Efficient Facial Emotion Recognition Using Residual Squeeze-and-Excitation and Enhanced Inception Networks

This article introduces a new deep convolutional neural network model to read facial emotions by fusing the residual squeeze-and-excitation blocks with enhanced inception modules to realize a better performance at a lower computational cost. The proposed DeepCNN network involves depthwise separable convolutions in inception blocks and residual links boosted with channel attention systems by use of squeeze-and- excitation blocks. Experimental analysis on the FER2013 data shows that the model attains 93.31% accuracy using only 4.1 million parameters, which is much more parameter efficient than the existing state of the art approaches. Its architecture uses data augmentation algorithms such as random rotation, horizontal flipping, and affine transformations to enhance generalization. Extensive testing demonstrates that the algorithm is highly consistent with high precision and recall in all seven classes of emotions and is computationally efficient enough to implement in real-time systems.