CN102413378A - Adaptive neural network-based lost packet recovery method in video transmission - Google Patents
Adaptive neural network-based lost packet recovery method in video transmission Download PDFInfo
- Publication number
- CN102413378A CN102413378A CN2011103413360A CN201110341336A CN102413378A CN 102413378 A CN102413378 A CN 102413378A CN 2011103413360 A CN2011103413360 A CN 2011103413360A CN 201110341336 A CN201110341336 A CN 201110341336A CN 102413378 A CN102413378 A CN 102413378A
- Authority
- CN
- China
- Prior art keywords
- video
- network
- value
- critic
- input
- 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
Images
Abstract
The invention relates to an adaptive neural network-based lost packet recovery method in video transmission. The lower layer of a protocol which is used in the existing network transmission has an error correcting code, however, only the problem of error code in the packet rather than the problem of lost packet can be solved. The method comprises the following steps of: taking video information as the state of the current environment to be input; outputting a video encoding rate which is to be used under the input states, wherein a main network lost packet model uses a back propagation adaptive neutral network-based AHC (Actor-Critic) model; and leading the Critic to receive reward feedback of an environment in the process of studying, updating Value (x), and handing a reward forecast error to an Actor module in the manner of external feedback for guiding the Actor to correct a strategy of a selecting action. The influence of the network lost packet to internet protocol (IP) video call is effectively reduced, and the stability of video interaction is improved.
Description
Technical field
The invention belongs to the video data technology field, relate to a kind of video transmission loss recovery technology based on adaptive neural network.
Background technology
Along with the development of 3G technology and broadband network, and constantly the weeding out the old and bring forth the new of terminal, video communication service will obtain significant progress, thus Video transmission system to real-time and accuracy require increasingly high.In the video transmission process of network, the principal element that influences audio frequency tonequality has time delay, electrostatic interference, packet loss and shake etc.Wherein, the packet loss problem is a most crucial factor that influences video interactive service quality always.Although the lower protocol layers of using in the Network Transmission has error correcting code, can only solve error code in the bag, can't solve the packet loss problem.The wrong retransmission mechanism of Transmission Control Protocol can guarantee the correctness of network data flow; But it can't satisfy the real-time requirement of video signal transmission; Particularly under the situation of many group broadcasting, different recipients has different packet loss retransmission requesting, and this not only can increase the consumption of the network bandwidth; And cause Network Transmission to be stopped up conversely, cause bigger data-bag lost.Therefore, need study, to improve the quality of video transmission to the loss recovery treatment technology of application layer.
Backpropagation adaptive neural network (Back Propagation Neural Network is called for short the BP network) is to use the most extensive and successful a kind of artificial neural net at present.Mainly form by input layer, hidden layer and output layer.Feedforward network is meant, only exists the weights between the different layers neuron to connect, and do not exist with the connection between one deck neuron, and also the connection between the neuron, the just connection between adjacent layer, and the situation that interlayer connects does not appear.It is existing that oneself proves: the single hidden layer feedforward network with S type activation primitive; As long as hidden layer has fully many neurons; It just can approach any known function with precision arbitrarily, thereby shows that the BP neural net can be used as a general function approximation device.The BP Application of Neural Network among the video transmission packet loss, can be overcome the defective of traditional packet loss detection video recovery technology, thereby has good application and promotion prospect.
Summary of the invention
The present invention proposes a kind of video transmission loss recovery technology based on adaptive neural network; Comprise the Network Packet Loss model and based on backpropagation adaptive neural network (Back Propagation Neural Network; The BP network) loss recovery technology, the real-time and the accuracy of raising Video transmission system.
The present invention adopts following technical solution: at first video information is as the state input of current environment, the video frequency coding rate that output should be used under these input states.Its main Network Packet Loss model has used the AHC model (or being called the Actor-Critic model) based on the backpropagation adaptive neural network; Wherein Actor is responsible for producing an action under the current state condition; Critic then is responsible for can getablely rewarding under the study prediction current state condition, and Value (x) is Critic and environment is rewarded the function of predicting.In the learning process, Critic accepts the award feedback of environment, upgrades Value (x), and will reward predicated error and give Actor module with the form of external world's feedback, is used to instruct Actor to correct the strategy of choosing action.Wherein, Value (x) learns according to the Q-learning update rule, and Actor then adopts Gaussian ASLA function of movement to realize, Actor and Critic use BP (Back-Propagation) neural net to carry out extensive work.
