CN109815364A - A kind of massive video feature extraction, storage and search method and system - Google Patents
A kind of massive video feature extraction, storage and search method and system Download PDFInfo
- Publication number
- CN109815364A CN109815364A CN201910047518.3A CN201910047518A CN109815364A CN 109815364 A CN109815364 A CN 109815364A CN 201910047518 A CN201910047518 A CN 201910047518A CN 109815364 A CN109815364 A CN 109815364A
- Authority
- CN
- China
- Prior art keywords
- video
- feature
- massive
- convolution
- convolutional neural
- 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
- 238000000034 method Methods 0.000 title claims abstract description 54
- 238000003860 storage Methods 0.000 title claims abstract description 27
- 238000000605 extraction Methods 0.000 title claims abstract description 17
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 29
- 239000000284 extract Substances 0.000 claims abstract description 8
- 238000005070 sampling Methods 0.000 claims abstract description 6
- 230000000052 comparative effect Effects 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 230000006870 function Effects 0.000 claims description 4
- 230000011218 segmentation Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 230000003100 immobilizing effect Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 239000004576 sand Substances 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 claims description 2
- 230000001537 neural effect Effects 0.000 claims 3
- 239000000203 mixture Substances 0.000 claims 1
- 238000011176 pooling Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 239000012634 fragment Substances 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/73—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to technical field of video processing, specifically, it is related to a kind of massive video feature extraction, storage and search method and system, feature extraction passes through training depth convolutional neural networks and extracts, storage method includes video slicing, interval sampling, terrestrial reference identification, depth convolutional neural networks handle to obtain Hash coding and feature, selection key frame and etc., search method includes landmark information and picture retrieval, when inputting picture, Hash coding and feature are extracted by depth convolutional neural networks first, uncommon coding quick-searching in hash index library, then it is accurately matched in video features library with feature, video information is obtained in video information library after obtaining similarly target index;The present invention has preferably accuracy, and precision is high.
Description
Technical field
The present invention relates to technical field of video processing, specifically, being related to a kind of massive video feature extraction, storage and inspection
Rope method and system.
Background technique
Image retrieval technologies can be divided into text based retrieval method and content-based retrieval method, text based
Search method inputs the mark of keyword and all images firstly the need of text marking is added in advance for the image in all search libraries
Explanatory notes is originally matched.The search method retrieval flow of image content-based is similar with text based retrieval method, different
Be content-based retrieval mode be description using the visual signature of image as image, Visual Feature Retrieval Process have at present sift,
The conventional methods such as hog, harr, gist, retrieval mode are to extract characteristic matching retrieval in Image Visual Feature, with feature database.
Current video retrieval method accuracy rate is low, and precision is small, is not able to satisfy the demand of people.
Summary of the invention
The contents of the present invention are to provide a kind of massive video feature extraction, storage and search method and system, can gram
Take certain or certain defects of the prior art.
A kind of massive video feature extracting method according to the present invention comprising following steps:
A, by video input to depth convolutional neural networks;
B, depth convolutional neural networks carry out feature extraction to video;
C, video features are obtained.
2, a kind of massive video feature extracting method according to claim 1, it is characterised in that: the depth convolution
For neural network based on ResNet101, conv-5 includes c5-1, c5-2 and c5-3, carries out attention behaviour in conv-5
Make as follows:
One layer of convolution kernel of c5-1 connection be the convolution of 3x3 and export 512 characteristic patterns, then connects the convolution of one layer of 3x3 and defeated
A characteristic pattern out, attention map of this feature figure as the direction spatial;
The convolution of one layer of 1x1 of c5-2 connection keeps dimension 2048 to tie up constant, and does global poolization operation, obtains channel
The attention map in direction;
The calculation formula of feature extraction is as follows:
Wherein, Fc-3Indicate c5-3 layers of characteristic pattern,The operation of representing matrix element multiplication,Representing matrix element addition
Operation, MsAnd McRespectively indicate spatial attention operation and channel attention operation;Attention is last
Feature vector of the one layer of convolution by 2048 dimension of Generalized-mean pooling output, and use L2 normalization characteristic.
Preferably, the loss function of depth convolutional neural networks is contrastive loss, training set is
Structure-from-motion, after the completion of training, immobilizing foundation model parameter, trains the Hash coding layer of 512 dimensions respectively,
And support the full articulamentum of 4784 terrestrial references.
