A Neighborhood Dependent Approach (NDA) for Power-aware Design of a Real-time Video Enhancement System

摘要

In this paper, a design of a high performance, power-aware video enhancement system based on a nonlinear enhancement technique is proposed. The main contribution is a novel partitioning and gating technique to reduce switching activities in a 2-D convolution module. The proposed technique includes an efficient on-chip detection scheme for a parallel architecture of a 2-D convolution unit. Bit representations of the pixel values within a neighborhood are considered in the detection scheme. In addition, a data forwarding technique is developed to incorporate symmetry in the convolution kernel mask with an on-chip buffering method to reduce the number of multiplications by 75%. It is observed that the proposed enhancement system design reduces dynamic power consumption by more than 29% compared to the baseline approach.