Facial Emotion Recognition Using Swarm Optimized Multi-Dimensional DeepNets with Losses Calculated by Cross Entropy Function

Abstract

The human face forms a canvas wherein various non-verbal expressions are communicated. These expressional cues and verbal communication represent the accurate perception of the actual intent. In many cases, a person may present an outward expression that might differ from the genuine emotion or the feeling that the person experiences. Even when people try to hide these emotions, the real emotions that are internally felt might reflect as facial expressions in the form of micro expressions. These micro expressions cannot be masked and reflect the actual emotional state of a person under study. Such micro expressions are on display for a tiny time frame, making it difficult for a typical person to spot and recognize them. This necessitates a place for Machine Learning, where machines can be trained to look for these micro expressions and categorize them once they are on display. The study’s primary purpose is to spot and correctly classify these micro expressions, which are very difficult for a casual observer to identify. This research improves upon the accuracy of the recognition by using a novel learning technique that not only captures and recognizes multimodal facial micro expressions but also has features for aligning, cropping, and superimposing these feature frames to produce highly accurate and consistent results. A modified variant of the deep learning architecture of Convolutional Neural Networks combined with the swarm-based optimality technique of the Artificial Bee Colony Algorithm is proposed to effectively get an accuracy of more than 85% in identifying and classifying these micro expressions in contrast to other algorithms that have relatively less accuracy. One of the main aspects of processing these expressions from video or live feeds is aligning the frames homographically and identifying these concise bursts of micro expressions, which significantly increases the accuracy of the outcomes. The proposed swarm-based technique handles this in the research to precisely align and crop the subsequent frames, resulting in much superior detection rates in identifying the micro expressions when on display.