Concrete grammar of the present invention may further comprise the steps:
Input variable comprises three types of network parameter information and video information, and described three types of network parameter information are respectively 4 time delay state values, 4 dither state values, 1 packet loss state value;
Described video information is 6 frame complexity state values;
The video frequency coding rate of output variable for using.
Step 2. data normalization is handled, and the data in the input and output sequence is carried out normalization handle, and specifically is to the processing of standardizing of the state variable parameter of time delay, shake and frame complexity, is converted into the value of [0,1] scope.
Step 3. video delivery network packet loss model has used the AHC model (Actor-Critic model) based on the backpropagation adaptive neural network, and concrete grammar is:
Through using three neural nets to be used for the Value (x) among the match Critic, μ (x) and the α (x) among the Actor respectively, wherein μ (x) and α (x) represent the average and the standard deviation of output valve respectively;
Described Actor is responsible for producing an action under the current state condition,
Described Critic then is responsible for can getablely rewarding under the study prediction current state condition;
Described Value (x) rewards the function of predicting for Critic to environment;
The input of three neural nets is identical, all is normalization variable state value afterwards, and it is Inputl-Inputl5 that input layer has 15 input units; Each neural net has a latent layer, and latent layer has three neurons to consist of Hiden1-Hiden3; It is Out that output layer has an output neuron; The neuronic threshold function of in the network each uses be can be little the Sigmoid function.
In the learning process of step 4. training BP neural net; Concrete grammar is:
Actor selects a code check when moment t-1, then can receive the feedback incentive message of environment, promptly is the video quality information after proofreading and correct here; This moment, Critic need use the Q-learning update rule, upgraded the award anticipation function of oneself.
The Q-learning update rule is the deviation that anticipation function is predicted when moment t-1, shown in 1.
R wherein
T-1Be the award that the action of t-1 is constantly received, Value
tWith Value
T-1The output valve of function Value (x) after Critic assesses environment of living in when being respectively moment t and t-1, γ is a learning parameter;
Formula 1 is used the actual value of the award that the current predicted value of anticipation function Value (x) replaces can obtaining in the future, adds by moment t-1 to the award of gained the t constantly the award that should obtain during moment t-1 exactly then.
Step 5. test b P neural net;
BP neural net to training is accomplished is tested; Historical data is formed input information according to network parameter information in the step (1) and video information; Carrying out normalization according to step (2) again handles; Through obtaining 14 dateouts altogether after the normalization, add packet loss in addition like this, it is corresponding with outl-outl5 respectively to have 15 Input; Directly call the sim function in the MatLab Neural Network Toolbox according to step (3), test matrix carried out emulation, wherein in the network each neuronic threshold function all be can be little the Sigmoid function, correspond to the packet loss predicted value of video delivery network.
The anti-normalization of step 6. data is handled;
The video code rate value that output is under the current state is carried out anti-normalization processing, uses the obedience average to produce output valve as μ (x), standard deviation as the Gaussian distribution randomizer of α (x).
The beneficial effect that the present invention has is: through the video loss recovery technology of research based on the backpropagation adaptive neural network; Permission provides significantly improved normal video quality in having the network of packet loss; With of the influence of effective reduction Network Packet Loss, improve the stability of video interactive to the IP video call.Make when receiving more relatively multimedia data information based on the video call system receiving terminal; Also can recover accurate data information more; Thereby can improve the video displaying quality; For providing, video user guarantees service quality (Quality of Service, video transmission technologies QoS).
Description of drawings
Fig. 1 is based on the sketch map of the loss recovery method of backpropagation adaptive neural network.
Embodiment
Below in conjunction with accompanying drawing the present invention is described further.
As shown in Figure 1, the present invention includes following steps:
Input variable comprises three types of network parameter information and video information, and described three types of network parameter information are respectively 4 time delay state values, 4 dither state values, 1 packet loss state value; Described video information is 6 frame complexity state values, the video frequency coding rate of output variable for using;
Step 2. data normalization is handled:
Because network delay has very big distribution, arrive the hundreds of millisecond greatly for a short time to being lower than 1 millisecond, delay variation also is like this, the scope of packet loss is then very little, can get the value between little 1 in theory, in fact is generally less than 0.01.In addition, most of Frame complexity is positioned at 10
5Below, there is few frames can reach 4*10
5More than.Directly the state value that these spans are widely different is very inappropriate as input, needs to use normalized method, the data in the input and output sequence is carried out normalization handle.
