CN109784280A - Human bodys' response method based on Bi-LSTM-Attention model - Google Patents
Human bodys' response method based on Bi-LSTM-Attention model Download PDFInfo
- Publication number
- CN109784280A CN109784280A CN201910048015.8A CN201910048015A CN109784280A CN 109784280 A CN109784280 A CN 109784280A CN 201910048015 A CN201910048015 A CN 201910048015A CN 109784280 A CN109784280 A CN 109784280A
- Authority
- CN
- China
- Prior art keywords
- lstm
- vector
- model
- network
- attention
- 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
Abstract
The present invention provides a kind of Human bodys' response method based on Bi-LSTM-Attention model, the following steps are included: step S1, the video frame of extraction is inputted into InceptionV3 model, network parameter is reduced while increasing convolutional neural networks depth using InceptionV3 model, the depth characteristic for sufficiently extracting video frame, obtains relevant feature vector;The obtained feature vector of step S1 is passed in Bi-LSTM neural network and handles by step S2, sufficiently learns the temporal aspect between video frame by Bi-LSTM neural network;The temporal aspect vector that step S2 is obtained is passed to attention Mechanism Model and adaptively perceives the network weight for having larger impact to recognition result by step S3, and the relevant feature of these network weights is more paid close attention to.The present invention can be improved the discrimination of human body behavior.
Description
Technical field
The present invention relates to video analysis and identification field, especially a kind of people based on Bi-LSTM-Attention model
Body Activity recognition method.
Background technique
For Human bodys' response, most of early stage is to extract video features using the method for engineer.A kind of scheme
The characteristics of human body under complex background is extracted using the method for space-time interest points, this method is by calculating each position in video sequence
Power and space-time interest points are found by the method for very big value filtering.WANG W et al. is learnt quiet using the method for sparse coding
State feature, and histogram is carried out to feature with the time domain pyramid structure based on maximum pond indicate, finally divided using SVM
Class.Another scheme proposes a kind of hierarchical cluster multi-task learning (HC-MTL) method, reinforces shared row by objective function
Human bodys' response is realized with specific behavioural characteristic is learnt for relationship.Method based on manual features extraction is in Activity recognition
Aspect achieves many excellent achievements, however there is also some insoluble problems, the method for engineer tends not to
The substantive characteristics of movement is given expression to, and due to the diversity of movement, is often easy to ignore some important features, for row
There is large effect for identification.
JI S et al. has been put forward for the first time a kind of 3D CNN algorithm, and this method is by using 3D volumes to the video frame on time shaft
Product core is used to identify human body behavior to capture the room and time information of video.B.Mahasseni et al. is tieed up by construction human body 3
Then skeleton is used to identify human body behavior using the timing information that LSTM study human body 3 ties up skeleton.CNN net is utilized in Ullah A
Network extracts the further feature of video frame, and carries out the timing information in learning characteristic sequence by two-way LSTM, finally by
Softmax classifier is classified.J.Donahue et al. proposes a kind of long-term cyclic convolution network, and the network is from 2D CNN
Middle extraction feature simultaneously learns the ordinal relation between these features by LSTM network.The CNN and LSTM in Activity recognition
Using greatly improving the precision of identification, and reduce workload.But the depth of CNN has the feature extraction of video frame
Large effect: the low depth characteristic for being not easy to show image of network layer is easy poor fitting;Profound network model is easy
Gradient dispersion is generated to be difficult to optimize network mould.LSTM can not effectively learn the temporal aspect of movement, lack autonomous adaptability.
Present document relates to term:
SVM: support vector machines;
3D CNN:3D convolutional neural networks;
LSTM: long memory network in short-term.
Attention: attention.
Summary of the invention
It is an object of the present invention to overcome the shortcomings of the prior art and provide one kind to be based on Bi-LSTM-
The Human bodys' response method of Attention model, this method can be with the timing information in learning characteristic sequence, and passes through attention
Power mechanism trains network weight, reaches better performance, reduces identification error.
The technical solution adopted by the present invention is that:
A kind of Human bodys' response method based on Bi-LSTM-Attention model, comprising the following steps:
Step S1, inputs InceptionV3 model for the video frame of extraction, increases convolution using InceptionV3 model
Network parameter is reduced while neural network depth, the depth characteristic of video frame is sufficiently extracted, obtains relevant feature vector;
The obtained feature vector of step S1 is passed in Bi-LSTM neural network and handles, passes through Bi- by step S2
LSTM neural network sufficiently learns the temporal aspect between video frame;
The temporal aspect vector that step S2 is obtained is passed to attention Mechanism Model and adaptively perceived to knowledge by step S3
Other result has the network weight of larger impact, and the relevant feature of these network weights is more paid close attention to.
