Convolutional Features-Based Broad Learning With LSTM for Multidimensional Facial Emotion Recognition in Human–Robot Interaction

Convolutional feature-based broad learning with long short-term memory (CBLSTM) is proposed to recognize multidimensional facial emotions in human–robot interaction. The CBLSTM model consists of convolution and pooling layers, broad learning (BL), and long- and short-term memory network. It aims to obtain the depth, width, and time scale information of facial emotion through three parts of the model, so as to realize multidimensional facial emotion recognition. CBLSTM adopts the structure of BL after processing was done at the convolution and pooling layer to replace the original random mapping method and extract features with more representation ability, which significantly reduces the computational time of the facial emotion recognition network. Moreover, we adopted incremental learning, which can quickly reconstruct the model without a complete retraining process. Experiments on three databases are developed, including CK+, MMI, and SFEW2.0 databases. The experimental results show that the proposed CBLSTM model using multidimensional information produces higher recognition accuracy than that without time scale information. It is 1.30% higher on the CK+ database and 1.06% higher on the MMI database. The computation time is 9.065 s, which is significantly shorter than the time reported for the convolutional neural network (CNN). In addition, the proposed method obtains improvement compared to the state-of-the-art methods. It improves the recognition rate by 3.97%, 1.77%, and 0.17% compared to that of CNN-SIPS, HOG-TOP, and CMACNN in the CK+ database, 5.17%, 5.14%, and 3.56% compared to TLMOS, ALAW, and DAUGN in the MMI database, and 7.08% and 2.98% compared to CNNVA and QCNN in the SFEW2.0 database.