Preferably, increase the full articulamentum FC of one layer of 2048 dimension after training depth convolutional neural networks extract feature, and
Concat operation is carried out to feature and FC, 4096 dimensional features are formed, by pca dimensionality reduction to 2048, as final feature.
The present invention extracts video features by depth convolutional neural networks, becomes apparent from feature, more accurately.
The present invention also provides a kind of massive video Feature Extraction System based on depth convolutional neural networks, the depth
Convolutional neural networks based on ResNet101, the conv-5 module on ResNet101 include c5-1, c5-2 and c5-3 with
And attention operation module, attention operation module include spatial attention operation module, channel
Attention operation module and final operation module, in which:
Spatial attention operation module is used to handle the convolution that one layer of convolution kernel of c5-1 connection is 3x3 and output
512 characteristic patterns, then connect the convolution of one layer of 3x3 and export a characteristic pattern;
Channel attention operation module is used to handle the convolution of one layer of 1x1 of c5-2 connection, and dimension 2048 is kept to tie up
It is constant, and do global poolization operation;
The relationship that final operation module is used to handle between c5-1, c5-2 and c5-3 obtains final feature.
A kind of massive video storage method according to the present invention comprising following steps:
One, by video slicing, it is divided into the segment being made of camera lens by Shot Detection, and be stored in video information library;
Two, segment interval sampling generate sample frame, sample frame input depth convolutional neural networks after obtain Hash coding and
Feature;
Three, terrestrial reference identification is carried out in sample frame input terrestrial reference identification model, is as a result stored in video terrestrial reference storehouse;
Four, the correlation between comparative feature selects key frame, and the feature of key frame is stored in video features library, crucial
The Hash coding deposit hash index library of frame.
Preferably, video slicing is carried out in video task scheduling system in step 1, tracking stream and the overall situation are utilized
Color of image continuity comprehensive characteristics determine camera lens time boundary, by Video segmentation are the segment being made of camera lens.
Preferably, the information in segment deposit video information library includes start frame number, terminates frame number, frame length in step 1
With affiliated video.
Preferably, segment is sampled according to 30 frame periods in step 2.
Preferably, ground identifies method for distinguishing are as follows: if some classification softmax in terrestrial reference identification model in step 3
Value is greater than 0.8 and thinks that the ground of current lens is designated as the category, and result is stored in video terrestrial reference storehouse.
Preferably, in step 4, the method for the correlation between comparative feature are as follows: high using threshold value 0.55 as boundary
Think similar in the threshold value, similar feature only retains first feature, to obtain one or more pass for representing the camera lens
Key frame.
Preferably, the separability that the building of Hash coding reinforces coding by increasing sigmoid in training, with 0.5 threshold
Value is divided into 0,1 coding.
The present invention also provides a kind of massive video storage systems, use a kind of above-mentioned sea based on deep learning
Measure video storage method.
The present invention also provides a kind of massive video search methods comprising following steps:
(1), landmark information or image are inputted;
(2), when inputting landmark information, into being searched in video terrestrial reference storehouse;When inputting picture, pass through depth convolution mind first
Hash coding and feature are extracted through network;
(3), Hash coding quick-searching in hash index library, is then accurately matched in video features library with feature,
Video information is obtained in video information library after obtaining similarly target index.
Preferably, after extracting Hash coding and feature, Hash coding is searched in history search library is in step (2)
It is no to have similar retrieval, if so, directly accurately being matched with feature, if nothing, carry out step (3).
Preferably, accurate matched method are as follows: obtained aspect indexing after being retrieved according to Hash obtains candidate feature
Number N, form [N, 2048] eigenmatrix, with retrieving image feature carry out COS distance calculating, according to matching similarity into
After row sequence, video segment information is inquired from video information library, cooperation screenshot is output to interface.
The present invention also provides a kind of massive video searching systems, use a kind of above-mentioned massive video search method.
The present invention include video lens segmentation, key-frame extraction, characteristic storage matched with index, quick-searching, terrestrial reference with
The technologies such as identification fully describe the terrestrial reference feature in image using the depth characteristic of depth convolutional neural networks, before increasing terrestrial reference
The accuracy rate of phase identification and later period index.By building hash index library and video features library, while guaranteeing quick-searching,
Improve the matching precision of retrieval.The present invention provide it is a kind of based on deep neural network for landmark image feature extraction, deposit
Storage, the frame indexed adapt to the terrestrial reference retrieval and identification of extensive video.
