CN102724543B - HMM-based method for implementing media quality analysis evaluation control in IP network - Google Patents

Abstract

The invention relates to an HMM (Hidden Markov Model)-based method for implementing media quality analysis evaluation control in an IP (Internet Protocol) network, which belongs to the field of IP network technology. According to the method, the media quality analysis system firstly utilizes the HMM learning to generate the state transition matrix, real-timely monitor the quality parameters of the media in the IP network and acquire parameters queues of each of the layers in the IP network to generate observing state queues of each of the layers, then decodes the observing state queues of each of the layers to acquire layers hidden observing status queues to be served as the media quality ratings of each of the layers, and finally, performs weighted average on the media quality ratings of each of the layers to acquire media quality analysis results queues in the IP network. By adopting the method provided by the invention, the current media quality parameters model is changed into the observing state, then the HMM is utilized to decode and acquire the most possible hidden state to be served as the media quality analysis evaluation result, so that the subjective evaluation of media quality from a user can be greatly reflected in the situation that taking the objective media quality parameters into account is ensured.

Description

In IP network, based on HMM, realize the method that media quality analysis and evaluation is controlled

Technical field

The present invention relates to IP network technical field, particularly IP network media quality assessment technology field, specifically refers in a kind of IP network and realizes based on HMM the method that media quality analysis and evaluation is controlled.

Background technology

Along with constantly improving and the development of internet, applications of the Internet, the resource-sharing of carrying out on the Internet carrier and information interchange mode gradually from single text mode to sound, look, picture, the polynary mode transition of text.Under this background, the audio/video information amount of the Internet bearer is increasing.And this carrier of the Internet itself has uncertainty and unreliability in transfer of data, often complicated network environment brings the mistake, damage of transmission data and shared resource, the internet use person such as imperfect not to wish the result of seeing.

For traditional text interactive mode, common network quality diagnostic tool is enough, by diagnosing some network performance parameters, coordinate summary, quality condition that algorithm that the data integrity such as signature is verified can be diagnosed out object text resource (only having complete, disappearance, mistake, these four kinds of situations of redundancy on ordinary meaning).Yet for multimedia resource, above-mentioned quality diagnosis instrument just cannot be satisfied the demand, reason is as follows:

1, the data volume of multimedia resource is very large, and possesses not equity, and in a large amount of data, a part is relatively important, and another part is relatively less important.

2, multimedia resource is high to requirement of real-time, not take the complete resource of collecting as successfully identifying, and also will add this concept in time.

3, the performance that multimedia resource is used is also subject to the restriction of local resource, for a simple example, the same video file the possibility of result of different player plays is completely different, and same in the different terminal of performance, the embodiment of multimedia resource also has a lot of individual differences.

4, audient's viewed status impact in the very large meaning of the quality of multimedia quality, same media resource may allow same person watch twice drawn result also not identical.

Because above reason makes the quality analysis more complicated of network audio-video resource.

In order to address this problem, a large amount of work and assessment have been done by standardization body, and credit rating when audio and video resources is play is stipulated.According to GB/T7401-1987 standard, video quality generally gives a mark to describe by 1-5 grade in macroscopic view, and 5 are divided into the highlyest, and 1 is divided into minimumly, and this is widely accepted.Be respectively:

1, the completely not discernable scoring 5 of infringement;

2, damage discernable, single not horrible scoring 4;

3, damage discernablely, make us slightly disliking scoring 3;

4, damage discernable, horrible scoring 2;

5, damage discernablely, cannot stand scoring 1 completely.

Existing media resource mass analysis method, mainly comprises following several:

1, parameter threshold analytical method

By calculating the various parameters in audio video transmission, reception, playing process, and each parameter threshold value is under normal circumstances carried out to record, when changing parameter exceeding normality threshold when media play, presentation medium mass formation impact, the number of anomaly parameter is evaluated the credit rating of media resource.This method can also be expanded, and obviously for different parameters, its impact that exceeds that threshold value brings is not identical, for same parameter, exceeds impact that the side-play amount of threshold value causes obviously also not necessarily linear.This just need to assess parameters weight, and the influence curve of each parameter is carried out to modeling makes algorithm reasonable.

Obviously, this mode is partial to objective indicator.More famous example (be also that relevant parameter is less, and extensively approved) be G.1070 video quality model of ITU-T.This model, for the application of different video and audio standards, has reasonable applicability to other network flow-mediums from video conference to IPTV.This model is represented by formula below, has considered coding and Internet Transmission simultaneously:

$\mathrm{Vq}=1+I_{\mathrm{coding}}\exp (-\frac{P_{\mathrm{plv}}}{D_{p_{\mathrm{plv}}}});$

P wherein _plvrepresent Network Packet Loss, represent video and the tolerance index of service to packet loss, I _codingother data damage of presentation code level.

