Comparative Study of Deep Learning Techniques to Detect Emotion from Baby Faces

Emotion identification on baby faces has great potential for use in pediatric specialists and peri-partum technology. This study offers a detailed comparison of deep learning approaches to identify three emotional states in baby faces: Crying, Happy, and Normal. All the data sets used here are collected from the Kuwait Bangladesh Friendship Govt. Hospital as a real-time data set and personal photography. Where, the primary dataset was around 600 and data augmentation done for better perfermance. In this study, we applied three Transfer learning models along with our own customized Convolutional Neural Network where MobileNetV2 performed so well and give the highest accuracy of 99.30% and our customized CNN model shows with the accuracy of 96.52% on emotion detection from baby faces. With a highest accuracy, MobileNetV2 outperformed Xception and DenseNet201. The custom CNN also performed well, outperforming VGG19. The performances shows the efficiency of MobileNetV2 and other advanced deep learning models for baby facial emotion identification. In future, we aim to increase our primary datasets since we intended to implement those dataset for an IoT Based Baby Monitoring System.