|M.Sc Student||Hait Naama|
|Subject||Model-Based Transrating of Coded Video|
|Department||Department of Electrical Engineering||Supervisor||Professor Emeritus David Malah|
|Full Thesis text|
Transrating of coded video is the process of reducing the bit rate of a high-quality pre-encoded video to match user-specific bit rate requirement. In this research work, we examine model-based transrating via requantization of the transform coefficients, in the state of the art H.264 coder.
Previous works, related to previous standards, chose the optimal requantization steps via an iterated Lagrangian optimization that minimizes the distortion subject to a rate constraint. However, these works evaluated the rate in each iteration and therefore required an exhaustive search at a high computational load. Moreover, these methods cannot be applied for requantization in H.264 as is, due to its advanced coding features such as intra spatial prediction and context adaptive entropy coding.
The goal of a transrating system is to reduce the bit rate of an encoded video sequence, at low complexity, while preserving a high quality video. To reduce the computational complexity, the proposed transrating system reuses as much input coding decisions as possible (e.g. motion vectors). The model-based requantization further reduces the computational burden by alleviating the repetitive quantization and coding required during the search for the optimal step sizes. The models incorporated in this work relate the rate and the distortion to the fraction of zeroed quantized transform coefficients, ρ, rather than to the step size itself.
The intra spatial prediction in H.264 introduces block dependencies that pose two algorithmic problems. First, to avoid a drift error, the intra-coded frame should be fully decoded. Second, estimating the relation between ρ and the step size becomes a challenging task, for which we propose a novel statistical-based model.
For optimal requantization in inter-coded frames, we propose two novel modifications of previous work. First, an extended Lagrangian optimization is proposed, to improve the subjective quality by regulating the changes in the quantization step sizes throughout the frame. Second, the ρ-domain rate-distortion models suggested in previous works are not suitable for macroblock level coding in H.264. The macroblock level rate-distortion models developed in this work are adapted to H.264 requantization and consider its context adaptive entropy coding.
Overall, as compared to re-encoding, the proposed transrating system reduces the computational complexity by a factor of about 4, at an average PSNR loss of 0.75[dB]. In comparison with a simple one-pass requantization, the proposed algorithm achieves better performance both objectively (PSNR gain of up to 1.6[dB]) and subjectively, at the cost of twice the complexity.