Modeling of 3-D DCT coefficients for fast video encoding

Abstract

-The discrete cosine transform (DCT) plays a key role in image/video coding due to its powerful energy concentration ability. On the other hand, since many DCT coefficients will be quantized to zeros, a huge number of redundant computations are introduced. This paper proposes an analytical model to predict the zero-quantized DCT (ZQDCT) coefficients for fast 3-D DCT based video encoding. The theoretical analysis is first performed to study the sufficient condition under which each DCT coefficient becomes zero. Then, the threshold for each ZQDCT coefficient is mathematically determined. Finally the transform and quantization operations of the predicted ZQDCT coefficients are omitted. We show, by the experimental results, that the proposed model can significantly reduce the redundant computations and achieve better real-time performance than the original encoder without any quality degradation. Computational reduction also implies longer battery lifetime and energy economy for digital applications.Keywords-discrete cosine transform (DCT); quantization; video coding; computational complexity I .I NTRODUCTIONAs an alternative approach to motion compensation in video coding standards (e.g., MPEG-1/2/4 and H.261/3), three-dimensional (3-D) DCT has been investigated by many researchers. The main advantage of 3-D DCT is the low computational complexity. This is very important for many real-time applications and digital devices as they are still suffering most from the restriction of battery lifetime. Although there is significant research in reducing the computations for H.264 and [1]-[5], the complexity is still far from the requirements of most portable devices. Therefore, there is a great need now to develop sufficient 3-D DCT algorithms for such applications.Previously, the efforts to advance the 3-D DCT algorithms are mainly focused on the improvement of coding efficiency. These proposed approaches can reduce the correlations more effectively by using the scene change detector [6], threshold for different local activity [7]- [8] and multiple description method [9]. The latest results [10] indicate that 3-D DCT coding outperforms MPEG-2 while is slightly worse than MPEG-4 at high bit rates.Although 3-D DCT coding is much superior to H.264 and MPEG-4 in terms of computational complexity, it is still desired to further reduce the computations without visual quality degradation. As the most complex part, threedimensional DCT transform consumes more than half of the computation in 3-D DCT encoding while a large number of transformed coefficients will be finally quantized to zeros. Thus, if we can reduce these redundant computations, the encoding process will be much faster. This also implies longer battery lifetime for portable digital applications.In this paper, we propose an analytical model to skip redundant DCT and quantization without visual quality degradation through a theoretical analysis on the dynamic range of DCT coefficients. Although the proposed model is implemented based on the baseline 3-D DCT ...