Abstract
Human's facial expressions and emotions have direct impact on their action and decision-making abilities. Basic CNN models are complexity of speeding up the operation to minimize the complexity. In this paper, we have proposed a Deep Convolutional Neural Networks along with Bi-Long Short Term Memory, which is followed by a single and cross-fusion attention mechanism for gathering both spatial and channel information from feature vector maps. Piecewise Cubic Polynomial and linear activation function was used to speed up Interactive Learning Information (ILI). Global Average Pooling (GAP) computes weights for feature vector maps; softmax classifier is used to classify images into 7 classes based on the expression present on the input images. The proposed model's performance was compared with benchmarking methods like NGO-BiLSTM, ICNN-BiLSTM and HCNN-LSTM. The proposed model resulted with better accuracy than other methods with 82.89%, 96.78%, 95.78%, and 95.87% on FER 2013, CK+, RAF-DB and JAFFE datasets and also resulted in lower False Recognition Rate (FAR) of 7.23%, 1.42%, 1.96% and 1.78% on all four datasets respectively. The proposed model has performed well than other benchmarking models with high Genuine Recognition Rate (GAR) of 88.57% on FER2013, 97.23% on CK+, 96.87% on RAF-DB and 96.32% on JAFFE datasets respectively.