A Convolutional Neural Network-Based Web Application for Facial Emotion Detection in Special Needs School Students

Abstract

Facial expressions are among the most dominant nonverbal cues used to communicate emotion, yet students with special needs in extraordinary schools (Sekolah Luar Biasa, SLB) often struggle to recognize and express their own emotions accurately, which can hinder social interaction. This study aimed to design and evaluate a Convolutional Neural Network (CNN)-based web application that helps teachers introduce basic emotions to SLB students in Madiun City through automatic facial emotion detection. A mixed-method approach was applied: a residual CNN architecture was trained on a combined dataset consisting of the public FER2013 corpus and locally collected facial images of SLB students, processed through resizing to 48×48 grayscale, pixel normalization, augmentation, and a 70:15:15 train–validation–test split; the trained model was then integrated into a Streamlit-based web interface and evaluated through black-box testing and teacher interviews. The model reached a training accuracy of about 71% and a validation accuracy of about 65%, with an overall test accuracy of 67% across seven emotion classes, performing best on the happy class (F1-score 0.87) and weakest on the fear class (F1-score 0.48). Black-box testing confirmed that all seven core functional scenarios operated as specified (100% acceptance), and teacher feedback indicated that the interface was simple, informative, and useful as a supportive rather than diagnostic tool for observing students’ emotional expressions before and during lessons. These findings suggest that a CNN-based facial emotion detection system is a feasible and accessible educational aid for special-needs classrooms, although confusion among visually similar emotions and reliance on single-frame capture indicate directions for further improvement.