Detailed description of the invention
Fig. 1 is a kind of flow chart of massive video feature extracting method in embodiment 1;
Fig. 2 is a kind of structural frames of the massive video Feature Extraction System based on depth convolutional neural networks in embodiment 1
Figure;
Fig. 3 is a kind of flow chart of massive video storage method in embodiment 1;
Fig. 4 is a kind of flow chart of massive video search method in embodiment 1.
Specific embodiment
To further appreciate that the contents of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments.It should be understood that
, embodiment be only to the present invention explain and and it is non-limiting.
Embodiment 1
As shown in Figure 1, present embodiments providing a kind of massive video feature extracting method, comprising the following steps:
A, by video input to depth convolutional neural networks;
B, depth convolutional neural networks carry out feature extraction to video;
C, video features are obtained.
In the present embodiment, for the depth convolutional neural networks based on ResNet101, conv-5 includes c5-1, c5-
It is as follows to carry out attention operation in conv-5 by 2 and c5-3:
One layer of convolution kernel of c5-1 connection be the convolution of 3x3 and export 512 characteristic patterns, then connects the convolution of one layer of 3x3 and defeated
A characteristic pattern out, attention map of this feature figure as the direction spatial;
The convolution of one layer of 1x1 of c5-2 connection keeps dimension 2048 to tie up constant, and does global poolization operation, obtains channel
The attention map in direction;
The calculation formula of feature extraction is as follows:
Wherein, Fc-3Indicate c5-3 layers of characteristic pattern,The operation of representing matrix element multiplication,Representing matrix element addition
Operation, MsAnd McRespectively indicate spatial attention operation and channel attention operation;Attention is last
Feature vector of the one layer of convolution by 2048 dimension of Generalized-mean pooling output, and use L2 normalization characteristic.
In the present embodiment, the loss function of depth convolutional neural networks is contrastive loss, and training set is
Structure-from-motion, after the completion of training, immobilizing foundation model parameter, trains the Hash coding layer of 512 dimensions respectively,
And support the full articulamentum of 4784 terrestrial references.
In the present embodiment, after training depth convolutional neural networks extract feature, increase the full articulamentum FC of one layer of 2048 dimension,
And concat operation is carried out to feature and FC, 4096 dimensional features are formed, by pca dimensionality reduction to 2048, as final feature.
As shown in Fig. 2, present embodiments providing a kind of massive video feature extraction system based on depth convolutional neural networks
System, the depth convolutional neural networks based on ResNet101, the conv-5 module on ResNet101 include c5-1,
C5-2 and c5-3 and attention operation module, attention operation module include spatial attention operation mould
Block, channel attention operation module and final operation module, in which:
Spatial attention operation module is used to handle the convolution that one layer of convolution kernel of c5-1 connection is 3x3 and output
512 characteristic patterns, then connect the convolution of one layer of 3x3 and export a characteristic pattern;
Channel attention operation module is used to handle the convolution of one layer of 1x1 of c5-2 connection, and dimension 2048 is kept to tie up
It is constant, and do global poolization operation;
The relationship that final operation module is used to handle between c5-1, c5-2 and c5-3 obtains final feature.
As shown in figure 3, present embodiments providing a kind of massive video storage method comprising following steps:
One, by video slicing, it is divided into the segment being made of camera lens by Shot Detection, and be stored in video information library;
Two, segment interval sampling generates sample frame, after sample frame inputs depth convolutional neural networks (V-DIR i.e. in figure)
Obtain Hash coding and feature;
Three, terrestrial reference identification is carried out in sample frame input terrestrial reference identification model, is as a result stored in video terrestrial reference storehouse;
Four, the correlation between comparative feature selects key frame, and the feature of key frame is stored in video features library, crucial
The Hash coding deposit hash index library of frame.Massive video is analyzed in advance by depth convolutional neural networks, with certain in video
Interval sampling frame, and according to feature selecting key frame, characteristic storage only selects key frame to be stored, and accounts for reduce space
With.
