Many of the current emotion recognitions are intra-subject, the reason is that recognize cross-subject emotions have always been difficult. In this paper, we proposed a decision tree classifier based on sequential backward selection (DT-SBS), and "leave-one-subject-out" verification strategy was used for evaluating performance of cross-subject emotion recognition. Power spectral density (PSD) was extracted from EEG signals as feature set. The emotion recognition performance of the method we proposed was examined on EEG data of a group of subjects. The pictures were utilized to stimulate the emotion of 12 subjects and we collected their EEG signals (PEEG), the mean recognition accuracy of 65.8%. The dataset for emotion analysis using physiological signals (DEAP) was utilized to verify the method we proposed in this paper and the mean recognition accuracy of 65.2%. Compared with the performance reported in the existing literature, the method we proposed in this paper was effective in cross-subject emotion recognition. In addition, the emotion recognition performance of different rhythms were explored, we found that the mean recognition accuracy of PSD in the gamma rhythm achieved the best mean recognition accuracy compared with the other rhythms; the high frequency bands contained higher emotional information than the low frequency bands.