ANALYSING DEEP CNN COMPONENTS FOR EMOTION CLASSIFICATION: AN ABLATION STUDY ON ICML FACE DATASET

Abstract

Facial emotion recognition (FER) is a crucial application of deep learning, with significant implications in human-computer interaction, mental health analysis, and affective computing. This study conducts an ablation analysis on various architectural components of a Deep Convolutional Neural Network (CNN) for emotion classification using the ICML Face Dataset. We investigate the impact of Batch Normalization, Dropout, and Network Depth on model performance. Our baseline CNN achieves 48.05% accuracy, while removing Batch Normalization unexpectedly improves performance to 54.25%, suggesting its potential inefficacy in this dataset. Conversely, removing Dropout reduces accuracy to 47.26%, indicating its importance in generalization. A shallower network further degrades performance to 46.27%, highlighting the necessity of deeper architectures for complex feature extraction. Finally, an optimized CNN integrating L2 regularization, Batch Normalization, and 50% Dropout achieves 80.38% accuracy, demonstrating substantial improvements. These findings provide insights into architectural design choices for enhancing facial emotion recognition models and highlight the significance of regularization techniques in achieving robust generalization.