EmoSense: Pioneering Facial Emotion Recognition with Precision Through Model Optimization and Face Emotion Constraints

Facial emotions are prime characteristics of humans that reflect the attributes of a human emotional condition. Facial Emotional Recognition (FER) is a significant research area since it has several significant applications such as crime detection, security, government sectors, surveillance, etc. Many researchers have addressed FER and put efforts into improving recognition using various methods such as Deep Convolutional Neural Networks (DCNN), Local Binary Pattern Convolutional Neural Networks (LBPCNN) and Micro Expression Recognition (MER). Nevertheless, there is a dearth of better Convolutional Neural Networks (CNN) for better accuracy of FER. Various methods have already proven their accuracies on datasets such as Fer_2013, CK48 and Legend. There are many challenges such as varying positions of images, illumination and noise etc. are not resolved yet. In this direction, the proposed modified CNN model has a combination of CNN layers to train the model and achieved better training and validation accuracy on datasets. The proposed method has three-fold contributions (1) to propose an efficient CNN model for FER and (2) to train the proposed model on three standard datasets 3) to test on real-time images. The proposed model is trained on three standard datasets (Fer_2013, CK48, and Legend) and achieves significantly higher training accuracy 92.34%, 99.84%, and 89.86%, respectively as compared to previous methods. Furthermore, the model demonstrates impressive generalization ability by achieving even higher test accuracy on real-world images 91.54%, 96.27%, and 87.47%, respectively. Also, the proposed novel CNN model has addressed all current challenges of FER with low computational complexity and prediction of facial expression more efficiently.