Aiming at the problems of the existing single image super-resolution (SISR) reconstruction methods, such as insufficient feature utilization and weak ability to recover image high-frequency details, we propose a multi-scale feature feedback network (MFN) for SISR. First, multi-scale convolutional kernels are used to extract features of different dimensions in the image. Thus, the ability of the network to extract features of different scales early is improved. Then, we combine the dense network to design a group of multi-scale projection units to explore the correlation between high and low-resolution image features from multiple scales. Each group of projection units consists of a multi-scale up-sampling block and a multi-scale attention block (MAB). The MAB focuses on high-frequency features related to reconstructed image edges and textures. Finally, we used a constrained recurrent neural network to implement the feedback mechanism to improve the feature extraction capability of the network. Extensive experimental results demonstrate the effectiveness of our approach, which has advantages over other state-of-the-art SISR methods in terms of restoration quality and network complexity.