Therefore, to the processing of standardizing of the state variable parameter of time delay (getting 4 values), shake (getting 4 values), three parts of frame frame complexity (getting 6 values), the real number value x of each input is converted into the output that is positioned at value between [0,1].Considering above three state variables mentioning, because the span of packet loss between 0-1, does not need normalization.Like this through the normalization after obtain 14 dateouts altogether, add packet loss in addition, have 15 respectively with Fig. 1 in outl-outl5 corresponding.These 15 data are used for the state of descriptive system place environment, as Actor and Critic.The input of employed neural net in the fit procedure, respectively with Fig. 1 in Inputl-Inputl5 corresponding one by one.
Step 3. video delivery network packet loss model has used the AHC model (Actor-Critic model) based on the backpropagation adaptive neural network, and concrete grammar is:
Through using three neural nets to be used for the Value (x) among the match Critic, μ (x) and the α (x) among the Actor respectively, wherein μ (x) and α (x) represent the average and the standard deviation of output valve respectively;
Described Actor is responsible for producing an action under the current state condition,
Described Critic then is responsible for can getablely rewarding under the study prediction current state condition;
Described Value (x) rewards the function of predicting for Critic to environment;
The input of three neural nets is identical, all is normalization variable state value afterwards, and it is Inputl-Inputl5 that input layer as shown in fig. 1 has 15 input units; Each neural net has a latent layer (Hidden Layer), and latent layer has three neurons to consist of Hiden1-Hiden3; It is Out that output layer has an output neuron; The neuronic threshold function of in the network each uses be can be little the Sigmoid function;
In the learning process of step 4. training BP neural net; Concrete grammar is:
Actor selects a code check when moment t-1, then can receive the feedback incentive message of environment, promptly is the video quality information after proofreading and correct here; This moment, Critic need use the Q-learning update rule, upgraded the award anticipation function of oneself;
The Q-learning update rule is the deviation that anticipation function is predicted when moment t-1, shown in 1.
R wherein
T-1Be the award that the action of t-1 is constantly received, Value
tWith Value
T-1The output valve of function Value (x) after Critic assesses environment of living in when being respectively moment t and t-1, γ is a learning parameter;
Formula 1 is used the actual value of the award that the current predicted value of anticipation function Value (x) replaces can obtaining in the future, adds by moment t-1 to the award of gained the t constantly the award that should obtain during moment t-1 exactly then;
Step 5. test b P neural net;
BP neural net to training is accomplished is tested; Historical data is formed input information according to network parameter information in the step (1) and video information; Carrying out normalization according to step (2) again handles; Through obtaining 14 dateouts altogether after the normalization, add packet loss in addition like this, it is corresponding with outl-outl5 respectively to have 15 Input; Directly call the sim function in the MatLab Neural Network Toolbox according to step (3), test matrix carried out emulation, wherein in the network each neuronic threshold function all be can be little the Sigmoid function, correspond to the packet loss predicted value of video delivery network;
The anti-normalization of step 6. data is handled;
The video code rate value that output is under the current state is carried out anti-normalization processing, uses the obedience average to produce output valve as μ (x), standard deviation as the Gaussian distribution randomizer of α (x).
Claims (1)
1. video transmission loss recovery method based on adaptive neural network is characterized in that this method may further comprise the steps:
Step 1. is selected the input variable and the output variable of video loss recovery model, confirms training sample; Concrete grammar is:
Input variable comprises three types of network parameter information and video information, and described three types of network parameter information are respectively 4 time delay state values, 4 dither state values, 1 packet loss state value;
Described video information is 6 frame complexity state values,
The video frequency coding rate of output variable for using;
Step 2. data normalization is handled, and the data in the input and output sequence is carried out normalization handle, and specifically is to the processing of standardizing of the state variable parameter of time delay, shake and frame complexity, is converted into the value of [0,1] scope;
Step 3. video delivery network packet loss model has used the AHC model (Actor-Critic model) based on the backpropagation adaptive neural network, and concrete grammar is:
Through using three neural nets to be used for the Value (x) among the match Critic, μ (x) and the α (x) among the Actor respectively, wherein μ (x) and α (x) represent the average and the standard deviation of output valve respectively;
Described Actor is responsible for producing an action under the current state condition;
Described Critic then is responsible for can getablely rewarding under the study prediction current state condition;
Described Value (x) rewards the function of predicting for Critic to environment;
The input of three neural nets is identical, all is normalization variable state value afterwards, and it is Inputl-Inputl5 that input layer has 15 input units; Each neural net has a latent layer, and latent layer has three neurons to consist of Hiden1-Hiden3; It is Out that output layer has an output neuron; The neuronic threshold function of in the network each uses be can be little the sigmoid function;
In the learning process of step 4. training BP neural net; Concrete grammar is:
Actor selects a code check when moment t-1, then can receive the feedback incentive message of environment, promptly is the video quality information after proofreading and correct here; This moment, Critic need use the Q-learning update rule, upgraded the award anticipation function of oneself;
The Q-learning update rule is the deviation that anticipation function is predicted when moment t-1, shown in the following formula
R wherein
T-1Be the award that the action of t-1 is constantly received, Value
tWith Value
T-1The output valve of function Value (x) after Critic assesses environment of living in when being respectively moment t and t-1, γ is a learning parameter;
Following formula uses the actual value of the award that the current predicted value of anticipation function Value (x) replaces can obtaining in the future, adds by moment t-1 to the award of gained the t constantly the award that should obtain during moment t-1 exactly then;
Step 5. test b P neural net;
BP neural net to training is accomplished is tested; Historical data is formed input information according to network parameter information in the step (1) and video information; Carrying out normalization according to step (2) again handles; Through obtaining 14 dateouts altogether after the normalization, add packet loss in addition like this, it is corresponding with outl-outl5 respectively to have 15 Input; Directly call the sim function in the MatLab Neural Network Toolbox according to step (3), test matrix carried out emulation, wherein in the network each neuronic threshold function all be can be little the Sigmoid function, correspond to the packet loss predicted value of video delivery network;
The anti-normalization of step 6. data is handled;
The video code rate value that output is under the current state is carried out anti-normalization processing, uses the obedience average to produce output valve as μ (x), standard deviation as the Gaussian distribution randomizer of α (x).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2011103413360A CN102413378A (en) | 2011-11-02 | 2011-11-02 | Adaptive neural network-based lost packet recovery method in video transmission |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2011103413360A CN102413378A (en) | 2011-11-02 | 2011-11-02 | Adaptive neural network-based lost packet recovery method in video transmission |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102413378A true CN102413378A (en) | 2012-04-11 |
Family
ID=45915175
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2011103413360A Pending CN102413378A (en) | 2011-11-02 | 2011-11-02 | Adaptive neural network-based lost packet recovery method in video transmission |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102413378A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107547914A (en) * | 2017-08-15 | 2018-01-05 | 浙江工业大学 | The substandard video segments based on KNN Q study of DASH obtain optimization method |
| CN107808004A (en) * | 2017-11-15 | 2018-03-16 | 北京百度网讯科技有限公司 | Model training method and system, server, storage medium |
| CN107992939A (en) * | 2017-12-06 | 2018-05-04 | 湖北工业大学 | Cutting force gear working method is waited based on depth enhancing study |
| CN108010533A (en) * | 2016-10-27 | 2018-05-08 | 北京酷我科技有限公司 | The automatic identifying method and device of voice data code check |
| CN108063961A (en) * | 2017-12-22 | 2018-05-22 | 北京联合网视文化传播有限公司 | A kind of self-adaption code rate video transmission method and system based on intensified learning |
| CN109218083A (en) * | 2018-08-27 | 2019-01-15 | 广州爱拍网络科技有限公司 | A kind of voice data transmission method and device |
| CN109726811A (en) * | 2017-10-27 | 2019-05-07 | 谷歌有限责任公司 | Use priority formation neural network |
| CN110006486A (en) * | 2019-04-01 | 2019-07-12 | 中清控(武汉)科技有限公司 | A kind of intelligence flow temperature acquisition control module and intelligence flow temperature measurement method |
| CN112511482A (en) * | 2019-09-16 | 2021-03-16 | 华为技术有限公司 | Media data transmission method, device and system |
| CN112532349A (en) * | 2020-11-24 | 2021-03-19 | 广州技象科技有限公司 | Data processing method and device based on decoding abnormity |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101808244A (en) * | 2010-03-24 | 2010-08-18 | 北京邮电大学 | Video transmission control method and system |
