Application of Bi-Laplacian Transparent Composite Model to HEVC: Residual Data Modelling and Rate Control
As an important tool in lossy multimedia compression, transform is employed to transfer signals from space/time domain into frequency domain. As such, transform can reduce the correlation among neighboring pixels and can concentrate energy on only a few coefficients. This is useful and beneficial for quantization and entropy coding.
Among various transforms, discrete cosine transform (DCT) is the most widely used one in multimedia compression technologies from the early image compression standard JPEG to video compression standard H.264 and High Efficiency Video Coding (HEVC) which is the newest video coding standard. During the development of lossy compression, a lot of interest has been attracted to understand the statistical distribution of DCT coefficients, which would be useful to design compression techniques, such as quantization, entropy coding and rate control.
Recently, a Bi-Laplacian Transparent Composite Model (BLTCM) has been developed to provide modeling of distribution of DCT coefficients with both simplicity and accuracy. It has been reported that for DCT coefficients obtained from original images, which is applied in JPEG, a TCM can provide better modeling than Laplacian.
In video compression, such as H.264/AVC, DCT is performed on residual images obtained after prediction with different transform sizes. What's more, in HEVC, besides DCT as the main transform tool, discrete sine transform (DST) and transform skip (TS) techniques are possibly performed on residual data in small blocks. As such, the distribution of transformed residual data differs from that of transformed original image data.
In this thesis, the distribution of coefficients, including those from all DCT, DST and TS blocks, is analyzed based on BLTCM. To be specific, firstly, the distribution of all the coefficients from the whole frame is examined. In HEVC, the entropy coding is implemented based on the new encoding concept, coefficient group (CG) with size 4_x4, where quantized coefficients are encoded with context models based on their scan indices in each CG. Secondly, coefficients at the same scan indices among different CGs are grouped together to form a set. Distribution of coefficients in each set is analyzed. Based on our result, BLTCM is better than other commonly used distributions, such as Laplacian and Cauchy distributions, in both _2 and KL-divergence testing.
Furthermore, unlike the way based on Laplacian and Cauchy distribution, the BLTCM can be used to model rate-quantization (R-Q) and distortion-quantization (D-Q) models without approximation expressions. R-Q and D-Q models based on BLTCM can reflect the distribution of coefficients, which are important in rate control. In video coding, rate control involves these two models to generate a suitable quantization parameter without multi-passes encoding in order to maintain the coding efficiency and to generate required rate to satisfy rate requirement. In this thesis, based on BLTCM, rate control in HEVC is revised with much increase in coding efficiency and decrease in rate fluctuation in terms of rate variance among frames for constant bit rate requirement.