Comparative Performance Analysis of a Custom CNN, VGG16, and the Proposed Emotion Net Model for Facial Emotion Recognition

Abstract

Facial emotion recognition (FER) plays a pivotal role in affective computing, healthcare monitoring, adaptive education, surveillance, and human computer interaction. This study benchmarks three deep learning solutions for seven class FER on the Extended Cohn Kanade (CK+) corpus: a compact custom convolutional neural network (CNN), a VGG16 model adapted through transfer learning, and a newly proposed architecture, EmotionNet. All three models are trained and evaluated under an identical preprocessing and augmentation protocol so that observed performance differences can be attributed to network design rather than to inconsistencies in data handling. The custom CNN is inexpensive to run, whereas VGG16 leverages pretrained visual features at the cost of considerably higher memory and computational demand. EmotionNet is built from stacked convolutional blocks with batch normalization, rectified linear unit (ReLU) activation, max pooling, and dropout, followed by fully connected classification layers, and it is embedded within an end  to  end real  time video inference pipeline. On the CK+ benchmark, the three models reach classification accuracies of 92.84%, 96.73%, and 98.86%, respectively, with EmotionNet additionally producing 98.54% precision, 98.31% recall, and a 98.42% F1 score using 42.70 million parameters and 7.82 GFLOPs of computation. These findings indicate that EmotionNet delivers the strongest recognition accuracy while maintaining a more favorable computation  to  accuracy trade  off than VGG16, whereas the custom CNN remains the most suitable choice when hardware resources are severely constrained.