2, subjective evaluation method

Select a certain amount of audient colony, to media resource, the quality condition of media resource is obtained in the assessment of giving a mark.

This mode is generally by formulating the media play form of standard, and the media audiences of standard experience mode, select various dissimilar audient colonies, and repeatedly test marking is asked for expectation the quality of media resource is made to judge.Obviously this is a kind of subjective assessment mode, belongs to human body sense organ (Human component) part.

3, parameter analysis and subjective assessment combine

This mode can regard that above-mentioned two kinds are passed judgment on the comprehensive of mode as, and the quality conclusion of coming combination parameter mode and subjective assessment mode to draw by certain logic, finally draws a unified mass parameter.

Obviously above-mentioned three kinds of methods respectively have superiority, but have some defects

Parameter analytic approach has the preciseness in mathematical meaning, utilizes a rigorous mathematic(al) representation by media quality marking value of the interrelated generation of parameter.Meanwhile, computational methods are also simple, by objective index calculating parameter, by parameter substitution expression formula, have good execution efficiency.But this is based upon on the basis of rigorous mathematic(al) representation.With regard to some current results of study, also without any mathematic(al) representation can comprehensively contain from media resource and receive the various parameters among broadcasting, backward is considered, immediately there is such mathematical expression, also be difficult to its proof reasonability, containing the more more difficult quilt of expression formula of multi-parameter extensively approves, it is also (between each parameter, also have various forms of associations, the independent parameter collection in its meaning is proved to be to find) of exponential form that parameter increase simultaneously causes the rising of expression formula complexity.Finally, the method is also difficult to solve the 4th of above-mentioned media quality analytical challenge---audient's viewed status impact in the very large meaning of quality of multimedia quality, this problem.

The result of subjective assessment mode is the most correct, by rational flow process, control, a large amount of audient's marking and the statistical method of science can be good at reflecting the quality of media itself, but the cost that the method is paid is too large, a resource will be experienced for several times in the circulation of audient colony, and huge the method that makes of the performance in efficiency and cost can only be evaluated and cannot promote the very important media resource of minority.

The mode that both are merged all can run into defect above-mentioned aspect parameter analysis and subjective assessment, therefore no matter and owing to taking into account the analysis of two aspects, be that the collection of parameter or the collection of subjective suggestion are compared and independent discussed middle mode and can simplify to some extent (process that generally can simplify subjective evaluation method reaches the object that reduces cost).

Markov model (Markov Model, or title " Markov model ") is a generate pattern system based on probability, and so-called generate pattern i.e. a time-discrete n status state machine, and output is switched between n state.

Generate pattern is divided again deterministic model and non-deterministic model, and Markov model belongs to the latter, is used for describing non-deterministic generate pattern.Define a n rank markoff process and be the process that transfer between state only depends on a front n state.If utilize Markov model to describe media quality analytic process, it is single order markoff process.

Hidden markov model (Hidden Markov Model, HMM, or title " hidden Markov model ") is on the basis of Markov model, to have considered another situation.The state of output might not be observed, experience by directly perceived in other words by intuitivism apprehension, and it is a kind of hidden state likely, is not easy perceived.

Hidden markov model expands to hidden state by the state in Markov model for this reason, has defined a new concept: observation state simultaneously.

Hidden state and observation state also do not require correspondence one by one, and observation state has represented the state set that is easy to be understood.But between observation state and hidden state, have relation,, under a hidden state, observation state can embody the regularity on some probability.

A single order hidden markov model is defined as a tlv triple on mathematics: (π, A, B).Wherein:

Π=(π _i): initialization probability vector;

A=(ai _j): state-transition matrix ;

B=(b _ij): confusion matrix Pr (y _i| x _j);

π is init state vector, and A is state-transition matrix, represents to be switched to i shape probability of state from i-1 state, and B is confusion matrix, and representative is the probability that under Y, observation state is X at hidden state.

HMM generally has three class application

1, assessment: a known HMM, assess the probability that one group of observation state queue occurs;

2, decoding: for a known observation state queue, and a definite HMM, draw most possible hidden state queue;

3, study: by a large amount of observation queues, learn to generate a HMM itself.

HMM adopts forward direction algorithm to solve evaluation problem.

Under a definite observation state, at t, the probability in hidden state i is called as the t local probability of state i constantly constantly, with t (j), represents.

The paths of the constantly all sensings of t (j)=Pr (observation state | hidden state j) * Pr(t j state).