Further, in step S1, different convolutional layers is incorporated in one by way of in parallel by InceptionV3 model
It rises, while convolution operation is carried out to video frame using various sizes of convolution kernel, finally by filter fused layer different volumes
The feature vector of product core processing is stitched together, and exports depth characteristic matrix by full articulamentum and is used for transmission Bi-LSTM nerve
In network.
Further, step S2 is specifically included:
wi(i=1 ... 6) indicates a layer network layer to the weight of another network layer;{…ht-1,ht,ht+1... indicate LSTM mind
Through the propagated forward layer in network, the input of propagated forward layer be ... xt-1,xt,xt+1... characteristic sequence from front to back;
{…ht+1',ht',ht-1' ... indicating back-propagating layer in LSTM neural network, the input of back-propagating layer is
{…xt+1,xt,xt-1... characteristic sequence from back to front;
X thereintIndicate that extracted video frame passes through the feature obtained after InceptionV3 model extraction depth characteristic
Vector;Such as following formula:
ht=f (w1xt+w2ht-1+b1) (1)
ht'=f (w3xt+w5ht+1+b2) (2)
ot'=g (w4ht+b3) (3)
ot"=g (w6ht'+b3) (4)
ot=(ot'+ot”)/2 (5)
Above formula (1), (2), (3), the f in (4) and g represent activation primitive, b1、b2、b3、b4Represent the biasing of hidden unit
Coefficient, o', o " are the result that two LSTM units handle the feature vector of Inceptionv3 layers of output at the corresponding moment respectively;
Two feature vectors at corresponding moment are added the temporal aspect vector summed and be averaged as output.
Further, step S3 is specifically included:
otIt indicates t-th of the temporal aspect vector exported from Bi-LSTM neural network, then temporal aspect vector is passed
Enter into attention Mechanism Model, obtains initial state vector S by the hidden layer in attention Mechanism Modelt;Weight coefficient αt
Indicate initial state vector StThe shared specific gravity size in the state vector Y of final output;Each initial state vector StWith power
Weight factor alphatProduct cumulative and obtain the state vector Y of final output;Calculation formula is as follows:
et=tanh (wtst+bt) (6)
Tanh indicates that excitation function, n indicate the quantity of video frame;etIndicate the state vector S of t-th of temporal aspect vectort
The energy value determined, wtAnd btIndicate weight and biasing;By formula (7) using e as the power of truth of a matter various pieces energy value therewith
The available weight coefficient for having much influences on classification results of ratio of the cumulative sum of the energy value of preceding part, it is thus achieved that
Conversion of the original state to state of attention;Finally as formula (8) obtains the state vector Y of final output.
The present invention has the advantages that feature extraction phases of the present invention in video frame, use InceptionV3 model extraction
Feature solves the problems, such as network depth, the timing information that then Bi-LSTM neural network can sufficiently between learning characteristic, finally
Attention mechanism the performance of network model can be made more preferable.By Action Youtobe and KTH human body behavioral data collection with
The methods of existing DB-LSTM, 3D CNN are compared, the experimental results showed that algorithm discrimination proposed by the invention reaches
94.38% and 95.67%.
Detailed description of the invention
Fig. 1 is the Activity recognition block schematic illustration of the invention based on Bi-LSTM-Attention model.
Fig. 2 is the schematic diagram of InceptionV3 model of the invention.
Fig. 3 is the schematic diagram of Bi-LSTM neural network of the invention.
Fig. 4 is the schematic diagram of attention Mechanism Model of the invention.
Specific embodiment
Below with reference to specific drawings and examples, the invention will be further described.
The present invention proposes a kind of Human bodys' response method (One Human based on Bi-LSTM-Attention model
Action Recognition Algorithm Based on Bi-LSTM-Attention model);
This method extracts 20 video frames first from each video, passes through InceptionV3 model extraction video frame
Then depth characteristic constructs the feature vector in Bi-LSTM neural network forwardly and rearwardly, followed by attention
(Attention) Mechanism Model adaptively perceives the network weight for having larger impact to recognition result, makes Bi-LSTM-
Attention model can realize more accurate identification according to the context of behavior, connect finally by one layer of full articulamentum
Softmax classifier classifies to video.
