CN108021927A - A kind of method for extracting video fingerprints based on slow change visual signature - Google Patents
A kind of method for extracting video fingerprints based on slow change visual signature Download PDFInfo
- Publication number
- CN108021927A CN108021927A CN201711087291.2A CN201711087291A CN108021927A CN 108021927 A CN108021927 A CN 108021927A CN 201711087291 A CN201711087291 A CN 201711087291A CN 108021927 A CN108021927 A CN 108021927A
- Authority
- CN
- China
- Prior art keywords
- video
- network
- lstm
- space characteristics
- training
- 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
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F16/78—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/783—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
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Library & Information Science (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the slow method for extracting video fingerprints for becoming visual signature, comprise the following steps:To the image of each image generation random distortion version of training set, space characteristics are trained with original image and distorted image, extract the parameter of network;The normalization that space and sequential are carried out to training set video data pre-processes, and handles and fixes frame number for fixed dimension;Training data, training LSTM networks are used as by the use of the characteristic sequence of space characteristics extraction network;Network is extracted from trained space characteristics, by obtained feature and LSTM cascades, extracts video finger print.This method simulates the visual perception principle of people, can extract the fingerprint of video, has the advantages that robust, efficient.Available for fields such as video copy detection, video frequency searchings.
Description
Technical field
The present invention relates to video copy detection field, more particularly to a kind of video finger print extraction based on slow change visual signature
Method.
Background technology
With the development of video sharing website and mobile Internet, the video data on network sharply increases, and thus brings
The problems such as infringement of copyright is propagated with illegal contents.Since data volume is huge, manpower search can not be relied on to illegally copy video.For
Solve the problems, such as this, some video copying detection methods are proposed successively in recent years.Video copy detection is in known source video
On the premise of, it is searched for from mass data and copies version.Video fingerprinting algorithms are the key technologies of copy detection, it is by video
Main contents be described as a brief synopsis, similar to mankind's fingerprint.Video copy detection technology can pass through ratio
Distinguish whether two sections of videos are homologous compared with video finger print.
The concept of video finger print is by Stanford University Indyk etc.[1]Itd is proposed in 1999.In recent years to video copy detection
The research of video fingerprinting algorithms of the demand driving of technology, and feature extraction is to design the key issue of video fingerprinting algorithms, directly
Connect the robustness and distinction for determining fingerprint.By feature extraction model split, video fingerprinting algorithms can be divided mainly into two classes:Frame
The space fingerprint cascade of rank and space-time joint fingerprint.
First kind method independently calculates the fingerprint of each frame of video, and then cascade obtains the fingerprint of whole section of video.Roover
Deng[2]The radial direction hash algorithm (Rash, Radial hashing) of proposition is the representative of this kind of method, and Rash algorithms are along per frame footpath
To direction calculating pixel Valued Statistics as fingerprint.There is robustness, Lee etc. using the center of gravity of gradient direction[3]It is proposed
Algorithm referred to the center of gravity (CGO, Centroid of Gradient Orientations) of gradient direction as feature calculation
Line.In addition, some algorithms are with area-of-interest[4]And sparse features[5]Based on design video fingerprinting algorithms.
The time-space relationship of second class method combination video data calculates video finger print.Such as Li et al.[6]The structure chart of proposition
Model (SGM, Structural Graphical Models) generates video finger print by the efficient dimensionality reduction in time domain and space.
Nie etc.[7]It is tensor by representation of video shot, it is proposed that a kind of video fingerprinting algorithms based on tensor resolution.Esmaeili etc.[8]It is proposed
Based on regarding for Three-dimensional DCT (3D-DCT, Three-Dimensional Discrete Cosine Transform)
Frequency fingerprint algorithm, it by the key visual information of video is brief fingerprint to concentrate characteristic using the energy of dct transform.
In the implementation of the present invention, discovery at least has the following disadvantages in the prior art and deficiency by inventor:
Most of conventional video fingerprint algorithm depends on the feature extracting method of hand-designed, is typically only capable to portray video
One side visual signature.Since the information expressed by image and video data is sufficiently complex, the model of hand-designed is difficult comprehensive
These information are portrayed, particularly some abstract characteristics.