When t=1, without any the path of pointing to current state.Therefore be positioned at the probability of current state during t=1, be initial probability, i.e. Pr (state|t=1)=Pr (state), therefore, the initial probability that local probability during t=1 equals current state is multiplied by relevant observation probability, that is: wherein π is initial vector, for observation state under hidden state j in confusion matrix equals k _t+1probability.Then the local probability of the t moment three kinds of states is all calculated, and the local probability of t+1 moment state j is that available following recurrence formula calculates so:

${\mathrm{\α}}_{t+1}\left(j\right)=b_{{\mathrm{jk}}_{t+1}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_t\left(i\right)a_{\mathrm{ij}};$

Wherein, (a _ij) be state-transition matrix, (b _ij) be confusion matrix, α _t+1(j) be the t+1 local probability of hidden state j constantly, n ∈ N.

Asking for forward variable of backward variable algorithm is similar, is recursion.Forward variable is used initial vector as the parameter probability valuing of t=1, backward variable-definition T is (0<t<=T) constantly, the backward variable of all hidden states is 1, and for hidden state queue arbitrarily, last state is accomplished fact.Inverse algorithms is the reverse deduction of forward direction algorithm.

Summary of the invention

The object of the invention is to have overcome above-mentioned shortcoming of the prior art, a kind of HMM that utilizes is provided, introduce observation state, and the current media quality parameter model of IP network is turned to observation state, by decoding, obtain most possible hidden state as media quality analysis and evaluation result, thereby can be in the situation that guaranteeing to consider objective media quality parameter, farthest embody user for the subjective assessment of media quality, and application mode is easy, with low cost, range of application also realizes based on HMM the method that media quality analysis and evaluation is controlled in IP network comparatively widely.

In order to realize above-mentioned object, in IP network of the present invention, based on HMM, realize method that media quality analysis and evaluation the controls media quality analytical system based in access IP network, the method comprises the following steps:

(0) described media quality analytical system utilizes HMM study to generate state-transition matrix;

(1) mass parameter of media in the IP network described in described media quality analytical system Real-Time Monitoring, the parameter queue of obtaining each layer in IP network;

(2) described media quality analytical system gathers each obtained layer parameter queue respectively, produces the observation state queue of each layer;

(3) state-transition matrix of the HMM described in described media quality analytical system utilization is decoded to the observation state queue of each layer, obtains the hidden state queue of each layer as the media quality scoring of this layer;

(4) described media quality analytical system is weighted the media quality scoring of each layer on average, obtains the queue of IP network media quality analysis result, thereby completes the operation of media quality analysis and evaluation.

2, in IP network according to claim 1, based on HMM, realize the method that media quality analysis and evaluation is controlled, it is characterized in that, described media quality analytical system utilizes HMM study to generate state-transition matrix, is specially:

Described media quality analytical system utilizes the backward Algorithm Learning of HMM to generate state-transition matrix.

3, in IP network according to claim 2, based on HMM, realize the method that media quality analysis and evaluation is controlled, it is characterized in that, described backward algorithm specifically comprises the following steps:

(01) described media quality analytical system init state transfer matrix, makes the t=T backward variable β of all states constantly _t(i) be 1, i.e. β _t(i)=1,1≤i≤N;

(02) described media quality analytical system is according to each time point t=T-1 of following formula recursive calculation, T-2 ..., the backward variable β of 1 o'clock _t(i),

${\mathrm{\&beta;}}_t\left(i\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right),t=T-1,T-2,...,\mathrm{1,1}\&le;i\&le;N$

A _ijfor state-transition matrix, b _j(O _t+1) equal O for observation sequence under hidden state j in confusion matrix _t+1probability, O _t+1for t+1 observation sequence constantly.

In this IP network, based on HMM, realize in the method for media quality analysis and evaluation control, the state-transition matrix of the HMM described in described media quality analytical system utilization is decoded to the observation state queue of each layer, the hidden state queue of obtaining each layer, is specially:

Described media quality analytical system utilizes HMM dimension bit algorithm to decode to the observation state queue of each layer, obtains the hidden state queue of each layer.

In this IP network, based on HMM, realize in the method for media quality analysis and evaluation control, the described hidden state queue that utilizes HMM dimension bit algorithm to decode and obtain each layer the observation state queue of each layer, is specially:

In described HMM dimension bit algorithm, the observation state queue decoding that utilizes following formula to pass through each layer obtains the hidden state queue of each layer:

${\mathrm{\&delta;}}_t\left(i\right)\underset{j}{\max }\left({\mathrm{\&delta;}}_{t-1}\left(j\right)a_{\mathrm{ij}}{b_{\mathrm{ik}}}_t\right);$

Wherein, (a _ij) be state-transition matrix, (b _ij) be confusion matrix, δ _t(i) for t arrives probability maximum in all sequences probability of hidden state i constantly, i.e. local path probability, for observation state under hidden state i in confusion matrix equals k _tprobability, max is for getting maximum.

