Neural decoding is still a challenge and hot topic in neurocomputing science. Recently, many studies have shown that brain network patterns containing rich spatial and temporal structure information, which represents the activation information of brain under external stimuli. %Therefore, the research of decoding stimuli from brain network received extensive more attention. The traditional method extracts brain network features directly from the common machine learning method, then puts these features into the classifier, and realizes to decode external stimuli. However, this method cannot effectively extract the multi-dimensional structural information, which is hidden in the brain network. The tensor researchers show that the tensor decomposition model can fully mine unique spatio-temporal structure characteristics in multi-dimensional structure data. This research proposed a stimulus constrained tensor brain model(STN)which involves the tensor decomposition idea and stimulus category constraint information. The model was verified on the real neuroimaging data sets (MEG and fMRI). The experimental results show that the STN model achieves more than 11.06% and 18.46% on accuracy matrix compared with others methods on two modal data sets. These results imply the superiority of extracting discriminative characteristics about STN model, especially for decoding object stimuli with semantic information.