Abstract: Recently, Convolutional Neural Network (CNN) has demonstrated powerful capabilities in Single Image Super Resolution (SISR) and has shown significant improvements over traditional methods. However, although these methods are very successful, they are not practical to be directly applied to some edge devices due to the large amount of computing resources required. To address this problem, a lightweight SISR network called Multi-path Fusion Enhancement Network (MFEN) is proposed. Specifically, a novel Fusion Attention Enhancement Block (FAEB) is proposed as the main building block of MFEN. The FAEB module consists of one main branch and two hierarchical branches. The main branch is composed of stacked Enhanced Pixel Attentional Blocks (EPABs), responsible for deep feature learning of the feature maps. The hierarchical branches are responsible for extracting and fusing feature maps with different sizes of receptive fields, achieving multi-scale feature learning. The fusion method for the hierarchical branches is based on the output of adjacent EPAB modules as branch inputs, and the Self-Adaptive Attention Module (SAAM) is designed to dynamically enhance the fusion degree of features with different sizes of receptive fields, achieving further feature information completion and thus achieving more comprehensive and accurate feature learning. Extensive experiments show that our MFEN achieves higher accuracy than other advanced methods on benchmark test sets.