Multi learning emotion recognition based on context awareness and attention mechanism

To improve the efficiency and accuracy of facial image emotion recognition, a multi learning emotion recognition network structure was proposed based on context awareness and attention mechanism．The proposed network was composed of three sub networks: scene feature extraction, body feature extraction and fusion decision-making．Single and double output structures were adopted to realize multi-label emotion classification and continuous spatial emotion regression．Improved focus loss function was proposed to assign more weights to small samples or samples that were difficult to classify, to improve efficiency of network training．Simulations using emotic data set showed that proposed improved focus loss and mean absolute error regression combination loss was better, average classification accuracy and regression average error rate were 28.5% and 0.098 respectively．It is concluded that the proposed method had better classification effect for small samples．