Abstract: An integrated method was proposed to achieve the objective accuracy of emotion recognition from the electroencephalograph (EEG) signal. The EEG signals from DEAP data set were decomposed into multiple intrinsic mode functions (IMFs) with the empirical mode decomposition (EMD) method. After that, the power spectrum density was extracted as the EEG feature from the IMFs with different time windows. The emotion estimated scales of the subjects were mapped in the Valence-Arousal emotion model to be clustered into 4 classes. Gaussian kernel function support vector machine (SVM) was adopted to classify the emotion states. "One versus one" model was employed within the SVM to deal with the multi-classification problem. Results show that the highest accuracy of the emotion classification obtained by the Gaussian kernel function SVM acquires is 90.9% (with subject 22), and the mean accuracy is 68.31%. The results explain that the method can recognize different emotion from EEG signal effectively, and that different participants have different emotion experiences with the same stimulus, and that the EEG features from high frequency IMFs achieve higher emotion classification accuracy than that from low frequency IMFs.