EESRNet: A Network for Energy Efficient Super-Resolution

Shijie Yue, Chenghua Li, Zhengyang Zhuge, Ruixia Song

Abstract: Recently, various high-performance video super-resolution methods have been proposed. However, deployment on mobile phones is cumbersome due to the limitations of mobile phones’ power consumption and computing power. We find methods that exploit temporal information in videos (e.g. optical flow) require huge energy consumption. Therefore, we use hidden features to preserve temporal information. Besides, the energy-efficient super-resolution network (EESRNet) is obtained by removing the residual connections in the Anchor-Based Plain Network (ABPN) [8]. Combining the two, we propose a Temporal Energy Efficient Super-Resolution Network (TEESRNet), which can efficiently utilize video spatio-temporal information with low energy consumption. Experiments show that for EESRNet, compared with ABPN, the latency is reduced by more than $$40\%$$ , while performance decreases slightly. Furthermore, for TEESRNet, the PSNR is improved by 0.24 dB and 1.19 dB compared to EESRNet and RRN [19] respectively, while still maintaining real-time (<30 ms).