In this IP network, based on HMM, realize in the method for media quality analysis and evaluation control, described method is further comprising the steps of afterwards in described step (4):

(5) described media quality analytical system is optimized the state-transition matrix (a in described HMM according to the hidden state queue of the observation state queue of each described layer and each layer _ij).

In this IP network, based on HMM, realize in the method that media quality analysis and evaluation controls the state-transition matrix (a in the HMM described in described optimization _ij), be specially:

Utilize the state-transition matrix (a in the HMM described in following formula optimization _ij):

=(from hidden state S _ito hidden state S _jcarry out the desired value of state switching)/(from hidden state S _icarry out the desired value of state switching),

That is, ${\stackrel{\&OverBar;}{a}}_{\mathrm{ij}}=\frac{\underset{t=1}{\overset{T-1}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t(i,j)}{\underset{t=1}{\overset{T-1}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t\left(i\right)}$

Wherein, t is positioned at hidden state S constantly _iprobability variable γ _t(i) be:

γ _t(i)=P(q _t=S _i|O,λ)，

? ${\mathrm{\&gamma;}}_t\left(i\right)=\frac{{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)}{P\left(O\right|\mathrm{\&lambda;})}=\frac{{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)}{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)},$

Wherein O is observation sequence, and λ is HMM, α _t(i) be the t local probability of hidden state i constantly, β _t(i) be t backward variable constantly;

And t is positioned at hidden state S constantly _iand t+1 is positioned at hidden state S constantly _jprobability variable ξ _t(i, j) is:

ξ _t(i,j)=P(q _t=S _i,q _t+1=S _j|O,λ)，

That is, ${\mathrm{\&xi;}}_t(i,j)=\frac{{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}{P\left(O\right|\mathrm{\&lambda;})},$

${\mathrm{\&xi;}}_t(i,j)=\frac{{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}.$

And described t is positioned at hidden state S constantly _iand t+1 is positioned at hidden state S constantly _jprobability variable ξ _t(i, j) and described t are positioned at hidden state S constantly _iprobability variable γ _t(i) meet following equation:

${\mathrm{\&gamma;}}_t\left(i\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t(i,j).$

In this IP network, based on HMM, realize in the method for media quality analysis and evaluation control, the level in described IP network comprises: network layer, media delivery key-course, media services layer, media encapsulated layer, media coding layer and media content layer.The hidden state queue of each described layer, is specially: the scoring of network layer media quality, the scoring of media delivery key-course media quality, the scoring of media services layer media quality, the scoring of media encapsulated layer media quality, the scoring of media coding layer media quality and the scoring of media content layer media quality.

Adopted in the IP network of this invention and realized based on HMM the method that media quality analysis and evaluation is controlled, first media quality analytical system utilizes HMM study to generate state-transition matrix, the mass parameter of media in IP network described in Real-Time Monitoring, obtain the parameter queue of each layer in IP network, and each layer parameter queue is gathered respectively, produce the observation state queue of each layer; Then utilize HMM to decode to the observation state queue of each layer, obtain the hidden state queue of each layer as the media quality scoring of this layer; Finally the media quality scoring of each layer is weighted on average, obtains the queue of IP network media quality analysis result, thereby complete the operation of media quality analysis and evaluation.The method turns to observation state by the current media quality parameter model of IP network, and then utilize HMM decoding to obtain most possible hidden state as media quality analysis and evaluation result, thereby can be in the situation that guaranteeing to consider objective media quality parameter, farthest embody user for the subjective assessment of media quality, and it is easy in IP network of the present invention, based on HMM, to realize the method application mode that media quality analysis and evaluation controls, with low cost, range of application is also comparatively extensive.

Accompanying drawing explanation

Fig. 1 realizes the flow chart of steps of the method for media quality analysis and evaluation control based on HMM in IP network of the present invention.

Fig. 2 utilizes method of the present invention to realize the process schematic diagram that media quality analysis and evaluation is controlled in practical application.

Fig. 3 is the method for the present invention confusion matrix schematic diagram of relation between presentation medium quality score and observation state packet loss in actual applications.

Embodiment

In order more clearly to understand the technology page of the present invention, especially exemplified by following examples, describe in detail.

Refer to shown in Fig. 1, for realize the flow chart of steps of the method for media quality analysis and evaluation control in IP network of the present invention based on HMM.

The media quality analytical system of method of the present invention based in access IP network, in one embodiment, the method comprises the following steps:

(0) described media quality analytical system utilizes HMM study to generate state-transition matrix;

(1) mass parameter of media in the IP network described in described media quality analytical system Real-Time Monitoring, the parameter queue of obtaining each layer in IP network;

(2) described media quality analytical system gathers each obtained layer parameter queue respectively, produces the observation state queue of each layer;

(3) state-transition matrix of the HMM described in described media quality analytical system utilization is decoded to the observation state queue of each layer, obtains the hidden state queue of each layer as the media quality scoring of this layer;

(4) described media quality analytical system is weighted the media quality scoring of each layer on average, obtains the queue of IP network media quality analysis result, thereby completes the operation of media quality analysis and evaluation.