This method mainly includes three big steps:
Step S1, inputs InceptionV3 model for the video frame of extraction, increases convolution using InceptionV3 model
Network parameter is reduced while neural network depth, the depth characteristic of video frame is sufficiently extracted, obtains relevant feature vector;
InceptionV3 model mainly carries out feature extraction to input video frame and these video frames is processed into Bi-
The feature vector form that LSTM neural network is able to receive directly and can handle;It is different from traditional CNN feature extracting method, it
Different convolutional layers is combined together by way of in parallel, while video frame is rolled up using various sizes of convolution kernel
Obtained feature vector, is then stitched together by product operation again;As shown in Fig. 2.
Therein 128 × 128 × 3 represent video frame size, and (128 × 128 represent pixel, and 3 represent the channel of rgb video
Number), 1 × 1,1 × n, n × 1 represent convolution kernel size, pool representativeization layer operation;Finally by Filter Concat (filtering
Device fused layer) feature vector of different convolution kernels processing is stitched together, exporting S*1024 by full articulamentum, (S represents video
Frame number, the depth characteristic matrix for 20) dimension is used for transmission in Bi-LSTM neural network herein.
The obtained feature vector of step S1 is passed in Bi-LSTM neural network and handles, passes through Bi- by step S2
LSTM neural network sufficiently learns the temporal aspect between video frame;As shown in figure 3,
Wherein wi(i=1 ... 6) indicates a layer network layer to the weight of another network layer;{…ht-1,ht,ht+1... indicate
Propagated forward layer in LSTM neural network, the input of propagated forward layer be ... xt-1,xt,xt+1... feature sequence from front to back
Column;
{…ht+1',ht',ht-1' ... indicating back-propagating layer in LSTM neural network, the input of back-propagating layer is
{…xt+1,xt,xt-1... characteristic sequence from back to front;
X thereintIndicate that extracted video frame passes through the feature obtained after InceptionV3 model extraction depth characteristic
Vector;Such as following formula:
ht=f (w1xt+w2ht-1+b1) (1)
ht'=f (w3xt+w5ht+1+b2) (2)
ot'=g (w4ht+b3) (3)
ot"=g (w6ht'+b3) (4)
ot=(ot'+ot”)/2 (5)
Above formula (1), (2), (3), the f in (4) and g represent activation primitive, b1、b2、b3、b4Represent the biasing of hidden unit
Coefficient, o', o " are the result that two LSTM units handle the feature vector of Inceptionv3 layers of output at the corresponding moment respectively;
Two feature vectors at corresponding moment are added the temporal aspect vector summed and be averaged as output, output result is one
S*1024 ties up matrix;Temporal aspect vector is finally sent to progress sensing network weight in attention Mechanism Model again;With biography
The individual event LSTM algorithm of system is compared, and Bi-LSTM algorithm due to that can learn in the past with the information in future to obtain more simultaneously
The temporal information of robust.
The temporal aspect vector that step S2 is obtained is passed to attention Mechanism Model and adaptively perceived to knowledge by step S3
Other result has the network weight of larger impact, and the relevant feature of these network weights is more paid close attention to.Such as Fig. 4
It is shown;
otIt indicates t-th of the temporal aspect vector exported from Bi-LSTM neural network, then temporal aspect vector is passed
Enter into attention Mechanism Model, obtains initial state vector S by the hidden layer in attention Mechanism Modelt;Weight coefficient αt
Indicate initial state vector StThe shared specific gravity size in the state vector Y (1024*1) of final output;Each original state
Vector StWith weight coefficient αtProduct cumulative and obtain the state vector Y of final output;Calculation formula is as follows:
et=tanh (wtst+bt) (6)
Tanh indicates that excitation function, n indicate the quantity of video frame;etIndicate the state vector S of t-th of temporal aspect vectort
The energy value determined, wtAnd btIndicate weight and biasing;By formula (7) using e as the power of truth of a matter various pieces energy value therewith
The available weight coefficient for having much influences on classification results of ratio of the cumulative sum of the energy value of preceding part, it is thus achieved that
Conversion of the original state to state of attention;Then as formula (8) obtains the state vector Y of final output;Finally by Y by connecting entirely
It connects layer to combine as an output valve, reducing feature locations influences classification bring, will by softmax classifier
The output of multiple neurons is mapped in (0,1) section, thus to carry out classify more.
The present invention is tested under GPU acceleration environment using python language, using keras deep learning frame, electricity
Brain is configured to Win10 system, 16GB memory, GTX1080 11G video memory.Network in training Bi-LSTM-Attention model
Parameter.