The content of the invention
It is of the invention from video learning the present invention provides a kind of based on the slow method for extracting video fingerprints for becoming visual signature
Video finger print, introduces slow change analysis principle, comprising the sequential relevant information in sequence of frames of video, while has good robust
Property, distinction and high efficiency, can be applied to the fields such as video frequency searching or copy detection accordingly, described below:
It is a kind of that following step is included based on the slow method for extracting video fingerprints for becoming visual signature, the method for extracting video fingerprints
Suddenly:
It is special with original image and distorted image training space to the image of each image generation random distortion version of training set
Sign, extracts the parameter of network;
The normalization that space and sequential are carried out to training set video data pre-processes, and handles and fixes frame number for fixed dimension;
Training data, training LSTM networks are used as by the use of the characteristic sequence of space characteristics extraction network;
Network is extracted from trained space characteristics, by obtained feature and LSTM cascades, extracts video finger print.
Wherein, it is described to be specially with original image and distorted image training space characteristics, the step of parameter for extracting network:
Space characteristics extraction network structure is made of full Connection Neural Network;
Using the different image of content as inhomogeneity, original image and its distorted version are as similar;Training space characteristics carry
Take network make in class away from it is small, class spacing is big.
Wherein, the normalization pretreatment that space and sequential are carried out to training set video data, is handled as fixed dimension
The step of fixed frame number is specially:
Each frame of video data is normalized, and is fixed size by size change over, by sampling each video
Frame number is changed into fixed frame number.
Further, the characteristic sequence by the use of space characteristics extraction network trains LSTM networks as training data
Step is specially:
Each frame of video after each pretreatment is inputted into trained space characteristics extraction network, is obtained and frame number
The feature of equal number;
The output of LSTM is set to reconstruct input frame, cost function includes two:Reconstructed error, the LSTM every adjacent two of LSTM
The difference of a hidden layer state;The network parameter of LSTM is trained by optimizing cost function.
Further, it is described to extract network from trained space characteristics, by obtained feature and LSTM cascades,
Extracting video finger print is specially:
Each input video input space feature extraction network after pretreatment, extracts the feature of each frame;
Obtained characteristic sequence is inputted in LSTM to the video finger print for extracting the video.
The beneficial effect of technical solution provided by the invention is:
1st, by the method for neural network learning video finger print, avoid traditional artificial algorithm for design and portraying video features
The limitation of aspect;
2nd, deep-neural-network is combined the space-time characteristic of extraction video with LSTM (long memory network in short-term), can be at the same time
Ensure robustness of the video finger print to spatial domain distortion (such as filter plus make an uproar) and sequential distortion (such as frame loss);
3rd, present procedure is simple, it is easy to accomplish, the fingerprint learning process computation complexity is low.
Brief description of the drawings
Fig. 1 is a kind of flow chart based on the slow method for extracting video fingerprints for becoming visual signature;
Fig. 2 is the flow chart that video finger print is extracted using space-time deep neural network;
Fig. 3 is the schematic diagram that space characteristics extract network structure;
Fig. 4 is LSTM structure diagrams.
Embodiment
To make the object, technical solutions and advantages of the present invention clearer, embodiment of the present invention is made below further
It is described in detail on ground.
Embodiment 1
In order to realize the robust of video finger print and efficient learning method, the embodiment of the present invention proposes a kind of based on slow change
The method for extracting video fingerprints of visual signature, it is described below referring to Fig. 1 and Fig. 2:
101:To the image of each image generation random distortion version of training set, trained with original image and distorted image empty
Between feature, extract the parameter of network;
Wherein, which is specially:
1) each image is added into random distortion so as to obtain corresponding distorted image;
It can be set for the type of distortion according to being actually needed, the embodiment of the present invention is not limited.
2) all images are normalized to n × n (i.e. including original image and distorted image), average 0, variance 1, and
As similar, space is trained by optimizing following cost function as inhomogeneity, each image and its distorted version for different images
Feature, extracts the parameter of network.
J=E d | l=+1 }-E d | l=-1 } (1)
In formula, E d | and l=+1 } represent that similar image (classification l=+1) obtains the phase of the distance between feature d by network
Hope, i.e., class it is interior away from.E d | and l=-1 } represent that inhomogeneity image (classification l=-1) obtains the distance between feature d's by network
It is expected, i.e. class spacing.
Pass through above-mentioned training so that the characteristic distance between similar image diminishes, the characteristic distance between inhomogeneity image
Become larger.