Depth convolutional neural networks are a kind of depth convolutional neural networks, and classification training characteristics extract model, Hash coding
Layer, terrestrial reference identification model.It is basic model with ResNet101, conv5 convolution is defeated by Generalized-mean pooling
The feature vector of 2048 dimensions out, training uses contrastive loss as loss function, after the completion of training, fixes base respectively
Plinth model parameter, the Hash coding layer of 512 dimension of training are made of one layer of full connection and sigmoid, and support 4784 terrestrial references
Full articulamentum, be made of one layer of full connection and softmax.
Terrestrial reference identification model is the latter linked full articulamentum of characteristic layer in depth convolutional neural networks model, using there is mark
What the landmark image of label was trained, classification is predicted by softmax mode.
In the present embodiment, in step 1, video slicing is to carry out in video task scheduling system, and complete video is pressed
Fragment is carried out according to 3500 frame lengths, determines camera lens time side using tracking stream and global image distribution of color continuity comprehensive characteristics
Video segmentation is the segment being made of camera lens by boundary.
In the present embodiment, in step 1, the information in segment deposit video information library includes start frame number, terminates frame number, frame
Long and affiliated video.
In the present embodiment, in step 2, segment is sampled according to 30 frame periods.
In the present embodiment, in step 3, ground identifies method for distinguishing are as follows: if some classification in terrestrial reference identification model
Softmax value is greater than 0.8 and thinks that the ground of current lens is designated as the category and then thinks that the ground of current lens is designated as the category, will tie
Fruit is stored in video terrestrial reference storehouse.
In the present embodiment, in step 4, the method for the correlation between comparative feature are as follows: compared with COS distance, with threshold
Value 0.55 is used as boundary, thinks similar higher than the threshold value, and similar feature only retains first feature, represents the camera lens to obtain
One or more key frame.It will be stored in the 2048 dimensional features deposit video features library of key frame, the Hash coding of 512 dimensions
It is stored in hash index library.
In the present embodiment, the separability that the building of Hash coding reinforces coding by increasing sigmoid in training, with 0.5
Threshold value is divided into 0,1 coding.
A kind of massive video storage system is present embodiments provided, a kind of above-mentioned massive video storage method is used.
As shown in figure 4, present embodiments providing a kind of massive video search method comprising following steps:
(1), landmark information or image are inputted;
(2), when inputting landmark information, into being searched in video terrestrial reference storehouse;When inputting picture, pass through depth convolution mind first
Hash coding and feature are extracted through network;
(3), Hash coding quick-searching in hash index library, is then accurately matched in video features library with feature,
Video information is obtained in video information library after obtaining similarly target index.
In the present embodiment, in step (2), after extracting Hash coding and feature, Hash coding is searched in history search library
Whether there is similar retrieval, if so, directly accurately being matched with feature, if nothing, carries out step (3).
In the present embodiment, accurate matched method are as follows: obtained aspect indexing after being retrieved according to Hash obtains candidate spy
Number N is levied, the eigenmatrix of [N, 2048] is formed, the feature with retrieving image carries out COS distance calculating, according to matching similarity
After being ranked up, video segment information is inquired from video information library, cooperation screenshot is output to interface.Made using depth Hash
For quick indexing mode, the COS distance for calculating whole features is not needed, Hash code storage in the database, uses
Hamming distance is matched, and quick-searching is guaranteed.After quick-searching terrestrial reference, and the uncertain terrestrial reference retrieved with it is defeated
Enter terrestrial reference matching, compares to further calculate COS distance by video features library, to guarantee precise search.
A kind of massive video searching system is present embodiments provided, a kind of above-mentioned massive video search method is used.
4784 classes marked are supported in storage, but retrieval is not limited thereto, and input the landmark image in non-4784 class,
Terrestrial reference in similar video can be equally retrieved, and is accurately positioned position in video.
Schematically the present invention and embodiments thereof are described above, description is not limiting, institute in attached drawing
What is shown is also one of embodiments of the present invention, and actual structure is not limited to this.So if the common skill of this field
Art personnel are enlightened by it, without departing from the spirit of the invention, are not inventively designed and the technical solution
Similar frame mode and embodiment, are within the scope of protection of the invention.
Claims (17)
1. a kind of massive video feature extracting method, it is characterised in that: the following steps are included:
A, by video input to depth convolutional neural networks;
B, depth convolutional neural networks carry out feature extraction to video;
C, video features are obtained.
