Deep Learning Based Facial Emotion Recognition System

DOI

https://www.doi.org/10.59256/indjcst.20260502053

Published In

Volume 5, Issue 2 (2026)

Published Online

May-August 2026

Pages

466-478

Article Type

Original Article

Check for updates

Facial Emotion Recognition (FER) constitutes a fundamental building block of affective computing and human-computer interaction, equipping machines with the capacity to interpret and respond to human emotional states in real time. Even with considerable advances in deep learning, contemporary FER systems continue to face difficulties related to cross-dataset gen-eralization, ambiguity among visually similar emotion categories, and severe class imbalance inherent in curated laboratory datasets. This study proposes a domain-adapted deep Convolutional Neural Network (CNN) trained entirely from scratch on a unified corpus assembled from the Extended Cohn-Kanade (CK+) and Karolinska Directed Emotional Faces (KDEF) benchmarks. The merged dataset comprises 6,530 labeled grayscale facial images at 48×48 pixel resolution, partitioned into 5,224 training samples and 1,306 test samples distributed across ten emotion categories. The proposed network consists of four progressively expanding convolutional blocks, each incorporating paired 3×3 convolution layers, batch normalization, ReLU activation, 2×2 max-pooling, and spatial dropout, culminating in global average pooling followed by two fully connected layers with a ten-way softmax output. The model is optimized over 40 epochs using the Adam algorithm (η = 10−3) with ReduceL ROnPlateau scheduling and early stopping. On the held-out test partition, the proposed model attains a peak accuracy of 90.51% and a macro-averaged F1-score of 0.92, outperforming a fine-tuned MobileNetV2 transfer-learning baseline (88.06%, macro F1 = 0.90) while requiring 36% fewer parameters (1.44 M versus 2.26 M). Comprehensive per-class precision, recall, F1-score, and confusion-matrix analyses confirm the benefits of domain-adapted architectural decisions over generic ImageNet-pretrained backbones. Supplementary mathematical derivations encompassing receptive-field growth, computational complexity, information-theoretic loss bounds, and regularization analysis further substantiate the proposed framework.