| CN102075312A (en) * | 2011-01-10 | 2011-05-25 | 西安电子科技大学 | Video service quality-based hybrid selective repeat method |
-
2011
- 2011-11-02 CN CN2011103413360A patent/CN102413378A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101808244A (en) * | 2010-03-24 | 2010-08-18 | 北京邮电大学 | Video transmission control method and system |
| CN102075312A (en) * | 2011-01-10 | 2011-05-25 | 西安电子科技大学 | Video service quality-based hybrid selective repeat method |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108010533A (en) * | 2016-10-27 | 2018-05-08 | 北京酷我科技有限公司 | The automatic identifying method and device of voice data code check |
| CN107547914A (en) * | 2017-08-15 | 2018-01-05 | 浙江工业大学 | The substandard video segments based on KNN Q study of DASH obtain optimization method |
| CN109726811A (en) * | 2017-10-27 | 2019-05-07 | 谷歌有限责任公司 | Use priority formation neural network |
| US11797839B2 (en) | 2017-10-27 | 2023-10-24 | Google Llc | Training neural networks using priority queues |
| CN107808004A (en) * | 2017-11-15 | 2018-03-16 | 北京百度网讯科技有限公司 | Model training method and system, server, storage medium |
| CN107992939A (en) * | 2017-12-06 | 2018-05-04 | 湖北工业大学 | Cutting force gear working method is waited based on depth enhancing study |
| CN108063961A (en) * | 2017-12-22 | 2018-05-22 | 北京联合网视文化传播有限公司 | A kind of self-adaption code rate video transmission method and system based on intensified learning |
| CN108063961B (en) * | 2017-12-22 | 2020-07-31 | 深圳市云网拜特科技有限公司 | Self-adaptive code rate video transmission method and system based on reinforcement learning |
| CN109218083A (en) * | 2018-08-27 | 2019-01-15 | 广州爱拍网络科技有限公司 | A kind of voice data transmission method and device |
| CN109218083B (en) * | 2018-08-27 | 2021-08-13 | 广州猎游信息科技有限公司 | Voice data transmission method and device |
| CN110006486A (en) * | 2019-04-01 | 2019-07-12 | 中清控(武汉)科技有限公司 | A kind of intelligence flow temperature acquisition control module and intelligence flow temperature measurement method |
| CN112511482A (en) * | 2019-09-16 | 2021-03-16 | 华为技术有限公司 | Media data transmission method, device and system |
| CN112532349A (en) * | 2020-11-24 | 2021-03-19 | 广州技象科技有限公司 | Data processing method and device based on decoding abnormity |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102413378A (en) | Adaptive neural network-based lost packet recovery method in video transmission | |
| CN101808244B (en) | Video transmission control method and system | |
| WO2021208771A1 (en) | Reinforced learning method and device | |
| US20180307979A1 (en) | Distributed deep learning using a distributed deep neural network | |
| CN112491714B (en) | Intelligent QoS route optimization method and system based on deep reinforcement learning in SDN environment | |
| CN110977967A (en) | Robot path planning method based on deep reinforcement learning | |
| Wei et al. | TRUST: A TCP throughput prediction method in mobile networks | |
| Mei et al. | Realtime mobile bandwidth prediction using lstm neural network | |
| CN107179700A (en) | A kind of intelligent home control system and method based on Alljoyn and machine learning | |
| Huang et al. | Learning tailored adaptive bitrate algorithms to heterogeneous network conditions: A domain-specific priors and meta-reinforcement learning approach | |
| Kamyod | End-to-end reliability analysis of an IoT based smart agriculture | |
| Muccini et al. | A machine learning-driven approach for proactive decision making in adaptive architectures | |
| CN113422751B (en) | Streaming media processing method and device based on online reinforcement learning and electronic equipment | |
| CN113748460A (en) | Bandwidth extension of incoming data using neural networks | |
| Hu et al. | Intelligent decision making framework for 6G network | |
| CN112269729A (en) | Intelligent load analysis method for large-scale server cluster of online shopping platform | |
| CN113761148A (en) | Conversation information acquisition method, device, equipment and storage medium | |
| CN113259657A (en) | DPPO code rate self-adaptive control system and method based on video quality fraction | |
| CN114024906B (en) | Flow control method, device and system for optimizing video transmission user experience quality | |
| CN115022684B (en) | Video stream self-adaptive transmission method based on deep reinforcement learning under QUIC protocol | |
| US20230336739A1 (en) | Rate control machine learning models with feedback control for video encoding | |
| CN116414951A (en) | Intelligent dialogue method, model training method, device, storage medium and equipment | |
| RU2447494C1 (en) | Intelligent controller with self-modification rules of educational and control neural networks | |
| Danner et al. | Decentralized machine learning using compressed push-pull averaging | |
| WO2023150943A1 (en) | Method for updating wireless channel model, and apparatus, device and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C12 | Rejection of a patent application after its publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20120411 |