2. a kind of massive video feature extracting method according to claim 1, it is characterised in that: the depth convolutional Neural
For network based on ResNet101, conv-5 includes c5-1, c5-2 and c5-3, carries out attention operation such as in conv-5
Under:
One layer of convolution kernel of c5-1 connection is the convolution of 3x3 and exports 512 characteristic patterns, then connects the convolution of one layer of 3x3 and export one
Open characteristic pattern, attention map of this feature figure as the direction spatial;
The convolution of one layer of 1x1 of c5-2 connection keeps dimension 2048 to tie up constant, and does global poolization operation, obtains the direction channel
Attention map;
Feature calculation formula is as follows:
Wherein, Fc-3Indicate c5-3 layers of characteristic pattern,The operation of representing matrix element multiplication,Representing matrix element add operation,
MsAnd McRespectively indicate spatial attention operation and channel attention operation;Attention the last layer volume
Feature vector of the product by 2048 dimension of Generalized-meanpooling output, and use L2 normalization characteristic.
3. a kind of massive video feature extracting method according to claim 2, it is characterised in that: the depth convolutional Neural
The loss function of network is contrastive loss, training set structure-from-motion, after the completion of training, point
Other immobilizing foundation model parameter, the Hash coding layer of 512 dimension of training, and support the full articulamentum of 4784 terrestrial references.
4. a kind of massive video feature extracting method according to claim 3, it is characterised in that: the training depth convolution
After neural network extracts feature, increase the full articulamentum FC of one layer of 2048 dimension, and concat operation, composition are carried out to feature and FC
4096 dimensional features, by pca dimensionality reduction to 2048, as final feature.
5. a kind of massive video Feature Extraction System based on depth convolutional neural networks, it is characterised in that: the depth convolution
Neural network based on ResNet101, the conv-5 module on ResNet101 include c5-1, c5-2 and c5-3 and
Attention operation module, attention operation module include spatial attention operation module, channel
Attention operation module and final operation module, in which:
Spatial attention operation module is used to handle the convolution that one layer of convolution kernel of c5-1 connection is 3x3 and exports 512
Characteristic pattern, then connect the convolution of one layer of 3x3 and export a characteristic pattern;
Channel attention operation module is used to handle the convolution of one layer of 1x1 of c5-2 connection, keeps the dimension of dimension 2048 not
Become, and does global poolization operation;
The relationship that final operation module is used to handle between c5-1, c5-2 and c5-3 obtains final feature.
6. a kind of massive video storage method, it is characterised in that: the following steps are included:
One, by video slicing, it is divided into the segment being made of camera lens by Shot Detection, and be stored in video information library;
Two, segment interval sampling generates sample frame, obtains Hash coding and feature after sample frame input depth convolutional neural networks;
Three, terrestrial reference identification is carried out in sample frame input terrestrial reference identification model, is as a result stored in video terrestrial reference storehouse;
Four, the correlation between comparative feature selects key frame, and the feature of key frame is stored in video features library, key frame
Hash coding deposit hash index library.
7. a kind of massive video storage method according to claim 6, it is characterised in that: in step 1, video slicing is
It is carried out in video task scheduling system, determines camera lens using tracking stream and global image distribution of color continuity comprehensive characteristics
Video segmentation is the segment being made of camera lens by time boundary.
8. a kind of massive video storage method according to claim 7, it is characterised in that: in step 1, segment deposit view
The information of frequency information bank includes start frame number, terminates frame number, frame length and affiliated video.
9. a kind of massive video storage method according to claim 8, it is characterised in that: in step 2, segment is according to 30
Frame period sampling.
10. a kind of massive video storage method according to claim 9, it is characterised in that: in step 3, terrestrial reference identification
Method are as follows: think that the ground of current lens is designated as such if some classification softmax value in terrestrial reference identification model is greater than 0.8
Not, result is stored in video terrestrial reference storehouse.
11. a kind of massive video storage method according to claim 10, it is characterised in that: in step 4, comparative feature
Between correlation method are as follows: using threshold value 0.55 as boundary, think similar higher than the threshold value, similar feature only retains head
A feature, to obtain one or more key frame for representing the camera lens.
12. a kind of massive video storage method according to claim 11, it is characterised in that: the building of Hash coding is by instructing
Increase the separability that sigmoid reinforces coding in white silk, 0,1 coding is divided into 0.5 threshold value.
13. a kind of massive video storage system, it is characterised in that: it uses one kind of any one of claim 6-12 to be based on
The massive video storage method of deep learning.
