Compared with other medical imaging methods, magnetic resonance imaging (MRI) requires a longer scanning time. Each pulse sequence needs to undergo multiple steps of radiofrequency (RF) excitation, gradient encoding, data acquisition and image reconstruction to generate an MRI image, and the time required for one MRI image is at least several hundred milliseconds. Therefore, reducing the scanning time of MRI images and improving the spatial resolution of images are important prerequisites for the expansion of clinical applications of MRI technology. In this paper, based on the introduction of the principle of MRI, the principle of magnetic resonance image reconstruction and the evaluation index of magnetic resonance image quality, a fast reconstruction technique of MRI based on convolutional neural network is proposed. Taking the cranial MRI as the research object, the designed convolutional neural network, after offline training, reconstructs the image and the undersampled reconstructed image of the untrained model from the online test with a large amount of known a priori information, and compares it with the reconstructed image of the full-sampled image, so as to realize the fast reconstruction method of MRI based on convolutional neural network (CNN). The experiments show the effectiveness of CNN for improving the quality of magnetic resonance reconstructed images with undersampled data, and the undersampled reconstructed images accelerate the speed of magnetic resonance image reconstruction.