Lightweight Facial Expressions Analysis with MobileNet

Facial expressions recognition is an exciting study area within Computer vision applications affecting estimating and human-calculating interaction. The detection and categorization of seven universal emotional states such as anger, neutrality, sad, fear, pleasure, disgust, surprise, face expression are essential in various computer vision applications. The primary focus of this study is on the extraction and categorization of these expressions using convolutional neural networks (CNNs). Over the last 10 years, CNNs and MobileNet have been a prominent focus of human-computer interaction research because of their superior feature-learning capabilities. Current models perform much worse on real-time, uncontrolled datasets than they do on controlled laboratory datasets while exhibiting remarkable accuracy in facial emotion recognition. Issues including deteriorated picture quality, occlusions, fluctuating illumination, and changes in head posture are frequently seen in photos taken from unstructured sources like the internet. By using deep learning approaches to interpret real-time photos, especially those with lesser quality, the goal here is to increase FER's recognition accuracy. For the experiment's purpose, FER2013 and Random dataset are utilized. The proposed method outperforms existing methods in accuracy, achieving a Train Accuracy of 50% and Validation Accuracy of 55% on benchmark datasets.