14. a kind of massive video search method, it is characterised in that: the following steps are included:
(1), landmark information or image are inputted;
(2), when inputting landmark information, into being searched in video terrestrial reference storehouse;When inputting picture, pass through depth convolutional Neural net first
Network extracts Hash coding and feature;
(3), Hash coding quick-searching in hash index library, is then accurately matched in video features library with feature, is obtained
Similarly video information is obtained in video information library after target index.
15. a kind of massive video search method as claimed in claim 7, it is characterised in that: in step (2), extract Hash and compile
After code and feature, Hash coding has searched whether similar retrieval in history search library, if so, directly with feature accurate
Match, if nothing, carries out step (3).
16. a kind of massive video search method as claimed in claim 8, it is characterised in that: accurate matched method are as follows: according to
Obtained aspect indexing after Hash retrieval, obtains candidate feature number N, forms the eigenmatrix of [N, 2048], same to retrieving image
Feature carry out COS distance calculating, after being ranked up according to matching similarity, inquire video clip from video information library
Information, cooperation screenshot are output to interface.
17. a kind of massive video searching system, it is characterised in that: it uses a kind of sea of any one of claim 14-16
Measure video features search method.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910047518.3A CN109815364B (en) | 2019-01-18 | 2019-01-18 | Method and system for extracting, storing and retrieving mass video features |
PCT/CN2020/072969 WO2020147857A1 (en) | 2019-01-18 | 2020-01-19 | Method and system for extracting, storing and retrieving mass video features |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910047518.3A CN109815364B (en) | 2019-01-18 | 2019-01-18 | Method and system for extracting, storing and retrieving mass video features |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109815364A true CN109815364A (en) | 2019-05-28 |
CN109815364B CN109815364B (en) | 2020-01-14 |
Family
ID=66603492
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910047518.3A Expired - Fee Related CN109815364B (en) | 2019-01-18 | 2019-01-18 | Method and system for extracting, storing and retrieving mass video features |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN109815364B (en) |
WO (1) | WO2020147857A1 (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110489685A (en) * | 2019-08-19 | 2019-11-22 | 腾讯科技(武汉)有限公司 | A kind of Webpage display process, system and relevant apparatus and storage medium |
CN110688524A (en) * | 2019-09-24 | 2020-01-14 | 深圳市网心科技有限公司 | Video retrieval method and device, electronic equipment and storage medium |
CN110738128A (en) * | 2019-09-19 | 2020-01-31 | 天津大学 | repeated video detection method based on deep learning |
CN110769291A (en) * | 2019-11-18 | 2020-02-07 | 上海极链网络科技有限公司 | Video processing method and device, electronic equipment and storage medium |
CN110769276A (en) * | 2019-11-07 | 2020-02-07 | 成都国腾实业集团有限公司 | Frame extraction slice video detection method based on MD5 |
WO2020147857A1 (en) * | 2019-01-18 | 2020-07-23 | 上海极链网络科技有限公司 | Method and system for extracting, storing and retrieving mass video features |
CN111444390A (en) * | 2020-04-02 | 2020-07-24 | 徐州工程学院 | Spark and depth hash based video parallel retrieval method |
CN112035701A (en) * | 2020-08-11 | 2020-12-04 | 南京烽火星空通信发展有限公司 | Internet short video source tracing method and system |
CN113297899A (en) * | 2021-03-23 | 2021-08-24 | 上海理工大学 | Video hash algorithm based on deep learning |
CN113313065A (en) * | 2021-06-23 | 2021-08-27 | 北京奇艺世纪科技有限公司 | Video processing method and device, electronic equipment and readable storage medium |
CN114090802A (en) * | 2022-01-13 | 2022-02-25 | 深圳市猿人创新科技有限公司 | Data storage and search method, device and equipment based on embedded equipment |
CN115442656A (en) * | 2021-06-04 | 2022-12-06 | ***通信集团浙江有限公司 | Method, device, equipment and storage medium for automatically detecting video titles and video trailers |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112001292B (en) * | 2020-08-18 | 2024-01-09 | 大连海事大学 | Finger vein indexing method based on multi-scale attention mechanism depth hash |
