Downsampling-based multiple description image coding using optimal filtering Yüksel Yapıcı Begüm Demir Sarp Ertürk Og ˘ uzhan Urhan University of Kocaeli Laboratory of Image and Signal Processing Department of Electronics & Telecommunications Engineering Veziroglu Campus 41040, Kocaeli, Turkey E-mail: urhano@kou.edu.tr Abstract. In this paper, a multiple description image coding scheme is proposed to facilitate the transmission of images over media with possible packet loss. The proposed method is based on finding the optimal reconstruction filter coefficients that will be used to reconstruct lost descriptions. For this purpose initially, the original image is downsampled and each subimage is coded using standard JPEG. These decoded images are then mapped to the original im- age size using the optimal filters. Multiple descriptions consist of coded down-sampled images and the corresponding optimal recon- struction filter coefficients. It is shown that the proposed method provided better results compared to standard interpolation filters (i.e., bicubic and bilinear). © 2008 SPIE and IS&T. DOI: 10.1117/1.2976420 1 Introduction The multiple description coding MDC approach is used for data transfer over packet networks, which are nowadays widespread. This approach fundamentally provides efficient transmission of multimedia data through these kinds of error-prone networks. MDC carries out this operation by coding and transmitting the original data using more than one bit stream and therefore reduces the influence of pos- sible packet loss. Various multimedia applications require data transmis- sion over error-prone networks in which part of the data might not arrive at the receiver. Automatic repeat requests executed by the receiver are not possible for real-time data, such as voice and video, because these will cause long delays. The MDC approach enables reconstruction of data at an acceptable quality level in the case of possible packet losses. Original data are coded at the encoder into more than one packet i.e., multiple descriptions such that each one is self-decodable. When all descriptions reach the re- ceiver, data are reconstructed at high quality; otherwise, acceptable quality data are still obtained. Nevertheless, coding efficiency is degraded due to redundancy introduced into descriptions. MDC schemes can be grouped according to their com- putational complexity and redundancy insertion ap- proaches. One of the first MDC methods makes use of mul- tiple description scalar quantization MDSQ, 1,2 which uses overlapping quantization steps to enable redundancy. At the decoder the intersection of received quantization steps are used for inverse quantization. In Refs. 3 and 4 redundancy insertion is carried out using transforms referred to as pair- wise correlating transform PCT–based approaches. These methods transform two input variables into two output vari- ables and then encode the transformed variables. If one of the transformed variables is not received, then it can be estimated at a certain accuracy using the other variable. Polyphase downsampling PD–based MDC approaches have been proposed in Refs. 5–8 The first type of PD-based MDC approaches quantize input data at two different quan- tization levels after downsampling. 5,6 The second type of PD-based MDC approaches perform oversampling, by making use of zero padding in the discrete cosine transform DCT domain in one or two dimensions, as presented in Refs. 7 and 8, respectively, before the descriptions are gen- erated. In PD-based MDC approaches, if one of the de- scriptions is lost at the receiver, then the other samples are used to reconstruct the lost data. MDC redundancy is intro- duced using frame expansion in. 9,10 Recently, wavelet- based MDC approaches have become popular. 11–15 For ex- ample, the MDSQ approach is applied in the wavelet domain, and it is shown that MDC performance is in- creased compared to image domain coding. The method presented in Ref. 12 uses biorthogonal filter structures to construct wavelet descriptions having lower redundancy. It uses a fast converging iterative convex optimization ap- proach to improve the quality at the receiver. A PD-based MDC approach is used in the wavelet domain in Ref. 13, and it is shown that the performance of this method is bet- ter than the MDSQ approach. MDC is directly used with a JPEG2000 coder in Ref. 14, where rate distortion charac- teristic of the input data is examined to introduce an adjust- able level of redundancy. PCT is combined with wavelet transform-based image coding in Ref. 15, and it is shown that this outperforms MDC in the DCT domain. Paper 07125RR received Jun. 27, 2007; revised manuscript received Apr. 30, 2008; accepted for publication May 2, 2008; published online Aug. 28, 2008. 1017-9909/2008/173/033018/9/$25.00 © 2008 SPIE and IS&T. Journal of Electronic Imaging 17(3), 033018 (Jul–Sep 2008) Journal of Electronic Imaging Jul–Sep 2008/Vol. 17(3) 033018-1