<p>The utilization of high-resolution Magnetic Resonance Imaging (MRI) images contributes to the early detection of diseases and subsequent accurate treatment. As the resolution of MRI images increases, image scanning and reconstruction time rises, which causes patient discomfort during scanning. To overcome these weaknesses, we propose an MRI image super-resolution (SR) reconstruction network based on frequency fusion. Firstly, the network takes deep feature extraction of high-frequency (HF) features and low-frequency (LF) features of the image, then HF and LF features are fused to obtain high-resolution images with clearer image structure and texture details. HF information is used to guide and enhance the reconstruction of details. In this paper, this network uses local and global residual designs to stabilize model training while improving network information flow. Additionally, it uses depthwise separable convolutions instead of regular convolutions, which enables the network to reduce the number of parameters and computational complexity while reconstructing approximately the same image quality. As a result, we conduct simulation experiments on the IXI dataset by performing high-resolution reconstruction of PD, T1, and T2 images. The experimental results indicate that the average reconstruction time of the network proposed in this paper is 49.64 ms/image, 33.56 ms/image, and 23.28 ms/image at scales of 2, 3, and 4, respectively. Comparing the reconstruction results with RCAN, HAT, SwinIR, and other networks, it is found that the network proposed in this paper exhibits better reconstruction results and faster reconstruction speed.</p>