CN113704532B (en) * | 2020-11-25 | 2024-04-26 | 天翼数字生活科技有限公司 | Method and system for improving picture retrieval recall rate |
CN114567798B (en) * | 2022-02-28 | 2023-12-12 | 南京烽火星空通信发展有限公司 | Tracing method for short video variety of Internet |
CN114842371B (en) * | 2022-03-30 | 2024-02-27 | 西北工业大学 | Unsupervised video anomaly detection method |
CN115017366B (en) * | 2022-07-11 | 2024-04-02 | 中国科学技术大学 | Unsupervised video hash retrieval method based on multi-granularity contextualization and multi-structure preservation |
CN117391150B (en) * | 2023-12-07 | 2024-03-12 | 之江实验室 | Graph data retrieval model training method based on hierarchical pooling graph hash |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105930402A (en) * | 2016-04-15 | 2016-09-07 | 乐视控股(北京)有限公司 | Convolutional neural network based video retrieval method and system |
CN107766447A (en) * | 2017-09-25 | 2018-03-06 | 浙江大学 | It is a kind of to solve the method for video question and answer using multilayer notice network mechanism |
WO2018086513A1 (en) * | 2016-11-08 | 2018-05-17 | 杭州海康威视数字技术股份有限公司 | Target detection method and device |
US20180329884A1 (en) * | 2017-05-12 | 2018-11-15 | Rsvp Technologies Inc. | Neural contextual conversation learning |
CN108985192A (en) * | 2018-06-29 | 2018-12-11 | 东南大学 | A kind of video smoke recognition methods based on multitask depth convolutional neural networks |
CN109087337A (en) * | 2018-11-07 | 2018-12-25 | 山东大学 | Long-time method for tracking target and system based on layering convolution feature |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107943938A (en) * | 2017-11-23 | 2018-04-20 | 清华大学 | A kind of large-scale image similar to search method and system quantified based on depth product |
CN108280233A (en) * | 2018-02-26 | 2018-07-13 | 南京邮电大学 | A kind of VideoGIS data retrieval method based on deep learning |
CN109086690B (en) * | 2018-07-13 | 2021-06-22 | 北京旷视科技有限公司 | Image feature extraction method, target identification method and corresponding device |
CN109815364B (en) * | 2019-01-18 | 2020-01-14 | 上海极链网络科技有限公司 | Method and system for extracting, storing and retrieving mass video features |
-
2019
- 2019-01-18 CN CN201910047518.3A patent/CN109815364B/en not_active Expired - Fee Related
-
2020
- 2020-01-19 WO PCT/CN2020/072969 patent/WO2020147857A1/en active Application Filing
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105930402A (en) * | 2016-04-15 | 2016-09-07 | 乐视控股(北京)有限公司 | Convolutional neural network based video retrieval method and system |
WO2018086513A1 (en) * | 2016-11-08 | 2018-05-17 | 杭州海康威视数字技术股份有限公司 | Target detection method and device |
US20180329884A1 (en) * | 2017-05-12 | 2018-11-15 | Rsvp Technologies Inc. | Neural contextual conversation learning |
CN107766447A (en) * | 2017-09-25 | 2018-03-06 | 浙江大学 | It is a kind of to solve the method for video question and answer using multilayer notice network mechanism |
CN108985192A (en) * | 2018-06-29 | 2018-12-11 | 东南大学 | A kind of video smoke recognition methods based on multitask depth convolutional neural networks |
CN109087337A (en) * | 2018-11-07 | 2018-12-25 | 山东大学 | Long-time method for tracking target and system based on layering convolution feature |
Non-Patent Citations (1)
Title |
---|
SUTASINEE CHIMLEK: "Landmark Image Searching with Inattentive Salient Regions", 《IEEE》 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020147857A1 (en) * | 2019-01-18 | 2020-07-23 | 上海极链网络科技有限公司 | Method and system for extracting, storing and retrieving mass video features |
CN110489685A (en) * | 2019-08-19 | 2019-11-22 | 腾讯科技(武汉)有限公司 | A kind of Webpage display process, system and relevant apparatus and storage medium |
CN110738128A (en) * | 2019-09-19 | 2020-01-31 | 天津大学 | repeated video detection method based on deep learning |
CN110688524A (en) * | 2019-09-24 | 2020-01-14 | 深圳市网心科技有限公司 | Video retrieval method and device, electronic equipment and storage medium |
CN110688524B (en) * | 2019-09-24 | 2023-04-14 | 深圳市网心科技有限公司 | Video retrieval method and device, electronic equipment and storage medium |
CN110769276A (en) * | 2019-11-07 | 2020-02-07 | 成都国腾实业集团有限公司 | Frame extraction slice video detection method based on MD5 |
