Frequency-based reconstruction of multi-description coded JPEG2000 images

Abstract

Multiple description coding (MDC), together with reconstruction, is a popular technique for concealing packets losses in a best-effort network like the Internet. In our previous study, we have developed a sample-based MDC scheme for subband images and an optimal reconstruction-based subband transform (ORB-ST) at senders that takes into account the reconstruction process at receivers. Due to the partitioning of a description into segments before coding the segments by the ORB-ST coder, our previous scheme has severe degradation in quality because, among other reasons, the coder cannot exploit redundancies across multiple segments in a description, especially at low bit rates. To address this issue, we propose in this paper to divide a description into subbands in the frequency domain, instead of segmenting a description in the sample domain. Based on the observation that the corresponding subbands across two descriptions after inverse transformation are correlated, we present a scheme that reconstructs a lost subband using the corresponding subband received in the other description. We show that our scheme is more effective than reconstruction in the full band; that it has similar quality and lower delays in low-loss scenarios and slightly degraded quality in high-loss scenarios, when compared to MDC, unsegmented, and JPEG2000-coded images sent by TCP; and that it has much better quality and similar delays when compared to MDC, segmented, and ORB-ST-coded images sent by UDP.