Partial Attention for Data-Driven Image Super Resolution under Low Computing Cost.

Low computation cost is crucial for a seamless experience on mobile consumer devices with limited resources. This study presents an efficient attention module for deep learning-based image super-resolution under low computing cost. We propose partial enhanced spatial attention (PESA) to achieve efficient and high-performing attention modules, which draws inspiration from partial convolution for feature extraction. Utilizing an efficient super-resolution network, our approach is assessed on two super-resolution datasets and contrasted with other attention strategies. PESA obtains the lowest computing cost and model parameters, as well as the top quantitative results on both datasets.