Efficient Hard-Decision Quantization Using an Adaptive Deadzone Offset Model for Video Coding

Abstract

In video coding, rate distortion optimized quantization (RDOQ), a popular version of softdecision quantization (SDQ), achieves superior coding performance, however is ill-suited for hardware implementation due to its inherent sequential processing. On the other hand, deadzone hard-decision quantization (HDQ) is friendly to hardware implementation, however suffers from non-negligible coding performance degradation. This paper proposes a content-adaptive deadzone offset model to improve the coefficient-wise deadzone HDQ by imitating the behavior patterns of RDOQ. The contributions of this paper are characterized by twofold. On one hand, this work formulates seeking optimal deadzone offset model as a problem of binary classification, and analyzes the distribution characteristics of the optimal deadzone offsets obtained from samples by fully imitating RDOQ, and then derives adaptive deadzone offset model by maximizing the right classification probability of offset-induced rounding in HDQ. On the other hand, the distribution parameters of DCT coefficients are measured in a position-wise way, and the adaptive deadzone model is built by applying Maximum a posterior estimation method according to three characteristic parameters, i.e. quantization step size, parameter of DCT coefficients, and quantization remainder, in the sense of rate distortion optimization. Simulation results verify that the proposed adaptive HDQ algorithm, in comparison with fixed-offset HDQ, achieves 0.54% and 0.52% bit rate saving on average with almost negligible complexity increment. Simultaneously, the proposed algorithm only sacrifices smaller than 0.55% and 0.54% increment in terms of BD-BR in comparison with RDOQ. The proposed HDQ is well-suited for hardwired video coding implementation.

Context-Adaptive Inverse Quantization for Inter-Frame Coding

“…Yan et al [19] utilized spatial and temporal characteristics to establish a spatiotemporal perception-aware model to adjust the CTUlevel QP offset. In [10], [20]- [22], the information loss and compression artifacts introduced by quantization can be effectively reduced by accurately modeling the distribution of DCT coefficients. In [8], [9], [23], the temporal adaptive quantization methods were used to reduce the inter-frame dependency and achieve a significant global optimization by using a group of neighboring frames.…”

“…Under this strategy, blocks with uniform size still need to share the same inverse transform kernels. To avoid the cost of transmitting parameters, some studies [8]- [10] have proposed adaptive quantization, which indirectly achieves the purpose of reducing quantization distortion by dynamically adjusting the quantization level at the encoder, but the decoder still follows a unified inverse quantization rule without having the ability of adaptive compensation. The in-loop filters [2] further reduce the quantization distortion by using the pixel correlations.…”