A lightweight multi-scale feature integration network for real-time single image super-resolution

Zheng He, Kai Liu, Zitao Liu, Qingyu Dou, Xiaomin Yang

Abstract: Recently, numerous methods based on convolutional neural networks (CNNs) have been proposed to attain satisfactory performance in single image super-resolution (SISR). Meanwhile, diverse lightweight CNN-based networks have been proposed to achieve applicability in real-time scenarios. However, the receptive fields in lightweight networks are limited because they do not make good use of multi-scale information. In this paper, we propose a lightweight multi-scale feature integration network (MFIN) to address the above issue. Specifically, to expand the receptive fields for global features, MFIN is constructed by cascading the multi-scale feature integration blocks (MFIBs) in a serial manner. Each MFIB contains a multi-scale feature extraction module (MFEM) and a feature integration unit (FIU). To enlarge the receptive fields at a granular level, the features in MFEM are cascaded in a parallel manner. To capture the full-image dependencies, FIU incorporates the dense and pixel-wise correlations from the outputs of MFEM efficiently. The conducted experiments demonstrate that our method outperforms state-of-the-art methods in quantitative and qualitative evaluation. Notably, the experimental results on running time state that our method can achieve real-time performance.