There is the problem of motor imagery recognition in rehabilitation training. To address this issue, a feature extraction method of bilevel brain functional network based on the corticomuscular coupling node selection is proposed. According to the strength of corticomuscular coupling, the core nodes under each movement of the subjects are selected. Based on the core nodes and the prior knowledge of motor sensory brain region in neurophysiology, the motor sensory core node regional network is constructed and the features are extracted. By utilizing the whole network characteristic of minimum spanning tree, the diameter and average eccentricity of the minimum spanning tree are combined with the average node degree, average clustering coefficient and average path length of the core brain network. In this way, the comprehensive characteristics of global and regional functional brain network are constructed by the bilevel brain functional network. The support vector machine is selected as the classification method, and the average accuracy of two types motion imagery is 86. 96% , which confirms that the proposed bilevel brain functional network analysis method has excellent feature expression ability and can effectively extract the inherent neural-muscle correlation features. It provides a new idea for motor imagination recognition.