A new algorithm for reducing the requantization loss in video transcoding

Video transcoders are devices that convert one video bitstream into another type of bitstream, either with or without standard format conversion. One step to be applied in video transcoders is the requantization of the transform coefficients, if an adaptation to a lower data rate is necessary. During this step, the quality is in most cases degraded compared to a single quantization. This is a consequence of nonoverlapping quantization characteristics of the input and the output quantizer. In this work we propose a new choice of the reconstruction level for the requantization step depending on the effective quantization curve of both quantization parameters involved. The reconstruction level is calculated such that it is centered in each effective quantization interval after requantization. Compared to the standard midpoint requantization this leads to quality gains of 3 dB PSNR for most pairs of input and output quantization parameters (QP). The algorithm is useful for intra- and inter-frame coding.

1. Introduction

It is often necessary to adapt encoded video bitstreams according to the transmission environment in order to support as many end devices as possible. This can be done by video transcoding technologies, for which a good overview can be found in [1]. A typical approach for bitrate adaptation is homogeneous transcoding in the domain of the Discrete Cosine Transform (DCT). Efficient implementations are investigated very well for the MPEG-2 standard, e.g. [2], [3]. Here, the bitstream is decoded only up to the reconstruction of the transform coefficients. Then requantization is applied on these coefficients. This avoids the time consuming step of conversion to the pixel domain using the inverse DCT. The requantization leads to a smaller number of levels to be encoded and therefore to a lower data rate. For drift-free transcoding of inter frames (P-frames), additionally the difference between the reconstructed input and output coefficients has to be calculated.