Experiment shows that the precision of network model proposed in this paper reaches 94.38% on Action Youtobe data set,
1 Action Youtobe data set of table is compared with other model algorithms
As it can be seen from table 1 in Action Youtobe data set, it is proposed by the present invention to be based on Bi-LSTM-
The Human bodys' response method of Attention model is available to be better than Binary after combining InceptionV3 model
CNN-Flow, Discriminative representation, tri- kinds of Proposed DB-LSTM based on deep learning algorithm
Precision can also be obtained also superior to other three kinds of traditional algorithms based on manual feature extraction: Hierarchical clustering
multi-task,Fisher vectors,3D spatio-temporal.Meanwhile the present invention under same model to LSTM,
Two kinds of algorithms of Bi-LSTM are tested, the experimental results showed that, using Bi-LSTM-Attention model to accuracy of identification band
Carry out 4.85% and 1.57% promotion.
The present invention also uses LSTM, Bi-LSTM, and Bi-LSTM-Attention combination InceptionV3 three kinds of methods of model exist
It is tested on KTH data set, average accuracy of identification of the method proposed by the present invention on KTH can achieve 95.67%.Than
The result of LSTM and Bi-LSTM algorithm has been higher by 5.33% and 1%.
2 KTH data set of table is compared with other model algorithms
The Human bodys' response proposed by the invention based on Bi-LSTM-Attention model it can be seen from upper table 2
Method still has good performance on KTH data set, it was demonstrated that the feasibility of algorithm proposed in this paper.
It should be noted last that the above specific embodiment is only used to illustrate the technical scheme of the present invention and not to limit it,
Although being described the invention in detail referring to example, those skilled in the art should understand that, it can be to the present invention
Technical solution be modified or replaced equivalently, without departing from the spirit and scope of the technical solution of the present invention, should all cover
In the scope of the claims of the present invention.
Claims (4)
1. a kind of Human bodys' response method based on Bi-LSTM-Attention model, which is characterized in that including following step
It is rapid:
Step S1, inputs InceptionV3 model for the video frame of extraction, increases convolutional Neural using InceptionV3 model
Network parameter is reduced while network depth, is sufficiently extracted the depth characteristic of video frame, is obtained relevant feature vector;
The obtained feature vector of step S1 is passed in Bi-LSTM neural network and handles, passes through Bi-LSTM by step S2
Neural network sufficiently learns the temporal aspect between video frame;
The temporal aspect vector that step S2 is obtained is passed to attention Mechanism Model and adaptively perceived to identification knot by step S3
Fruit has the network weight of larger impact, and the relevant feature of these network weights is more paid close attention to.
2. as described in claim 1 based on the Human bodys' response method of Bi-LSTM-Attention model, feature exists
In,
In step S1, different convolutional layers is combined together by InceptionV3 model by way of in parallel, while using not
Convolution kernel with size carries out convolution operation, the feature that different convolution kernels are handled finally by filter fused layer to video frame
Vector is stitched together, and exports depth characteristic matrix by full articulamentum and is used for transmission in Bi-LSTM neural network.
3. as described in claim 1 based on the Human bodys' response method of Bi-LSTM-Attention model, feature exists
In,
Step S2 is specifically included:
wi(i=1 ... 6) indicates a layer network layer to the weight of another network layer;{…ht-1,ht,ht+1... indicate LSTM nerve net
Propagated forward layer in network, the input of propagated forward layer be ... xt-1,xt,xt+1... characteristic sequence from front to back;
{…ht+1',ht',ht-1' ... indicate LSTM neural network in back-propagating layer, the input of back-propagating layer be ...
xt+1,xt,xt-1... characteristic sequence from back to front;
X thereintIndicate that extracted video frame passes through the feature vector obtained after InceptionV3 model extraction depth characteristic;
Such as following formula:
ht=f (w1xt+w2ht-1+b1) (1)
h′t=f (w3xt+w5ht+1+b2) (2)
o′t=g (w4ht+b3) (3)
o″t=g (w6h′t+b3) (4)
ot=(o 't+o″t)/2 (5)
Above formula (1), (2), (3), the f in (4) and g represent activation primitive, b1、b2、b3、b4The biasing coefficient of hidden unit is represented,
O', o " are the result that two LSTM units handle the feature vector of Inceptionv3 layers of output at the corresponding moment respectively;Corresponding
Two feature vectors at moment are added the temporal aspect vector summed and be averaged as output.