In a kind of more preferably execution mode, described media quality analytical system utilizes HMM study to generate state-transition matrix, is specially: described media quality analytical system utilizes the backward Algorithm Learning of HMM to generate state-transition matrix.This backward algorithm specifically comprises the following steps:

(01) described media quality analytical system init state transfer matrix, makes the t=T backward variable β of all states constantly _t(i) be 1, i.e. β _t(i)=1,1≤i≤N;

(02) described media quality analytical system is according to each time point t=T-1 of following formula recursive calculation, T-2 ..., the backward variable β of 1 o'clock _t(i),

${\mathrm{\&beta;}}_t\left(i\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right),t=T-1,T-2,...,\mathrm{1,1}\&le;i\&le;N$

A _ijfor state-transition matrix, b _j(O _t+1) equal O for observation sequence under hidden state j in confusion matrix _t+1probability, O _t+1for t+1 observation sequence constantly.

In a kind of more preferably execution mode, the state-transition matrix of the HMM described in described media quality analytical system utilization is decoded to the observation state queue of each layer, obtains the hidden state queue of each layer, is specially:

Described media quality analytical system utilizes HMM dimension bit algorithm to decode to the observation state queue of each layer, obtains the hidden state queue of each layer.

In a kind of further preferred embodiment, the described hidden state queue that utilizes HMM dimension bit algorithm to decode and obtain each layer the observation state queue of each layer, is specially:

In described HMM dimension bit algorithm, the observation state queue decoding that utilizes following formula to pass through each layer obtains the hidden state queue of each layer:

${\mathrm{\&delta;}}_t\left(i\right)\underset{j}{\max }\left({\mathrm{\&delta;}}_{t-1}\left(j\right)a_{\mathrm{ij}}{b_{\mathrm{ik}}}_t\right);$

Wherein, (a _ij) be state-transition matrix, (b _ij) be confusion matrix, δ _t(i) for t arrives probability maximum in all sequences probability of hidden state i constantly, i.e. local path probability, for observation state under hidden state i in confusion matrix equals k _tprobability, max is for getting maximum.

At another kind, more preferably in execution mode, described method is further comprising the steps of afterwards in described step (4):

(5) described media quality analytical system is optimized the state-transition matrix (a in described HMM according to the hidden state queue of the observation state queue of each described layer and each layer _ij).

In a kind of further preferred embodiment, the state-transition matrix (ai in the HMM described in described optimization _j), be specially:

Utilize the state-transition matrix (a in the HMM described in following formula optimization _ij):

=(from hidden state S _ito hidden state S _jcarry out the desired value of state switching)/(from hidden state S _icarry out the desired value of state switching),

That is, ${\stackrel{\&OverBar;}{a}}_{\mathrm{ij}}=\frac{\underset{t=1}{\overset{T-1}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t(i,j)}{\underset{t=1}{\overset{T-1}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t\left(i\right)}$

Wherein, t is positioned at hidden state S constantly _iprobability variable γ _t(i) be:

γ _t(i)=P(q _t=S _i|O,λ)，

? ${\mathrm{\&gamma;}}_t\left(i\right)=\frac{{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)}{P\left(O\right|\mathrm{\&lambda;})}=\frac{{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)}{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)},$

Wherein O is observation sequence, and λ is HMM, α _t(i) be the t local probability of hidden state i constantly, β _t(i) be t backward variable constantly;

And t is positioned at hidden state S constantly _iand t+1 is positioned at hidden state S constantly _jprobability variable ξ _t(i, j) is:

ξ _t(i,j)=P(q _t=S _i,q _t+1=S _j|O,λ)，

That is, ${\mathrm{\&xi;}}_t(i,j)=\frac{{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}{P\left(O\right|\mathrm{\&lambda;})},$

${\mathrm{\&xi;}}_t(i,j)=\frac{{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}.$

And described t is positioned at hidden state S constantly _iand t+1 is positioned at hidden state S constantly _jprobability variable ξ _t(i, j) and described t are positioned at hidden state S constantly _iprobability variable γ _t(i) meet following equation:

${\mathrm{\&gamma;}}_t\left(i\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t(i,j).$

In a kind of preferred execution mode, the level in described IP network comprises: network layer, media delivery key-course, media services layer, media encapsulated layer, media coding layer and media content layer.The hidden state queue of each described layer, is specially: the scoring of network layer media quality, the scoring of media delivery key-course media quality, the scoring of media services layer media quality, the scoring of media encapsulated layer media quality, the scoring of media coding layer media quality and the scoring of media content layer media quality.

