|M.Sc Student||Gendler Ora|
|Subject||On Transcoding Optimization Using Requantization|
|Department||Department of Electrical and Computer Engineering||Supervisor||ASSOCIATE PROF. Moshe Porat|
Multimedia applications involve transmission and storage of coded images and video. Often, the data to be transmitted is initially coded and stored in high quality. Later on, when transmission is required, in many cases the bit-rate of the data has to be reduced to meet the limitations of the transmission media and the available resources of the end users. This raises the need for bit-rate reduction, known as transcoding or transrating. For real-time applications, it is crucial for the transcoding to be of low computational complexity, while ensuring low distortion. A straightforward approach to transcoding is to use requantization with a coarser quantization step. In earlier works it was empirically demonstrated that the performance of requantization depends mainly on the ratio between the quantization step used in the initial compression and the step used for requantization.
In this work, requantization for transrating of video and still images is analyzed. We establish the dependency of the rate and the distortion of requantized images and video streams on the ratio between the quantization steps. Our analysis is based on the structure of the Uniform Threshold Quantizer and the Laplace-like distribution of the DCT coefficients in sub-band coding. We also show that the rounding policy has a major effect on the requantization performance. We analyze this phenomenon and provide guidelines for the preferred rounding policy. In addition, we examine a generalized transcoding approach that incorporates image downscaling with requantization. This method is based on the trade off between the number of quantization levels and the number of pixels used to code an image. We propose a method to select the coding mode, when the possibilities are requantization or downscaling. The proposed combined coding model outperforms requantization.
We also show how to incorporate the proposed requantization approach in a rate control scheme, using the rho-model as an example. We show that the rho-model can be used to accurately predict the resulting bit-rate for various requantization steps. Consequently, we analyze MPEG video sequences and JPEG still images, which could be adapted also to other compression standards, such as baseline H.263. For more advanced standards, such as H.264, further analysis is required. Our conclusion is that the proposed approach could be instrumental in achieving a required bit-rate when recompressing still images and video at low distortion, while allowing real-time implementation.