The embodiment of the present invention is not limited the concrete structure and training method of space characteristics extraction network.
102:The normalization that space and sequential are carried out to training set video data pre-processes, and handles as fixed dimension anchor-frame
Number;
Wherein, which is specially:
Each frame size of video data is transformed to n × n, and it is 0 that will be normalized to average, variance 1.Pass through sampling
The frame number of each video is changed into fixed frame number m.For n, the value of m, and the method for sampling can be according to the need of practical application
Set, the embodiment of the present invention is not limited.
103:Training data, training LSTM networks are used as by the use of the characteristic sequence of space characteristics extraction network;
Wherein, which is specially:
1) each frame input space feature extraction network in pretreated video, each video are obtained into m feature
Sequence (x1,x2...xm)。
LSTM is trained using the characteristic sequence of generation in formula (1), to optimize following cost function as target, training LSTM
Network parameter:
In formula, xtRepresent (x1,x2...xm) in t-th of feature;htRepresent xtHidden layer state during by LSTM units;
ytRepresent t-th of output, ytBy htIt is calculated:WyAnd byRespectively the weight matrix of LSTM networks and
Bias vector,For activation primitive;λ is constant;||·||2Represent two norms.In cost functionFor it is adjacent when
The similarity between hidden layer state is carved, by minimizing this LSTM can be made to learn the slow change vision on to video along time orientation
Feature, helps to improve the robustness of video finger print.
104:Network is extracted from trained space characteristics, by obtained feature and LSTM cascades, extracts video
Fingerprint.
Wherein, which is specially:
Video input space feature extraction network after pretreatment, extracts the feature of each frame, obtains characteristic sequence
(x1,x2...xm).Then by (x1,x2...xm) in each feature sequentially input in LSTM, last obtained hidden layer shape
State hmAs video finger print.
Refer in conclusion the embodiment of the present invention is realized from video learning video by above-mentioned steps 101- steps 104
Line, the video finger print combine the space characteristics and temporal aspect of video, while have good robustness and distinction, accordingly
It can be applied to the fields such as video frequency searching or copy detection.
Embodiment 2
Below by taking a hands-on process as an example, the scheme in embodiment 1 is carried out with reference to specific calculation formula detailed
Video finger print learning method thin to introduce, that embodiment that the present invention will be described in detail provides, it is described below:
201:Image preprocessing;
Training data of 15000 images as space characteristics extraction network is randomly selected from ImageNet, by image
Standard size 512 × 512 is normalized to, and passes through mean filter.
202:Generate the image of random distortion version;
To every image application a kind of random distortion or conversion, the type of these distortions or conversion includes:JPEG is damaged
Compression, Gaussian noise, rotation, medium filtering, histogram equalization, gamma correction, addition speckle noise, circulation filtering.Altogether
30000 pictures.
203:Training space characteristics extraction network;
The space characteristics extraction network used in the present embodiment is three layers of fully-connected network, is 1024- per node layer number
800-400.Every figure (32 × 32) is transformed to the vector that length is 1024, the vector I of original image generationiRepresent, IiIt is right
The vector use for the distorted image generation answeredRepresent.IiWithThe feature obtained by space characteristics extraction network is respectively f (Ii)
WithThis example is using gradient descent method training space characteristics extraction network, wherein each batch size is 1000, every time
Randomly select 1000 original graph (I1,I2...I1000) and corresponding distortion mapIt is excellent as training data
Change following cost function:
204:The normalization that space and sequential are carried out to training set video data pre-processes, and pre-processes and is fixed for fixed dimension
Frame number;
Wherein, the embodiment of the present invention uses training number of 400 video sequences downloaded from YouTube as LSTM
According to.Space and sequential normalization pretreatment are carried out to video data, the video size after processing is fixed as:32 × 32 × 10, often
Frame pixel average after normalization is 0, variance 1.
205:With the characteristic sequence of space characteristics extraction video;
Video features sequence is extracted in 400 video datas after being handled from step 204:By each frame of each video
(32 × 32) input step 203 obtains network, therefore obtains the characteristic sequence that 400 sizes are 400 × 10.