In actual applications, HMM has three elements: initial vector π, state switching matrix A and confusion matrix B.First state probability of initial vector π representative model wherein, in theory, in a long time span, no matter initial vector why, the output sequence of Hidden Markov all can tend towards stability (with observation state contact to some extent stable, but not absolute stability), such according to general model, when detecting some media resource quality, always can give tacit consent to it is, might as well assumed initial state be therefore vector [0,0,0,0,1], representative scoring is that 5 probability is 100%, be that initial hidden state is definite scoring 5, optimum state.Confusion matrix B has represented under each hidden state, observation state occurrence probability, and Fig. 3 has provided a very simple confusion matrix, is the relational matrix between media quality score (1-5) and observation state packet loss.When wherein the first row represents to be 5 when marking, packet loss state is necessarily in without packet loss (without packet loss state probability 100%).The every a line sum of this matrix is necessary for 1, and each row not requiring.

Relation between any one parameter and media quality score can represent with similar confusion matrix.Confusion matrix can obtain with subjective estimate method.First the observation parameter in this model is sorted out, generate corresponding state, such as example above, can define packet loss=0 for without packet loss state, packet loss <0.2% and be defined as on a small quantity for a long time packet loss etc. in continual data package dropout state (a front n period is all in packet loss state).Then by the media resource under various observation states, the mode of operating specification, selects the audient of some repeatedly to experience, and records each audient scoring given to the media resource of each state.Then assessment result is done and gathered, in statistics total experience result, the state that must be divided into 5 media article can obtain the first row of above-mentioned matrix, and same mode, can be perfect by whole confusion matrix.

Subjective estimate method has been used in the generation of confusion matrix, guarantees that this matrix is believable, can represent (by selecting rational audient's group can make result infinite approach normal person's evaluation result) of audient's perception.

Finally consider the problem of lower state-transition matrix, in HMM, this matrix cannot be with a scene description of easily thinking about it, because it is inherited from Markov model, and Markov model thinks that to meet the things state of its model only relevant with several states before this things, this obviously and video quality analyze this process and be not inconsistent.

In fact, HMM, by introducing observation variable, has made state switching matrix change under objective reality and the impact of measurable observation variable, it can be interpreted as to a kind of statistical probability set after objective variable impact and be applied.

In specific implementation process of the present invention, clearly, five states of 1-5, represent respectively the scoring to media quality to the hidden state in HMM, need to observe vectorial modelling simultaneously.

Use one group to observe vector state collection and obviously have weak point, in the process of media play, the parameter major part obtaining be quantitative such as: packet loss unit is %, and network jitter unit be millisecond etc.After quantitative parameter discrete, each variable can generate a state set, such as above-mentioned packet loss can be divided into 5 states.And combinations of states of multi-parameter so will be a very huge set, and such as each parameter has 4 states, 10 parameter combinations just have 4 ten power kind states, and this obviously can cause the index of observation state to expand, and is that algorithm becomes in efficiency infeasible.

The method solving is that parameter is sorted out, and merges, and forms several important states.

Media resource transmitting procedure in analyzing IP network, can utilize existing network hierarchy model that parameter layering is divided:

One, network layer, parameter comprises:

1, IP message dropping rate

2, IP message transmissions time delay

3, IP message arrive shake

4, TCP number of retransmissions

5, TCP connect replacement number of times

Two, media delivery key-course, parameter comprises:

1, RTP message dropping rate

2, RTP message arrive time delay

3, RTP message arrive shake

4, RTP packet out-ordering rate

5, RTP message retransmission rate

6, RTSP log-on count

7, RTSP eartbeat interval

Three, media services layer, parameter comprises:

1, MDI-df (the time delay factor)

2, MDI-MLR (MDI packet loss)

3, MLT-15 (past 15 minutes MLR statistics)

4, MLT-24 (past 24 hours MLR statistics)

5, the asynchronous counting of audio frequency and video

Four, media encapsulated layer, parameter comprises:

1, the preliminary rate of audio-visual synchronization

2, video I frame loss ratio

3, video B frame loss ratio

4, video P frame loss ratio

5, audio frame loss ratio

6, gop structure

7, container buffer memory (or playing buffer memory) overflow number of times

8, container buffer memory (or playing buffer memory) underflow number of times

Five, media coding layer, parameter comprises:

1, video resolution

2, video frame rate/field rate

3, PSNR (channel Y-PSNR)

4, dct transform matrix size

5, coded quantization step-length

6, macroblock size

7, syntactic analysis mistake

Six, media content layer, parameter comprises:

1, static ratio

2, Fuzzy Exponential

3, blocking effect index

4, Smoothness Index

Below from bottom to up the design parameter that can detect in the IP network transmission of media resource has been done to a division, then need the parameter of every one deck to gather, demonstrate a state set, can be designed as every one deck and be divided into 10 states (5 twice guarantees to reach requirement at observation state dividing precision).Such as 3 parameters of network layer, each quantizes to be divided into 10 grades, and packet loss 0 is grade 1, and packet loss is grade 2 etc. below 0.1%, and the state grade of each gain of parameter is got to maximum (grade of worst) as this layer state grade output.Obtain thus 6 observation state set, in each observation state set, had 10 states.

