Linear modeling for MPEG-4 intra frame decoding complexity prediction based on statistical analysis

Video decoding complexity prediction plays an important role in energy efficient applications, such as dynamic voltage scaling and workload reshaping. This paper presents a novel linear model for MPEG-4 intra frame decoding complexity prediction. Detailed experiments are conducted to exploit the statistical relationship between frame length and decoding complexity for various video contents under different bitrates. The experiments show that decoding complexity is linear related to frame length, the parameters of linear model vary slightly in terms of video sequences and bitrates, and the model parameters for different size video are proportional to the ratio of video size. Based on above principles, the linear model for CIF format video are fitted offline and utilized to predict both CIF and 4CIF format video sequences' intra frame decoding complexity on the fly. The probability density function of prediction error appeared normal distributed and the average prediction error is 0.47%. The maximal prediction error is 2.94% and the runtime overload of the proposed method is 54 cycles/frame on TI TMS320DM642 platform.