206:Training LSTM networks;
The input node number of LSTM is 400 defined in the present embodiment, the number of hidden nodes 128.By the spy of each video
Sign sequence is expressed as (x1,x2...x10), t-th of feature xtIt is h to be input to the hidden layer state obtained after LSTM networkst, define yt
=sig (Wyht+by), WyAnd byRespectively weight matrix and bias vector, sig () are sigmoid functions, ytLength be
400.Weight selection attenuation coefficient λ=0.5, optimizes following cost function and obtains the network parameter of LSTM.
This example chooses stochastic gradient descent method training LSTM networks.
207:Extract the video finger print of video.
By any one input video after the pretreatment of step 204,32 × 32 × 20 video data is obtained.By this
The space characteristics extraction network that the training of 20 two field picture input steps 203 obtains, obtains the characteristic sequence that size is 400 × 10.Will
The LSTM networks obtained in this 10 video sequence input steps 206, and by last hidden layer state h20As video finger print,
The length of 128.
In conclusion the embodiment of the present invention combines the space of video from video learning video finger print, the video finger print
Feature and temporal aspect;And introduce it is slow become analysis principle, make video finger print include video change in time it is slow important
Information, while there is good robustness and distinction, it can be applied to the fields such as video frequency searching or copy detection accordingly.
Embodiment 3
With reference to specific experimental data, feasibility verification is carried out to the scheme in Examples 1 and 2, it is as detailed below to retouch
State:
In above-described embodiment 2, choose 600 video sequences downloaded from YouTube and other 201 come from
Totally 801 videos are different and non-overlapping with training data two-by-two between video as test video for the video sequence of TRECVID.
Every section of video applies 9 kinds of common contents and keeps distortion, and different degree is chosen in every kind of distortion, as shown in the table:
1 type of distortion of table and parameter setting
Each original video produces 17 copy versions after these distortion processings, the original video in test library with copy
The total number of shellfish is 14418.Each video is normalized to obtain 32 × 32 × 20 using the method for step 204, is classified as two
A 32 × 32 × 10 video sequence simultaneously obtains the video features of two a length of 128 using the method for step 207, connects the two
Video features obtain a length of 256 video finger print.According to video finger print, the original version of every distorted picture of inquiry, statistics is looked into
Whether correct ask result.
According to accuracy rate P:The correct positive example data of query result account for the ratio for being predicted as positive example data;
Recall rate R:Query result be the data of positive example account for be actually positive example data ratio, calculate F1Index:
F1=2/ (1/P+1/R)=2 × P × R/P+R (5)
As a result 0.995, close to ideal value 1.The video finger print for understanding to learn has very strong robustness.
In conclusion the embodiment of the present invention combines the space of video from video learning video finger print, the video finger print
Feature and temporal aspect, make video finger print include video along temporal slow change visual signature, and demonstrate in test its Shandong
Rod and distinction, can be applied to the fields such as video frequency searching or copy detection accordingly.
Bibliography
[1]Indyk P,Iyengar G,Shivakumar N.Finding pirated video sequences on
the internet[R].Technical report,Stanford University,1999.
[2]De Roover C,De Vleeschouwer C,Lefebvre F,et al.Robust video
hashing based on radial projections of key frames[J].IEEE Transactions on
Signal processing,2005,53(10):4020-4037.
[3]Lee S,Yoo C D.Robust video fingerprinting for content-based video
identification[J].IEEE Transactions on Circuits and Systems for Video
Technology,2008,18(7):983-988.
[4]Li J,Guo X,Yu Y,et al.A robust and low-complexity video
fingerprint for multimedia security[C]//Wireless Personal Multimedia
Communications(WPMC),2014International Symposium on.IEEE,2014:97-102.
[5]Wu B,Krishnan S S,Zhang N,et al.Compact and robust video
fingerprinting usingsparse represented features[C]//Multimedia and Expo
(ICME),2016IEEE International Conference on.IEEE,2016:1-6.
[6]Li M,Monga V.Compact video fingerprinting via structural graphical
models[J].IEEE transactions on information forensics and security,2013,8(11):
1709-1721.
[7]Nie X,Yin Y,Sun J,et al.Comprehensive feature-based robust video
fingerprinting using tensor model[J].IEEE Transactions on Multimedia,2017,19
(4):785-796.
[8]Esmaeili M M,Fatourechi M,Ward R K.A robust and fast video copy
detection system using content-based fingerprinting[J].IEEE Transactions on
information forensics and security,2011,6(1):213-226.