According to 6 groups of observation state set, set up respectively 6 hidden markov models subsequently.When carrying out media quality analysis, in each detection time, in granularity, the media quality objective parameter recording is aggregated into 6 momentary status values, here define again a time value, be called the sense cycle time, under test environment, minimum detection granularity is configured to 1 second, and sense cycle is set to 5 seconds, obviously the input of each sense cycle is 6 observation state queues, and each queue length is 5.Each is observed in the Markov model that queue is input to respective layer and calculates the maximum probability hidden state queue under every one deck objective parameter impact, afterwards these output state queues are weighted to average operation, obtain meeting the Media Analysis result that real user is experienced.

According to realize the process of the method for media quality analysis and evaluation control in the IP network of the present invention shown in Fig. 2 based on HMM, each HMM is wherein independently, divide for 6 independent models, if the parameter monitoring difficulty to some levels wherein, or the media quality of confirming certain one deck is intact certainly, can remove the Markov model computing of certain one deck, acquired results cannot embody the impact on media quality result of certain layer parameter of removing.

Certainly, same, also can self-defined new argument, by algorithm above, carry out the foundation of Markov model, the given weighting weight that newly adds parameter set, expands this media quality analytical system simultaneously.

The advantage that method of the present invention is carried out the control of media quality analysis and evaluation is:

1, media play objective indicator and subjective analysis result are combined during quality analysis, taken into full account both features.

2, the disposable subjective analysis result of utilizing is carried out modeling, can reuse efficiently.

3, model has learning ability, and only needing abundant study model can be evolved with observation sequence sample is a more rational state.

4, system itself can be expanded and cutting, adapts to various application needs.

Adopted in the IP network of this invention and realized based on HMM the method that media quality analysis and evaluation is controlled, first media quality analytical system utilizes HMM study to generate state-transition matrix, the mass parameter of media in IP network described in Real-Time Monitoring, obtain the parameter queue of each layer in IP network, and each layer parameter queue is gathered respectively, produce the observation state queue of each layer; Then utilize HMM to decode to the observation state queue of each layer, obtain the hidden state queue of each layer as the media quality scoring of this layer; Finally the media quality scoring of each layer is weighted on average, obtains the queue of IP network media quality analysis result, thereby complete the operation of media quality analysis and evaluation.The method turns to observation state by the current media quality parameter model of IP network, and then utilize HMM decoding to obtain most possible hidden state as media quality analysis and evaluation result, thereby can be in the situation that guaranteeing to consider objective media quality parameter, farthest embody user for the subjective assessment of media quality, and it is easy in IP network of the present invention, based on HMM, to realize the method application mode that media quality analysis and evaluation controls, with low cost, range of application is also comparatively extensive.

In this specification, the present invention is described with reference to its specific embodiment.But, still can make various modifications and conversion obviously and not deviate from the spirit and scope of the present invention.Therefore, specification and accompanying drawing are regarded in an illustrative, rather than a restrictive.

Claims (5)

1. in IP network, based on HMM, realize the method that media quality analysis and evaluation is controlled, the media quality analytical system of the method based in access IP network, is characterized in that, described method comprises the following steps:

(0) described media quality analytical system utilizes HMM study to generate state-transition matrix;

Described media quality analytical system utilizes HMM study to generate state-transition matrix, is specially:

Described media quality analytical system utilizes the backward Algorithm Learning of HMM to generate state-transition matrix;

Described backward algorithm specifically comprises the following steps:

(01) described media quality analytical system init state transfer matrix, makes the t=T backward variable β of all states constantly _t(i) be 1, i.e. β _t(i)=1,1≤i≤N;

(02) described media quality analytical system is according to each time point t=T-1 of following formula recursive calculation, T-2 ..., the backward variable β of 1 o'clock _t(i),

${\mathrm{\&beta;}}_t\left(i\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right),t=T-1,T-2,...,\mathrm{1,1}\&le;i\&le;N$

A _ijfor state-transition matrix, b _j(O _t+1) equal O for observation sequence under hidden state j in confusion matrix _t+1probability, O _t+1for t+1 observation sequence constantly;

