In recent years, for the effect of image reconstruction, the research of single image Super-resolution (SISR) using Convolutional Neural Networks (CNN) has made significant progress, and is superior to traditional methods. Aiming at improving network performance, residual connection plays a key role in these CNN-based methods. With the increase of the depth of these CNN-based networks, the residual features more aggregate the information of multiple different levels of the input image, which is more advantageous for the reconstruction of image details. However, inmost deep CNN-based SR models, the hierarchical features of the original Low-resolution (LR) image are not fully utilized. In order to solve this problem. We propose Residual Dense Feature Aggregation Block (RDFAB). Where RDFAB is the unit of residual blocks and the features extracted from each residual block are aggregated to the end of the module. As a result, RDFAB is able to produce more representative features. To further strengthen the effectiveness of RDFAB, we propose a multi-scale feature aggregation (MFA) framework, and our final multi-scale residual dense feature aggregation network (MRDFAN) combines RDFAB and MFa together with the addition of skip connections.Comprehensive experiments demonstrate the necessity of the MF A framework and the advantages of MRDFAN over the state of the art SISR methods.