CN110769291B (en) * | 2019-11-18 | 2022-08-30 | 上海极链网络科技有限公司 | Video processing method and device, electronic equipment and storage medium |
CN110769291A (en) * | 2019-11-18 | 2020-02-07 | 上海极链网络科技有限公司 | Video processing method and device, electronic equipment and storage medium |
CN111444390A (en) * | 2020-04-02 | 2020-07-24 | 徐州工程学院 | Spark and depth hash based video parallel retrieval method |
CN112035701A (en) * | 2020-08-11 | 2020-12-04 | 南京烽火星空通信发展有限公司 | Internet short video source tracing method and system |
CN113297899A (en) * | 2021-03-23 | 2021-08-24 | 上海理工大学 | Video hash algorithm based on deep learning |
CN113297899B (en) * | 2021-03-23 | 2023-02-03 | 上海理工大学 | Video hash algorithm based on deep learning |
CN115442656A (en) * | 2021-06-04 | 2022-12-06 | ***通信集团浙江有限公司 | Method, device, equipment and storage medium for automatically detecting video titles and video trailers |
CN115442656B (en) * | 2021-06-04 | 2023-08-15 | ***通信集团浙江有限公司 | Video head and tail automatic detection method, device, equipment and storage medium |
CN113313065A (en) * | 2021-06-23 | 2021-08-27 | 北京奇艺世纪科技有限公司 | Video processing method and device, electronic equipment and readable storage medium |
CN114090802A (en) * | 2022-01-13 | 2022-02-25 | 深圳市猿人创新科技有限公司 | Data storage and search method, device and equipment based on embedded equipment |
Also Published As
Publication number | Publication date |
---|---|
WO2020147857A1 (en) | 2020-07-23 |
CN109815364B (en) | 2020-01-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109815364A (en) | A kind of massive video feature extraction, storage and search method and system | |
CN109948425B (en) | Pedestrian searching method and device for structure-aware self-attention and online instance aggregation matching | |
CN109857889B (en) | Image retrieval method, device and equipment and readable storage medium | |
Chaudhuri et al. | Multilabel remote sensing image retrieval using a semisupervised graph-theoretic method | |
CN107885764B (en) | Rapid Hash vehicle retrieval method based on multitask deep learning | |
US10635949B2 (en) | Latent embeddings for word images and their semantics | |
Karami et al. | Automatic plant counting and location based on a few-shot learning technique | |
US9251434B2 (en) | Techniques for spatial semantic attribute matching for location identification | |
US9165217B2 (en) | Techniques for ground-level photo geolocation using digital elevation | |
CN110956185A (en) | Method for detecting image salient object | |
CN109829467A (en) | Image labeling method, electronic device and non-transient computer-readable storage medium | |
CN103324677B (en) | Hierarchical fast image global positioning system (GPS) position estimation method | |
CN110807434A (en) | Pedestrian re-identification system and method based on combination of human body analysis and coarse and fine particle sizes | |
Sumbul et al. | Informative and representative triplet selection for multilabel remote sensing image retrieval | |
CN108960331A (en) | A kind of recognition methods again of the pedestrian based on pedestrian image feature clustering | |
Thompson et al. | finFindR: Automated recognition and identification of marine mammal dorsal fins using residual convolutional neural networks | |
CN105183857A (en) | Automatic picture training sample extracting method and system | |
CN113435329B (en) | Unsupervised pedestrian re-identification method based on video track feature association learning | |
Matzen et al. | Bubblenet: Foveated imaging for visual discovery | |
CN116935411A (en) | Radical-level ancient character recognition method based on character decomposition and reconstruction | |
CN116524263A (en) | Semi-automatic labeling method for fine-grained images | |
Hezel et al. | Video search with sub-image keyword transfer using existing image archives | |
CN117280338A (en) | Fine granularity visual content search platform | |
WO2015083170A1 (en) | Fine grained recognition method and system | |
CN113408356A (en) | Pedestrian re-identification method, device and equipment based on deep learning 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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200114 |
|
CF01 | Termination of patent right due to non-payment of annual fee |