Research On Video Decoding Complexity Modeling And Prediction

Multimedia applications, such as video decoding, always take heavy computational complexity and cause huge energy consumption. To save energy, state of the art hardware platforms tend to adjust their resource provision according to the actual decoding computational demand. The effectiveness of such methods depends on the prediction accuracy of decoding complexity. Based on the linear relationship between frame length and decoding complexity, this paper proposesFirstly, based on the statistical characteristic of the linear model coefficients, two principles are proposed to guide the establishment of linear model for intra frame:the coefficients of linear model vary slightly in terms of video sequences and bitrates, and the model coefficients for different size video are proportional to the ratio of video size. Then, an offline analysis assisted prediciton method is proposed to predict decoding complexity for intra frame. Experiments are conducted for both MPEG-4and H.264through DSP-based hardware platform and software-based simulation platform. Results show the probability density function of prediction error appeared normal distributed and the maximal prediction error is less than4%for all compressed stream. The runtime overload of the proposed method is54cycles/frame on TI TMS320DM642platform.Then an online predicting method which utilize State-Variable Analysis (SVA) for inter frame decoding complexity prediction is proposed. The state equation of decoding system is established as follows:first, the semantic meanings of state variables which directly affect the linear relationship between frame length and decoding complexity are exploited through analysis on decoding procedure. Then combined with the semantic correlation between neighboring frames in compressed streams, the state equation is established as a piecewise function which can reflect the variation of video content. Considering the different characteristics of state trajectory between Intra frame and Inter frame, we provide online decoding complexity predicting methods for them respectively. The mean values of state variables for intra frame are obtained through offline statistical analysis and then utilized directly during online process to predict decoding complexity. While for inter frame, the values of state variables are calculated using state equation in the online process and then the decoding complexity can be predicted.Finally, to cope with the demand for multimedia streaming system over cloud computing enviroment, algorithm that using kalman tilter to update contrlling parameters online is proposed. Requirements for the output function of decoding system are firstly illustrated from the perspective of compensating the hysteretic nature of Kalman Filter. Then the frame length is theoretically represented as the product of video contents, bitrate controll algorithm and rate distoration optimization strategy, and consequently represent the characteristic of compressed stream. So the linear model between decoding complexity and frame length is certainly make sense. Then based on the infrastucture of kalman filter, the linear model is defined as the output function of decoding system, model coefficients are defined as the state variable. State variables between neighboring frames are assumed identical in terms of the continuity of video sequences and the difference of them is defined as process noise. Using kalamn filter to establish the predicting procedure for decoding complexity predciiton, all the control parameters are calculated through online process. The proposed method is also implemented on both software-based simulation platform and DSP-based hardware platform, the decoding complexity of compressed streams which use either H.264or MPEG-4coding standards is predicted. The experimental results show the proposed method can give very accurate prediction for decoding complexity. The average prediction error is less than1%for intra frames, while it is5%for inter frames. Moreover, since the controlling parameters are updated without any offline process, this proposed method is suitable for multimedia system over cloud computing enviroment where the encoding parameters that create the compressed streams are unknown.