Key points are not available for this paper at this time.
Enhancing video quality is critical for achieving a boosted user experience on smart devices including mobiles, televisions, and monitors. Practical hardware designs should deliver high performance with minimal resources under the stringent limitations related to bandwidth, area and energy budget. The widespread usage of deep-learning algorithms in image processing tasks, including super-resolution (SR) and noise-reduction (NR), has further emphasized the necessity for energy-efficient hardware solutions. Therefore, an emerging critical requirement is to deploy these algorithms in real-time and high-resolution scenarios. However, achieving this goal presents several challenges, as illustrated in Fig. 20.1.1: 1) High-resolution network inference considerably increases power consumption due to its computational complexity, low sparsity and high-precision requirements; 2) frequent high-precision data transactions to external memory result in substantial power usage related to bandwidth usage; 3) efficient and flexible mechanisms are essential to support the diverse network structures and operations.
Shih et al. (Sun,) studied this question.