(1) mass parameter of media in the IP network described in described media quality analytical system Real-Time Monitoring, the parameter queue of obtaining each layer in IP network, the level in described IP network comprises: network layer, media delivery key-course, media services layer, media encapsulated layer, media coding layer and media content layer;

(2) described media quality analytical system gathers each obtained layer parameter queue respectively, produces the observation state queue of each layer;

(3) state-transition matrix of the HMM described in described media quality analytical system utilization is decoded to the observation state queue of each layer, obtains the hidden state queue of each layer as the media quality scoring of this layer;

Described media quality analytical system utilizes the state-transition matrix of HMM to decode to the observation state queue of each layer, obtains the hidden state queue of each layer, is specially:

Described media quality analytical system utilizes HMM dimension bit algorithm to decode to the observation state queue of each layer, obtains the hidden state queue of each layer;

The described hidden state queue that utilizes HMM dimension bit algorithm to decode and obtain each layer the observation state queue of each layer, is specially:

In described HMM dimension bit algorithm, the observation state queue decoding that utilizes following formula to pass through each layer obtains the hidden state queue of each layer:

${\mathrm{\&delta;}}_t\left(i\right)\underset{j}{\max }\left({\mathrm{\&delta;}}_{t-1}\left(j\right)a_{\mathrm{ij}}{b_{\mathrm{ik}}}_t\right);$

Wherein, (a _ij) be state-transition matrix, (b _ij) be confusion matrix, δ _t(i) for t arrives probability maximum in all sequences probability of hidden state i constantly, i.e. local path probability, for observation state under hidden state i in confusion matrix equals k _tprobability, max is for getting maximum;

(4) described media quality analytical system is weighted the media quality scoring of each layer on average, obtains the queue of IP network media quality analysis result, thereby completes the operation of media quality analysis and evaluation.

2. in IP network according to claim 1, based on HMM, realize the method that media quality analysis and evaluation is controlled, it is characterized in that, described method is further comprising the steps of afterwards in described step (4):

(5) described media quality analytical system is optimized the state-transition matrix (a in described HMM according to the hidden state queue of the observation state queue of each described layer and each layer _ij).

3. in IP network according to claim 2, based on HMM, realize the method that media quality analysis and evaluation is controlled, it is characterized in that, the state-transition matrix (a in the HMM described in described optimization _ij), be specially:

Utilize the state-transition matrix (a in the HMM described in following formula optimization _ij):

=(from hidden state S _ito hidden state S _jcarry out the desired value of state switching)/(from hidden state S _icarry out the desired value of state switching),

That is, ${\stackrel{\&OverBar;}{a}}_{\mathrm{ij}}=\frac{\underset{t=1}{\overset{T-1}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t(i,j)}{\underset{t=1}{\overset{T-1}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_t\left(i\right)}$

Wherein, t is positioned at hidden state S constantly _iprobability variable γ _t(i) be:

γ _t(i)=P(q _t=S _i|O,λ)，

? ${\mathrm{\&gamma;}}_t\left(i\right)=\frac{{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)}{P\left(O\right|\mathrm{\&lambda;})}=\frac{{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)}{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t\left(i\right){\mathrm{\&beta;}}_t\left(i\right)},$

Wherein O is observation sequence, and λ is HMM, α _t(i) be the t local probability of hidden state i constantly, β _t(i) be t backward variable constantly;

And t is positioned at hidden state S constantly _iand t+1 is positioned at hidden state S constantly _jprobability variable ξ _t(i, j) is:

ξ _t(i,j)=P(q _t=S _i,q _t+1=S _j|O,λ)，

That is, ${\mathrm{\&xi;}}_t(i,j)=\frac{{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}{P\left(O\right|\mathrm{\&lambda;})},$

${\mathrm{\&xi;}}_t(i,j)=\frac{{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t\left(i\right)a_{\mathrm{ij}}{b}_j\left(O_{t+1}\right){\mathrm{\&beta;}}_{t+1}\left(j\right)}.$

4. in IP network according to claim 3, based on HMM, realize the method that media quality analysis and evaluation is controlled, it is characterized in that, described t is positioned at hidden state S constantly _iand t+1 is positioned at hidden state S constantly _jprobability variable ξ _t(i, j) and described t are positioned at hidden state S constantly _iprobability variable γ _t(i) meet following equation:

${\mathrm{\&gamma;}}_t\left(i\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_t(i,j).$

5. in IP network according to claim 1, based on HMM, realize the method that media quality analysis and evaluation is controlled, it is characterized in that, the hidden state queue of described each layer, is specially:

The scoring of network layer media quality, the scoring of media delivery key-course media quality, the scoring of media services layer media quality, the scoring of media encapsulated layer media quality, the scoring of media coding layer media quality and the scoring of media content layer media quality.