A New Emotion Recognition Method based on Low-Rank Representation and Nonnegative Matrix Factorization

Successful interpersonal communication is inseparable from the ability of accurate emotions recognition. Since emotions are normally expressed by text, audio and video. If you want to get much more accurate emotion recognition results, you must make best use of the three forms of emotions information. In practical applications, there often exist two problems: noise pollution and data redundancy. To this end, we propose a new method called LNER, which is based on low-rank representation(LRR) and non-negative matrix factorization(NMF) of emotion recognition methods. Specifically, we firstly use low-rank representation method to remove noise data and then obtain the data personality characteristics; Then, the clean data is decomposed by non-negative matrix factorization to obtain the shared latent features among multiple modal data, that is, the common feature information; Finally, emotion recognition is carried out for individual feature information and common feature information. Extensive experiments on different emotions datasets confirm the effectiveness of the proposed method.