DIRECTIONAL MULTIPLE DESCRIPTION SCHEME FOR STILL IMAGES Gabriella OLMO * CERCOM Politecnico di TORINO Department of Electronics Corso Duca degli Abruzzi 24 10129 Torino (Italy) Tammam TILLO CERCOM Politecnico di TORINO Department of Electronics Corso Duca degli Abruzzi 24 10129 Torino (Italy) ABSTRACT Multiple description coding of images can be imple- mented by means of simple splitting of the pixels into sub- sets. This can be realized via a pre- and post-processing to most standards for image communications, with obvi- ous advantages. Unfortunately, this method suffers from high side distortion when only one description is received. In order to overcome this limitation, and to achieve bet- ter control of the central versus side distortion trade off, a method has been proposed which expands the image by zero padding in the two-dimensional DCT domain [4, 5]. However, the computational complexity of the pre- and post processing stages turns out to be considerable. In this paper, we propose a novel method based on one dimen- sional (either horizontal or vertical) image expansion; this technique allows for a better exploitation of the spatial data structures, and achieves a better estimation of lost descrip- tions. Simulation results show a noticeable performance improvement of the proposed method with respect to other zero padding techniques, in terms of both rate/distortion and computational complexity. 1. INTRODUCTION Multiple description coding (MDC) [1] is recognized as an effective method to protect multimedia information trans- mitted over networks subject to erasures. In the MDC ap- proach, two or more correlated descriptions of the same data are generated, which can be independently decoded, and yield mutually refinable information. Therefore, the quality of the recovered signal is dependent only on the number of received descriptions, and not on the specific loss pattern. Many methods have been proposed for the generation of multiple descriptions, among which MD scalar quantization [2], use of correlating transforms [3] and so on. However, these methods are not compatible with stan- dard image/video co-decoding tools; therefore their use * This work has been supported by CERCOM - Center for Multimedia Radio Communications can not be suitable for many multimedia applications. On the other hand, techniques which enable the creation of multiple descriptions of images by means of pre- and post- processing to standard co-decoders, such as JPEG2000 or JPEG, are very attractive; in fact, they allow to exploit a possible transmission diversity to achieve robustness, with- out substantial modification of wide spread standards. A straightforward method to achieve this goal it to split the pixels of the image to be encoded into two subsets, or sub- images, which are then passed on to the standard encoder. At the receiver side, in case both descriptions are correctly received, the sub-images are combined to achieve the full quality image. On the other hand, if one description is cor- rupted or lost, it can be estimated from the received one, exploiting the residual correlation between the two. Clearly enough, the quality of the decoded data is heav- ily dependent on the correlation between the descriptions. Over sampling of the original image prior to the descrip- tion generation can help improving the correlation char- acteristics; this has been proposed in [4, 5], where zero- padding in the two dimensional DCT domain is addressed, and also generalized to video sequences in [6]. A draw- back of this technique is the considerable computational complexity of the pre- and post-processing stages in charge of the description generation. Moreover, the method is not able to efficiently exploit the spatial structures of the im- age, in that images with more vertical or horizontal details are treated in the same manner. This means that the di- rectional correlation of the image is not exploited in an optimal way, and that better results could be achieved with the same amount of extra redundancy. In this paper, we propose the use of one directional image expansion. A proper zero-padding algorithm is em- ployed to expand either the rows or the columns of the in- put image, which is then splitted into the two sub-images representing the descriptions. A simple rule for the direc- tion selection is applied, which does not heavily impact on the computational complexity. This latter turns out to be low due to the simple one dimensional processing.