A DEEP LEARNING BASED APPROACH FOR FACIAL EXPRESSION RECOGNITION BY USING CROWD SOURCED LABELLED DATA

Abstract

Facial Expression Recognition (FER) has emerged as a vital area in the fields of affective computing and intelligent systems, enabling machines to interpret human emotions in diverse applications such as healthcare, education, surveillance, and human-computer interaction. This thesis presents a deep learning-based approach to FER that leverages crowd-sourced labeled data, which, while cost-effective and scalable, often suffers from annotation inconsistencies and label noise. To address these challenges, a Convolutional Neural Network (CNN) architecture was developed using the Keras framework, trained on a grayscale emotion dataset spanning seven fundamental emotional states. The methodology incorporates key techniques such as real-time data augmentation, dropout regularization, adaptive learning rate tuning, label smoothing, and early stopping to improve generalization and reduce overfitting. The model was trained using stratified data splits and evaluated using accuracy, loss, confusion matrix, and class-wise precision, recall, and F1-score metrics. Results show that the final model achieved a robust 96% classification accuracy, with particularly high F1-scores (above 94%) across most emotion classes, including happiness, sadness, and surprise. These findings indicate strong generalization capabilities and resilience to annotation noise. Overall, the proposed system demonstrates that a carefully designed CNN can effectively learn from imperfect crowd-sourced data and outperform traditional FER models, offering a scalable and reliable solution for real-world emotion recognition tasks.