FEEDBACK FREE DVC ARCHITECTURE USING MACHINE LEARNING J.L. Martínez 1 , G. Fernández-Escribano 1 , H. Kalva 2 , W.A.R.J.Weerakkody 3 , W.A.C.Fernando 3 , A.Garrido 1 1 Albacete Research Institute of Informatics. Universidad de Castilla-La Mancha. 02071 Albacete, Spain 2 Department of Computer Science and Engineering, Florida Atlantic University, Boca Raton, FL, USA 3 Center for Communications Systems Research, University of Surrey, GU2 7XH, United Kingdom ABSTRACT Most of the reported Distributed Video Coding (DVC) architectures have a serious limitation that hinders its practical application. The uses of a feedback channel between the encoder and the decoder require an interactive decoding procedure which is a limitation for applications such as offline processing. On the other hand, the decoder needs an efficient way to estimate the probability of error without assuming the availability of the original video at the decoder. In this paper we continue with our previous works into a more practical DVC architecture which solves both problems based on the use of machine learning. The proposed approach is based on extracting the relationships that exist between the residual frame and the number of requests over this feedback channel. We apply these concepts to pixel-domain Wyner-Ziv coding demonstrating significant savings in bitrates with a little loss of quality . Index Terms— DVC, Wyner-Ziv coding, Machine Learning, Feedback Channel, Turbo Codes. 1. INTRODUCTION DVC is a technique used to reduce the asymmetry in video codecs; the processing complexity of the encoders is reduced, leading to a low-cost implementation, while the majority of the computations are taken over by the decoders. The applications that are expected to benefit from this video coding architecture include wireless video surveillance, multimedia sensor networks, disposable video cameras, medical applications and mobile cameras phones. The theoretical framework of DVC was developed by Slepian- Wolf (SW) [1] for lossless Distributed Source Coding (DSC) and by Wyner-Ziv (WZ) [2] for the lossy case. One of the pioneering DVC approaches is the turbo based WZ coding scheme presented in [3], where the encoder is responsible for exploring the source statistics, and therefore achieving compression following the WZ paradigm. Most of the architectures available in the This work has been jointly supported by the Spanish MEC and European Commission FEDER funds under grants “Consolider Ingenio-2010 CSD2006-00046”' and “TIN2006-15516-C04-02”', and by JCCM funds under grant “PAI06-0106”. literature are based on [3] and make use of a reverse channel to the encoder requesting more information for the decoding process; this approach is referred to as feedback based architecture in the literature. Moreover, all architectures which are based on [3] use an ideal error estimation at the decoder which is able to measure the distortion between the original frame and the partial reconstructed one. The original frames are used to determine when a sufficient number of parity bits are received, in order to guarantee a residual BER (bit-error rate) below a given threshold, typically set to be equal to 10 −3 . This feature is also well- known in the literature and is referred to as rate control at the decoder. Such feedback based architectures offers significant problems in practical scenarios where the encoded video streams need to be stored for offline processing, and also when a bidirectional communication channel is not available. Feedback mechanics also have implications in terms of delay and decoder complexity since several iterative decoding operations may be needed to decode the data to a target quality level. Moreover, the ideal capacity inherent in these architectures is impractical for real implementations due to the dependence of the decoder on the original frame which is only available at the encoder. Basically, the two main drawbacks of current DVC architectures are the dependence on a feedback channel and the ideal correction detection based on original frames. Therefore, a solution without these disadvantages is essential to develop practical applications using DVC. In our previous work [4] we demonstrate that Machine Learning (ML) based approaches work well and have the potential to enable feedback free DVC architectures. The previous paper reported a framework that uses decision trees only to determine when a feedback to the encoder is necessary, i.e. a half-feedback approach. In this paper we present a full feedback free DVC architecture focus on eliminating the two main drawbacks mentioned above using a block based turbo encoder in Pixel Domain which exploits the correlation between the residual frame and the number of requests (bits parity) for each block. Also, in this work we use lossy coding for key frames instead of a lossless case used in the previous work which means a more realistic scenario. It is important to stress that 1140 978-1-4244-1764-3/08/$25.00 ©2008 IEEE ICIP 2008