4. as described in claim 1 based on the Human bodys' response method of Bi-LSTM-Attention model, feature exists
In step S3 is specifically included:
otIt indicates t-th of the temporal aspect vector exported from Bi-LSTM neural network, then temporal aspect vector is passed to
In attention Mechanism Model, initial state vector S is obtained by the hidden layer in attention Mechanism Modelt;Weight coefficient αtIt indicates
Initial state vector StThe shared specific gravity size in the state vector Y of final output;Each initial state vector StWith weight system
Number αtProduct cumulative and obtain the state vector Y of final output;Calculation formula is as follows:
et=tanh (wtst+bt) (6)
Tanh indicates that excitation function, n indicate the quantity of video frame;etIndicate the state vector S of t-th of temporal aspect vectortIt is determined
Fixed energy value, wtAnd btIndicate weight and biasing;By formula (7) using e as the power of truth of a matter various pieces energy value front therewith
The available weight coefficient for having much influences on classification results of ratio of the cumulative sum of the energy value divided, it is thus achieved that initially
Conversion of the state to state of attention;Then as formula (8) obtains the state vector Y of final output.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910048015.8A CN109784280A (en) | 2019-01-18 | 2019-01-18 | Human bodys' response method based on Bi-LSTM-Attention model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910048015.8A CN109784280A (en) | 2019-01-18 | 2019-01-18 | Human bodys' response method based on Bi-LSTM-Attention model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109784280A true CN109784280A (en) | 2019-05-21 |
Family
ID=66501453
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910048015.8A Pending CN109784280A (en) | 2019-01-18 | 2019-01-18 | Human bodys' response method based on Bi-LSTM-Attention model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109784280A (en) |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110222653A (en) * | 2019-06-11 | 2019-09-10 | 中国矿业大学(北京) | A kind of skeleton data Activity recognition method based on figure convolutional neural networks |
CN110245581A (en) * | 2019-05-25 | 2019-09-17 | 天津大学 | A kind of Human bodys' response method based on deep learning and distance-Doppler sequence |
CN110287820A (en) * | 2019-06-06 | 2019-09-27 | 北京清微智能科技有限公司 | Activity recognition method, apparatus, equipment and medium based on LRCN network |
CN110427834A (en) * | 2019-07-10 | 2019-11-08 | 上海工程技术大学 | A kind of Activity recognition system and method based on skeleton data |
CN110664412A (en) * | 2019-09-19 | 2020-01-10 | 天津师范大学 | Human activity recognition method facing wearable sensor |
CN110765956A (en) * | 2019-10-28 | 2020-02-07 | 西安电子科技大学 | Double-person interactive behavior recognition method based on component characteristics |
CN110929243A (en) * | 2019-11-22 | 2020-03-27 | 武汉大学 | Pedestrian identity recognition method based on mobile phone inertial sensor |
CN110991480A (en) * | 2019-10-31 | 2020-04-10 | 上海交通大学 | Attention mechanism-based sparse coding method |
CN110990608A (en) * | 2019-12-03 | 2020-04-10 | 哈尔滨工业大学 | Three-dimensional model retrieval method based on Simese structure bidirectional long-time and short-time memory network |
CN111079547A (en) * | 2019-11-22 | 2020-04-28 | 武汉大学 | Pedestrian moving direction identification method based on mobile phone inertial sensor |
CN111082879A (en) * | 2019-12-27 | 2020-04-28 | 南京邮电大学 | Wifi perception method based on deep space-time model |
CN111079599A (en) * | 2019-12-06 | 2020-04-28 | 浙江工业大学 | Human body complex behavior recognition method based on multi-feature fusion CNN-BLSTM |
CN111191663A (en) * | 2019-12-31 | 2020-05-22 | 深圳云天励飞技术有限公司 | License plate number recognition method and device, electronic equipment and storage medium |
CN111372123A (en) * | 2020-03-03 | 2020-07-03 | 南京信息工程大学 | Video time sequence segment extraction method based on local to global |
CN111401209A (en) * | 2020-03-11 | 2020-07-10 | 佛山市南海区广工大数控装备协同创新研究院 | Action recognition method based on deep learning |
CN111523410A (en) * | 2020-04-09 | 2020-08-11 | 哈尔滨工业大学 | Video saliency target detection method based on attention mechanism |
CN111597881A (en) * | 2020-04-03 | 2020-08-28 | 浙江工业大学 | Human body complex behavior identification method based on data separation multi-scale feature combination |