It will be appreciated by those skilled in the art that attached drawing is the schematic diagram of a preferred embodiment, the embodiments of the present invention
Sequence number is for illustration only, does not represent the quality of embodiment.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all the present invention spirit and
Within principle, any modification, equivalent replacement, improvement and so on, should all be included in the protection scope of the present invention.
Claims (5)
- It is 1. a kind of based on the slow method for extracting video fingerprints for becoming visual signature, it is characterised in that the method for extracting video fingerprints Comprise the following steps:To the image of each image generation random distortion version of training set, space characteristics are trained with original image and distorted image, Extract the parameter of network;The normalization that space and sequential are carried out to training set video data pre-processes, and handles and fixes frame number for fixed dimension;Training data, training LSTM networks are used as by the use of the characteristic sequence of space characteristics extraction network;Network is extracted from trained space characteristics, by obtained feature and LSTM cascades, extracts video finger print.
- It is 2. according to claim 1 a kind of based on the slow method for extracting video fingerprints for becoming visual signature, it is characterised in that institute State and be specially with original image and distorted image training space characteristics, the step of parameter for extracting network:Space characteristics extraction network structure is made of full Connection Neural Network;Using the different image of content as inhomogeneity, original image and its distorted version are as similar;Training space characteristics extraction net Network make in class away from it is small, class spacing is big.
- It is 3. according to claim 1 a kind of based on the slow method for extracting video fingerprints for becoming visual signature, it is characterised in that institute The normalization pretreatment that space and sequential are carried out to training set video data is stated, the step of fixing frame number for fixed dimension is handled and has Body is:Each frame of video data is normalized, and is fixed size by size change over, by sampling the frame number of each video It is changed into fixed frame number.
- It is 4. according to claim 1 a kind of based on the slow method for extracting video fingerprints for becoming visual signature, it is characterised in that institute The step of characteristic sequence stated by the use of space characteristics extraction network is used as training data, trained LSTM networks be specially:Each frame of video after each pretreatment is inputted into trained space characteristics extraction network, is obtained equal with frame number The feature of number;The output of LSTM is set to reconstruct input frame, cost function includes two:Reconstructed error, the LSTM of LSTM is every two neighboring hidden The difference of layer state;The network parameter of LSTM is trained by optimizing cost function.
- It is 5. according to claim 1 a kind of based on the slow method for extracting video fingerprints for becoming visual signature, it is characterised in that institute State and extract network from trained space characteristics, by obtained feature and LSTM cascades, extraction video finger print is specially:Each input video input space feature extraction network after pretreatment, extracts the feature of each frame;Obtained characteristic sequence is inputted in LSTM to the video finger print for extracting the video.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711087291.2A CN108021927A (en) | 2017-11-07 | 2017-11-07 | A kind of method for extracting video fingerprints based on slow change visual signature |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711087291.2A CN108021927A (en) | 2017-11-07 | 2017-11-07 | A kind of method for extracting video fingerprints based on slow change visual signature |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108021927A true CN108021927A (en) | 2018-05-11 |
Family
ID=62080607
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711087291.2A Pending CN108021927A (en) | 2017-11-07 | 2017-11-07 | A kind of method for extracting video fingerprints based on slow change visual signature |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108021927A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110826363A (en) * | 2018-08-09 | 2020-02-21 | 阿里巴巴集团控股有限公司 | Picture fingerprint generation method and device |
CN111143619A (en) * | 2019-12-27 | 2020-05-12 | 咪咕文化科技有限公司 | Video fingerprint generation method, video fingerprint retrieval method, electronic device and medium |
