Motion-refined rewriting of H.264/AVC-coded video to SVC streams

• Published in: Journal of visual communication and image representation 2011-07, Vol.22 (5), p.391-400
• Main Authors: De Cock, Jan, Notebaert, Stijn, Lambert, Peter, Van de Walle, Rik
• Format: Article
• Language: English
• Subjects: Architecture; Complexity; Computation; Distortion; H.264/AVC; Inter-layer prediction; Mathematical models; Motion refinement; Multi-layer control; Optimization; Quality scalability; Rewriting; Streams; SVC; Transcoding; Transforms
• ISSN: 1047-3203; 1095-9076
• DOI: 10.1016/j.jvcir.2011.03.001

► Motion-refined rewriting of single-layer H.264/AVC streams to SVC streams is discussed. ► We present a fast rate-distortion technique based on Lagrangian relaxation. ► The presented R&D model takes into account both base and enhancement layer rate and distortion. ► Motion-refined rewriting in the transform domain can increase rate-distortion performance. ► The presented architectures significantly reduce the computational complexity over reencoding. In this paper, we discuss motion-refined rewriting of single-layer H.264/AVC streams to SVC streams with multiple quality layers. First, we elaborate on techniques we developed for efficient rewriting of residual data from H.264/AVC to SVC. We investigate if rate-distortion performance can further be improved by extending these architectures with motion refinement techniques, which exploit the inter-layer motion prediction mechanisms available in SVC. For optimum performance, we discuss a fast rate-distortion technique based on Lagrangian relaxation. Although motion refinement in the transform-domain leads to extra distortion in the bitstream, we show that our rate-distortion model successfully takes into account both base and enhancement layer rate and distortion during optimization. Implementation results show that motion-refined rewriting in the transform domain can increase rate-distortion performance, with gains of up to 0.5 dB for the SVC base layer. The presented rewriting architectures significantly reduce the computational complexity when compared to reencoding, with a speed-up by a factor of forty or more, even in the case of motion refinement.