CN112037179A (en) * | 2020-08-11 | 2020-12-04 | 深圳大学 | Method, system and equipment for generating brain disease diagnosis model |
CN112131981A (en) * | 2020-09-10 | 2020-12-25 | 山东大学 | Driver fatigue detection method based on skeleton data behavior recognition |
CN112131972A (en) * | 2020-09-07 | 2020-12-25 | 重庆邮电大学 | Method for recognizing human body behaviors by using WiFi data based on attention mechanism |
CN112528733A (en) * | 2020-10-29 | 2021-03-19 | 西安工程大学 | Abnormal behavior identification method of network |
CN112528891A (en) * | 2020-12-16 | 2021-03-19 | 重庆邮电大学 | Bidirectional LSTM-CNN video behavior identification method based on skeleton information |
CN112597921A (en) * | 2020-12-28 | 2021-04-02 | 杭州电子科技大学 | Human behavior recognition method based on attention mechanism GRU deep learning |
CN112686211A (en) * | 2021-01-25 | 2021-04-20 | 广东工业大学 | Fall detection method and device based on attitude estimation |
CN112766420A (en) * | 2021-03-12 | 2021-05-07 | 合肥共达职业技术学院 | Human behavior identification method based on time-frequency domain information |
CN114120166A (en) * | 2021-10-14 | 2022-03-01 | 北京百度网讯科技有限公司 | Video question and answer method and device, electronic equipment and storage medium |
CN114155480A (en) * | 2022-02-10 | 2022-03-08 | 北京智视数策科技发展有限公司 | Vulgar action recognition method |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160314392A1 (en) * | 2015-03-26 | 2016-10-27 | Nokia Technologies Oy | Generating using a bidirectional rnn variations to music |
CN108363753A (en) * | 2018-01-30 | 2018-08-03 | 南京邮电大学 | Comment text sentiment classification model is trained and sensibility classification method, device and equipment |
CN108427670A (en) * | 2018-04-08 | 2018-08-21 | 重庆邮电大学 | A kind of sentiment analysis method based on context word vector sum deep learning |
CN108519890A (en) * | 2018-04-08 | 2018-09-11 | 武汉大学 | A kind of robustness code abstraction generating method based on from attention mechanism |
US20180288086A1 (en) * | 2017-04-03 | 2018-10-04 | Royal Bank Of Canada | Systems and methods for cyberbot network detection |
CN108763204A (en) * | 2018-05-21 | 2018-11-06 | 浙江大学 | A kind of multi-level text emotion feature extracting method and model |
AU2018101514A4 (en) * | 2018-10-11 | 2018-11-15 | Chi, Henan Mr | An automatic text-generating program for Chinese Hip-hop lyrics |
CN108829801A (en) * | 2018-06-06 | 2018-11-16 | 大连理工大学 | A kind of event trigger word abstracting method based on documentation level attention mechanism |
-
2019
- 2019-01-18 CN CN201910048015.8A patent/CN109784280A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160314392A1 (en) * | 2015-03-26 | 2016-10-27 | Nokia Technologies Oy | Generating using a bidirectional rnn variations to music |
US20180288086A1 (en) * | 2017-04-03 | 2018-10-04 | Royal Bank Of Canada | Systems and methods for cyberbot network detection |
CN108363753A (en) * | 2018-01-30 | 2018-08-03 | 南京邮电大学 | Comment text sentiment classification model is trained and sensibility classification method, device and equipment |
CN108427670A (en) * | 2018-04-08 | 2018-08-21 | 重庆邮电大学 | A kind of sentiment analysis method based on context word vector sum deep learning |
CN108519890A (en) * | 2018-04-08 | 2018-09-11 | 武汉大学 | A kind of robustness code abstraction generating method based on from attention mechanism |
CN108763204A (en) * | 2018-05-21 | 2018-11-06 | 浙江大学 | A kind of multi-level text emotion feature extracting method and model |
CN108829801A (en) * | 2018-06-06 | 2018-11-16 | 大连理工大学 | A kind of event trigger word abstracting method based on documentation level attention mechanism |
AU2018101514A4 (en) * | 2018-10-11 | 2018-11-15 | Chi, Henan Mr | An automatic text-generating program for Chinese Hip-hop lyrics |
Non-Patent Citations (4)
Title |
---|
CHRISTIAN SZEGEDY 等: "Rethinking the Inception Architecture for Computer Vision", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
HUANHOU XIAO 等: "Video Captioning using Hierarchical Multi-Attention Model", 《ACM》 * |
PENG ZHOU 等: "Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification", 《PROCEEDINGS OF THE 54TH ANNUAL MEETING OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS》 * |
高志强 等编著: "《深度学习 从入门到实战》", 30 June 2018 * |
