A Fuzzy Neural CBR Channel Rate Controller for MPEG2 Encoders* Maria C. F. De Castro, Fernando C. C. De Castro, Dalton S. Arantes and Dario F. G. Azevedo Faculdade de Engenharia Elétrica e de Computação Departamento de Comunicações Universidade Estadual de Campinas – UNICAMP Caixa Postal 6101 – 13.083-970 – Campinas – SP – BRASIL E-mail: cristina@decom.fee.unicamp.br, decastro@ee.pucrs.br, dalton@decom.fee.unicamp.br, dario@ee.pucrs.br Abstract- A fuzzy algorithm is used as control surface for the buffer occupancy of a MPEG2 (Moving Picture Experts Group) video encoder. Based on scene features, a supervised algorithm trains a Radial Basis Function Neural Network (RBFNN). The so trained RBFNN acts as a predictor for the number of bits generated in a frame, so that the predicted buffer occupancy can be determined. The predicted and present buffer occupancies are applied to the fuzzy-generated control surface which yields the encoder quantizer step parameter. We compare the obtained results with the Test Model 5 standard rate control scheme. 1. Introduction The MPEG-2 standard is a widely accepted video-compression technique that applies no standardization to the coding process, allowing the proposal of different approaches for the encoders. The MPEG coded data can be transmitted through constant bit rate (CBR) channels or through variable bit rate (VBR) channels. In the CBR case, due to the highly time-varying complexity of the video sequences, it is imperative the use of a buffer in order to properly stabilize the output rate to the channel [6]. As an example, perhaps one of the most demanding CBR buffer control tasks occurs in the transmission of medical images, where the requirements for image quality are severe and channel rates as low as 2 Mbps are not uncommon [8]. In a previous work [1] the authors developed a non-linear predictive video bit rate controller that uses a supervised RBFNN approach. The RBFNN training is carried out by updating the output neuron synapse weights, as well as the Gaussian centers, by means of the Stochastic Gradient (SG) algorithm [4]. The SG algorithm uses an error input derived from the desired output, which characterizes it as a supervised training process. The rate prediction at the RBFNN output is then applied to a non-linear mapping [2], which yields the encoder quantizer step (mquant parameter [3]). In this work we present a fuzzy approach to determine the encoder quantizer step. The fuzzy controller implements a family of piecewise lines, each one with a different slope. The line set is conceived to properly control the buffer occupancy for time-varying complexity video signals. The obtained results with such simple rate control technique are similar to those previously obtained by the authors using the referred non-linear mapping technique [1]. In comparison with the Test Model 5 (TM5) standard rate control scheme [3], superior performance was obtained for the buffer occupancy and signal-to-noise ratio (SNR). 2. RBF Neural Networks A RBF neural network is a universal approximator [4]. It consists of an input layer with M nodes, a hidden layer with K neurons and an output neuron layer. In our case, since we are interested in estimating just the scene complexity on a frame by frame basis, the output layer has only one output neuron, as can be seen in Figure 1 [5]. Figure 1: Radial Basis Function Network. * Partially supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) and Pontifícia Universidade Católica do Rio Grande do Sul (PUCRS).