Lightweight image super-resolution via overlapping back-projection feedback network for embedded devices

Super-resolution (SR) technology is widely used in embedded devices because it can improve image quality. However, to achieve improved performance, SR networks usually take a massive memory because of their large number of parameters_ ihey are not applicable for embedded devices with low power consumption. In this work, we propose an overlapping back-projection feedback network (LOBFN) for image SR, which is a lightweight network designed for embedded devices. First, a back-projection feedback block (FSS;) and recursive concatenation are used to leam the hierarchical representations of the network. Second, an overlapping back-projection suitable for lightweight network is proposed to minimize the reconstruction errors. Finally, a fusion attention module (FAM) is proposed to perceive infonmation-rich features. The final experiments proved that the proposed LOBFN significantly improved the SR performance of lightweight networks.