Cited By (43)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110245581A (en) * | 2019-05-25 | 2019-09-17 | 天津大学 | A kind of Human bodys' response method based on deep learning and distance-Doppler sequence |
CN110245581B (en) * | 2019-05-25 | 2023-04-07 | 天津大学 | Human behavior recognition method based on deep learning and distance-Doppler sequence |
CN110287820A (en) * | 2019-06-06 | 2019-09-27 | 北京清微智能科技有限公司 | Activity recognition method, apparatus, equipment and medium based on LRCN network |
CN110287820B (en) * | 2019-06-06 | 2021-07-23 | 北京清微智能科技有限公司 | Behavior recognition method, device, equipment and medium based on LRCN network |
CN110222653B (en) * | 2019-06-11 | 2020-06-16 | 中国矿业大学(北京) | Skeleton data behavior identification method based on graph convolution neural network |
CN110222653A (en) * | 2019-06-11 | 2019-09-10 | 中国矿业大学(北京) | A kind of skeleton data Activity recognition method based on figure convolutional neural networks |
CN110427834A (en) * | 2019-07-10 | 2019-11-08 | 上海工程技术大学 | A kind of Activity recognition system and method based on skeleton data |
CN110664412A (en) * | 2019-09-19 | 2020-01-10 | 天津师范大学 | Human activity recognition method facing wearable sensor |
CN110765956A (en) * | 2019-10-28 | 2020-02-07 | 西安电子科技大学 | Double-person interactive behavior recognition method based on component characteristics |
CN110765956B (en) * | 2019-10-28 | 2021-10-29 | 西安电子科技大学 | Double-person interactive behavior recognition method based on component characteristics |
CN110991480A (en) * | 2019-10-31 | 2020-04-10 | 上海交通大学 | Attention mechanism-based sparse coding method |
CN111079547A (en) * | 2019-11-22 | 2020-04-28 | 武汉大学 | Pedestrian moving direction identification method based on mobile phone inertial sensor |
CN111079547B (en) * | 2019-11-22 | 2022-07-19 | 武汉大学 | Pedestrian moving direction identification method based on mobile phone inertial sensor |
CN110929243B (en) * | 2019-11-22 | 2022-07-22 | 武汉大学 | Pedestrian identity recognition method based on mobile phone inertial sensor |
CN110929243A (en) * | 2019-11-22 | 2020-03-27 | 武汉大学 | Pedestrian identity recognition method based on mobile phone inertial sensor |
CN110990608A (en) * | 2019-12-03 | 2020-04-10 | 哈尔滨工业大学 | Three-dimensional model retrieval method based on Simese structure bidirectional long-time and short-time memory network |
CN111079599B (en) * | 2019-12-06 | 2022-04-05 | 浙江工业大学 | Human body complex behavior recognition method based on multi-feature fusion CNN-BLSTM |
CN111079599A (en) * | 2019-12-06 | 2020-04-28 | 浙江工业大学 | Human body complex behavior recognition method based on multi-feature fusion CNN-BLSTM |
CN111082879B (en) * | 2019-12-27 | 2022-02-01 | 南京邮电大学 | Wifi perception method based on deep space-time model |
CN111082879A (en) * | 2019-12-27 | 2020-04-28 | 南京邮电大学 | Wifi perception method based on deep space-time model |
CN111191663B (en) * | 2019-12-31 | 2022-01-11 | 深圳云天励飞技术股份有限公司 | License plate number recognition method and device, electronic equipment and storage medium |
CN111191663A (en) * | 2019-12-31 | 2020-05-22 | 深圳云天励飞技术有限公司 | License plate number recognition method and device, electronic equipment and storage medium |
CN111372123A (en) * | 2020-03-03 | 2020-07-03 | 南京信息工程大学 | Video time sequence segment extraction method based on local to global |
CN111372123B (en) * | 2020-03-03 | 2022-08-09 | 南京信息工程大学 | Video time sequence segment extraction method based on local to global |
CN111401209B (en) * | 2020-03-11 | 2023-11-07 | 佛山市南海区广工大数控装备协同创新研究院 | Action recognition method based on deep learning |
CN111401209A (en) * | 2020-03-11 | 2020-07-10 | 佛山市南海区广工大数控装备协同创新研究院 | Action recognition method based on deep learning |
CN111597881B (en) * | 2020-04-03 | 2022-04-05 | 浙江工业大学 | Human body complex behavior identification method based on data separation multi-scale feature combination |
CN111597881A (en) * | 2020-04-03 | 2020-08-28 | 浙江工业大学 | Human body complex behavior identification method based on data separation multi-scale feature combination |
CN111523410B (en) * | 2020-04-09 | 2022-08-26 | 哈尔滨工业大学 | Video saliency target detection method based on attention mechanism |
CN111523410A (en) * | 2020-04-09 | 2020-08-11 | 哈尔滨工业大学 | Video saliency target detection method based on attention mechanism |
CN112037179A (en) * | 2020-08-11 | 2020-12-04 | 深圳大学 | Method, system and equipment for generating brain disease diagnosis model |
