Real-Time Recognition of Three Facial Emotions (“Surprise, Neutral, Happy”) Based on CNN with Augmented Data

Emotion recognition from facial expressions has become essential for applications such as human-computer collaboration, robot communication, and interactive interfaces [1]. This work proposes a real-time recognition system (i.e., the system produces a prediction with a latency low enough for smooth interaction, typically less than 100 ms per image or greater than 10 frames per second) capable of classifying three emotions: surprise, neutrality, and joy from facial images. The model is based on a convolutional neural network (CNN) optimized by data augmentation techniques applied to the FER2013 dataset [2] (Data augmentation was applied only to the training subset, not before distribution). The CNN has three convolutional layers, four fully connected layers, and uses ReLU and Softmax functions. The proposed approach achieves a validation accuracy of 89%, maintains high and balanced recognition rates for each class, and is capable of processing slightly distorted faces (i.e., faces with small geometric or photometric variations, such as rotations, translations, scaling changes, or partial expressions) [3]. These results demonstrate the feasibility of fast, robust emotion recognition applicable to real-time interactive scenarios.