An Emotion and Gender Detection Using Hybridized Convolutional 2D and Batch Norm Residual Network Learning

Abstract

The deep learning algorithm called convolutional neural network (CNN) particularly with Residual Network (ResNet) receiving much attention from the research community in facial recognition recently. Unfortunately, the complexity of optimization problems in overfitting and vanishing gradient cause huge obstacles. More specifically, once the gradient is backpropagated in initial layers, repeated multiplication among layers constructs gradient infinitely small and causes the layers of the network to become deeper and degrade the performance. Moreover, the skip connection that comprises the residual network (ResNet) is not enough to solve the above-mentioned limitations, and this could downgrade the optimization of used layers and potentially further downgrade the accuracy. Therefore, a deep residual network (ResNet) with hybridized function i.e., convolutional-2D and Batch Norm is proposed as this could allow direct signal propagation from the initial to the final layer of the network for every single residual block deeply. Initially, the convolutional-2D and Batch Norm were constructed to overcome bias in-depth nets and propagate the gradients directly from the loss layers to any previous layers, while skipping intermediate weight layers deeply that have the potential to trigger vanishing or deterioration of the gradient signal. The proposed learning model has improved the degradation of accuracy drawback by decreasing the number of layers needed more in low level as compared to existing work for each block using batch normalization and convolutional-2D function.