CN112131972A (en) * | 2020-09-07 | 2020-12-25 | 重庆邮电大学 | Method for recognizing human body behaviors by using WiFi data based on attention mechanism |
CN112131981A (en) * | 2020-09-10 | 2020-12-25 | 山东大学 | Driver fatigue detection method based on skeleton data behavior recognition |
CN112528733A (en) * | 2020-10-29 | 2021-03-19 | 西安工程大学 | Abnormal behavior identification method of network |
CN112528733B (en) * | 2020-10-29 | 2024-03-22 | 西安工程大学 | Human body abnormal behavior identification method based on improved incapacity v3 network |
CN112528891A (en) * | 2020-12-16 | 2021-03-19 | 重庆邮电大学 | Bidirectional LSTM-CNN video behavior identification method based on skeleton information |
CN112597921A (en) * | 2020-12-28 | 2021-04-02 | 杭州电子科技大学 | Human behavior recognition method based on attention mechanism GRU deep learning |
CN112597921B (en) * | 2020-12-28 | 2024-02-02 | 杭州电子科技大学 | Human behavior recognition method based on attention mechanism GRU deep learning |
CN112686211A (en) * | 2021-01-25 | 2021-04-20 | 广东工业大学 | Fall detection method and device based on attitude estimation |
CN112766420A (en) * | 2021-03-12 | 2021-05-07 | 合肥共达职业技术学院 | Human behavior identification method based on time-frequency domain information |
CN114120166A (en) * | 2021-10-14 | 2022-03-01 | 北京百度网讯科技有限公司 | Video question and answer method and device, electronic equipment and storage medium |
CN114120166B (en) * | 2021-10-14 | 2023-09-22 | 北京百度网讯科技有限公司 | Video question-answering method and device, electronic equipment and storage medium |
CN114155480A (en) * | 2022-02-10 | 2022-03-08 | 北京智视数策科技发展有限公司 | Vulgar action recognition method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109784280A (en) | Human bodys' response method based on Bi-LSTM-Attention model | |
CN108717568B (en) | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network | |
Song et al. | Region-based quality estimation network for large-scale person re-identification | |
Deng et al. | Learning to predict crisp boundaries | |
Zhang et al. | Context encoding for semantic segmentation | |
CN108830157B (en) | Human behavior identification method based on attention mechanism and 3D convolutional neural network | |
CN109685819B (en) | Three-dimensional medical image segmentation method based on feature enhancement | |
CN109543697A (en) | A kind of RGBD images steganalysis method based on deep learning | |
CN111832516B (en) | Video behavior recognition method based on unsupervised video representation learning | |
CN108133188A (en) | A kind of Activity recognition method based on motion history image and convolutional neural networks | |
CN111950455B (en) | Motion imagery electroencephalogram characteristic identification method based on LFFCNN-GRU algorithm model | |
CN112784764A (en) | Expression recognition method and system based on local and global attention mechanism | |
CN109063719B (en) | Image classification method combining structure similarity and class information | |
CN110378208B (en) | Behavior identification method based on deep residual error network | |
CN108647599B (en) | Human behavior recognition method combining 3D (three-dimensional) jump layer connection and recurrent neural network | |
CN113496217A (en) | Method for identifying human face micro expression in video image sequence | |
CN104809469A (en) | Indoor scene image classification method facing service robot | |
CN113920581B (en) | Method for identifying actions in video by using space-time convolution attention network | |
CN108921047A (en) | A kind of multi-model ballot mean value action identification method based on cross-layer fusion | |
CN104063721A (en) | Human behavior recognition method based on automatic semantic feature study and screening | |
CN104408461A (en) | A method of identifying motion of local matching window based on sliding window | |
CN110322418A (en) | A kind of super-resolution image generates the training method and device of confrontation network | |
CN114241564A (en) | Facial expression recognition method based on inter-class difference strengthening network | |
CN110288603A (en) | Semantic segmentation method based on efficient convolutional network and convolution condition random field | |
CN113850182A (en) | Action identification method based on DAMR-3 DNet |
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: 20190521 |
|
RJ01 | Rejection of invention patent application after publication |