In order to improve the accuracy of EEG emotion recognition, extract richer feature information and improve the stability of network model, an improved EEG emotion recognition model based on multi-level attention mechanism is proposed. In the aspect of feature extraction, the original EEG signal was transformed into four-dimensional space spectrum time structure to extract rich EEG information. In the aspect of network model, a two-way convolution neural network was constructed to learn spatial and frequency information. It can effectively extract multi-scale features and increase the network width to learn richer feature information. After the convolution layer and pool layer, the batch normalization layer was integrated to prevent over fitting. Finally, a multi-level attention mechanism-bidirectional gated recurrent unit module was constructed to process the time characteristics and cooperate with Softmax classification. The bidirectional gated recurrent unit was used to learn more comprehensive upper and lower level feature information. The multi-level attention mechanism was used to correlate different time slices with the overall time slices in four-dimensional features. The evaluation experiments were carried out in two dimensions of arousal and potency of DEAP data set. The average accuracy of two classifications were 96.38% and 96.73% respectively, and the average accuracy of four classifications was 93.78%. The experimental results show that the average accuracy of this algorithm is improved compared with single channel convolutional neural network and other literature algorithms, which shows that this algorithm can effectively improve the performance of EEG emotion recognition.