. | Lightweight network for small target fall detection based on feature fusion and dynamic convolution | |
| Zhang et al. | Towards a fast and accurate road object detection algorithm based on convolutional neural networks |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181211 |
|
| RJ01 | Rejection of invention patent application after publication |