The invention discloses a magnetic resonance imaging method and device. The method combines a deep network model with a conventional acceleration reconstruction method in sequence. Firstly, the deep network model is used for recovering first imaging information with a high downsampling multiple to second imaging information with a low downsampling multiple, then, the conventional acceleration reconstruction method is used for completely reconstructing the second imaging information with the low downsampling multiple, and therefore, a final magnetic resonance image is obtained. Therefore, in the magnetic resonance imaging method provided by the invention, the output training sample of the deep neural network used for magnetic resonance quick imaging is not full sampling or super-full sampling data but is downsampling data, so that the output training sample of the training deep neural network can be obtained through a conventional downsampling method, the output training sample can be obtained through short collection time, and therefore, the deep neural network can be applied to a magnetic resonance imaging application scene with limited collection time, such as abdomen scanning.