CN111860053A (en) * | 2019-04-28 | 2020-10-30 | 北京灵汐科技有限公司 | Multimedia data identification method and device |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104077590A (en) * | 2014-06-30 | 2014-10-01 | 安科智慧城市技术(中国)有限公司 | Video fingerprint extraction method and system |
CN105279489A (en) * | 2015-10-13 | 2016-01-27 | 成都纽捷那科技有限公司 | Video fingerprint extraction method based on sparse coding |
CN106778571A (en) * | 2016-12-05 | 2017-05-31 | 天津大学 | A kind of digital video feature extracting method based on deep neural network |
CN106886768A (en) * | 2017-03-02 | 2017-06-23 | 杭州当虹科技有限公司 | A kind of video fingerprinting algorithms based on deep learning |
-
2017
- 2017-11-07 CN CN201711087291.2A patent/CN108021927A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104077590A (en) * | 2014-06-30 | 2014-10-01 | 安科智慧城市技术(中国)有限公司 | Video fingerprint extraction method and system |
CN105279489A (en) * | 2015-10-13 | 2016-01-27 | 成都纽捷那科技有限公司 | Video fingerprint extraction method based on sparse coding |
CN106778571A (en) * | 2016-12-05 | 2017-05-31 | 天津大学 | A kind of digital video feature extracting method based on deep neural network |
CN106886768A (en) * | 2017-03-02 | 2017-06-23 | 杭州当虹科技有限公司 | A kind of video fingerprinting algorithms based on deep learning |
Non-Patent Citations (1)
Title |
---|
汪冬冬 等: "基于时空深度神经网络的视频指纹算法", 《HTTP://KNS.CNKI.NET/KCMS/DETAIL/31.1690.TN.20170825.1539.018.HTML》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110826363A (en) * | 2018-08-09 | 2020-02-21 | 阿里巴巴集团控股有限公司 | Picture fingerprint generation method and device |
CN111860053A (en) * | 2019-04-28 | 2020-10-30 | 北京灵汐科技有限公司 | Multimedia data identification method and device |
WO2020220926A1 (en) * | 2019-04-28 | 2020-11-05 | 北京灵汐科技有限公司 | Multimedia data identification method and device |
CN111860053B (en) * | 2019-04-28 | 2023-11-24 | 北京灵汐科技有限公司 | Multimedia data identification method and device |
CN111143619A (en) * | 2019-12-27 | 2020-05-12 | 咪咕文化科技有限公司 | Video fingerprint generation method, video fingerprint retrieval method, electronic device and medium |
CN111143619B (en) * | 2019-12-27 | 2023-08-15 | 咪咕文化科技有限公司 | Video fingerprint generation method, search method, electronic device and medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Jing et al. | Hinet: Deep image hiding by invertible network | |
Guan et al. | DeepMIH: Deep invertible network for multiple image hiding | |
Walia et al. | Digital image forgery detection: a systematic scrutiny | |
Emam et al. | PCET based copy-move forgery detection in images under geometric transforms | |
Sun et al. | Video hashing based on appearance and attention features fusion via DBN | |
Bilal et al. | A robust technique for copy-move forgery detection from small and extremely smooth tampered regions based on the DHE-SURF features and mDBSCAN clustering | |
Liu et al. | A robust coverless steganography scheme using camouflage image | |
Ustubioglu et al. | A new copy move forgery detection technique with automatic threshold determination | |
CN106778571B (en) | Digital video feature extraction method based on deep neural network | |
Garg et al. | Neural style transfer for image steganography and destylization with supervised image to image translation | |
Ulutas et al. | A new copy move forgery detection method resistant to object removal with uniform background forgery | |
Singh et al. | SiteForge: Detecting and localizing forged images on microblogging platforms using deep convolutional neural network | |
Singh et al. | Comprehensive review of watermarking techniques in deep-learning environments | |
Yao et al. | Detecting copy-move forgery using non-negative matrix factorization | |
CN108021927A (en) | A kind of method for extracting video fingerprints based on slow change visual signature | |
Gao et al. | A robust video zero-watermarking based on deep convolutional neural network and self-organizing map in polar complex exponential transform domain | |
Liu et al. | Overview of image inpainting and forensic technology | |
Liu et al. | Image deblocking detection based on a convolutional neural network | |
Rani et al. | A robust watermarking scheme exploiting balanced neural tree for rightful ownership protection | |
Cao et al. | Screen-shooting resistant image watermarking based on lightweight neural network in frequency domain | |
CN116257648A (en) | Image retrieval defense method and system based on noise network and image preprocessing | |
Sun et al. | Deep perceptual hash based on hash center for image copyright protection | |
CN116383470A (en) | Image searching method with privacy protection function | |
Le-Tien et al. | A combination of super-resolution and deep learning approaches applied to image forgery detection | |
Zhu et al. | Recaptured image detection through enhanced residual-based correlation coefficients |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180511 |
|
WD01 | Invention patent application deemed withdrawn after publication |