EMOVISION Real-time multi-face emotion recognition using Deep Learning

As online video conferencing platforms have become more popular, it has become important to know how participants feel in order to improve engagement, communication, and behavior analysis. Most current emotion recognition systems need direct access to a webcam and often have trouble making predictions that are not always right and not being able to work well in real-time meeting settings. This paper presents EmoVision, a real-time multi-face emotion recognition system developed for video conferencing. The proposed system captures the meeting window directly from the screen instead of using the webcam. This makes it easy to use with existing platforms like Google Meet without needing to change the platform. We use two convolutional neural network (CNN) models that were trained on different input resolutions (48×48 and 64×64) to preprocess and analyze the detected facial regions from the captured frames. A decision fusion mechanism that uses confidence combines predictions from both models to make them more reliable and less likely to make mistakes. To fix problems with frame-by-frame predictions that aren't stable, a sliding window is used to smooth out the most recent frames.The system can find multiple people in real time and show emotion labels and confidence scores right on the video feed. Experimental findings indicate that EmoVision provides stable and dependable emotion recognition in dynamic video conferencing environments, rendering it appropriate for online education, virtual collaboration, and behavioral analytics. Keywords : Emotion recognition, Facial expression analysis, Video conferencing analytics, CNN ensemble, Temporal smoothing, Screen capture, OpenCV, TensorFlow/Keras, Human– computer interaction, Engagement analysis.