Feed-forward Nonlinear Network Approaches for Mpeg Video Rate Prediction

Conventional approaches generally assume that the compressed video has high correlation so that linear predictive methods can be applied. However , for realistic videos such as movies, sports and advertisements, there can be many exceptions since the correlation may be abnormally low. In this paper, we developed a feed-forward network-based rate control scheme which eeec-tively accommodates dramatic variations in video rate by introducing a nonlinear predictive technique. We investigated a radial basis function (RBF) network nonlinear video rate estimator. As a constituent quantiser control method, non-linear functional surfaces were also employed. The performance of the scheme was compared with the MPEG2 Test Model (TM) 5 and was assessed in terms of buuer occupancy, video data rate (bits/frame) and peak signal-to-noise ratio (PSNR). 1 INTRODUCTION An eeective rate control technique becomes more demanding when the video contains a long duration of rapid motion or frequent scene changes. A noteworthy fact for this type of video is that the occurrence of abrupt changes in visual information can hardly be predicted in purely stochastic ways. Hence, as a solution, it is necessary to introduce apriori knowledge to control the buuer and the quantiser in the MPEG video encoder. In this approach, we employed a nonlinear feed-forward predictive scheme in order to exploit short-term correlation of video and to improve the estimation performance. Attention was fo-cused on the RBF network as a nonlinear predictor 1]. A good rate control algorithm will keep the occupancy as steady as possible within the desired delay limit-by controlling the number of coded bits per unit time-with little quality degradation in PSNR. This paper is organised as follows: Sec. 2 gives an overview on the feed-forward video rate estimation approach. Sec. 3 describes the connguration of the nonlin-ear video rate estimation scheme. Sec. 4 discusses the nonlinear functional surfaces for the quantiser control. Sec. 5 includes the simulation results and discussions on the performance. Finally, Sec. 6 concludes this chapter. Any linear combiner can be used as a video rate estima-tor which takes scene change features as its input. First, we tested the linear estimator trained with recursive least squares (RLS) algorithm, which is widely used in the eld of adaptive control. A radial basis function (RBF) network was then adopted as a nonlinear estimator. The RBF network is known to have better estimation performance than linear predictors for non-stationary signals and has recently been used successfully in several …