Development of a Fusion-Based Facial Expression Dataset and a Machine Learning Approach for Emotion Recognition

Received: 31 December 2025 | Revised: 7 February 2026 and 25 February 2026 | Accepted: 27 February 2026 | Online: 4 April 2026

Corresponding author: Sunil S. Harakannanavar

Abstract

Facial Expression Recognition (FER) is a key part of affective computing, where it is still difficult to tell the difference between small changes in facial expressions. Many current FER methods depend on single feature descriptors, which do not always extract the extra structural and texture information that facial expressions have. To overcome this limitation, this paper introduces an adaptive weighted feature-level fusion framework that combines a modified Histogram of Oriented Gradients (HoG), Local Binary Patterns (LBP), and Fast Keypoint Detector with Binary Robust Independent Elementary Features (FKBD). The proposed method assigns discriminative weights to each descriptor instead of directly concatenating features. This lets the fusion process focus on important facial cues and ignore unnecessary ones. The weighted feature representation that comes out of this makes it easier to tell classes apart and makes it more resistant to changes in expression. Multiclass Support Vector Machine (SVM) and k-Nearest Neighbors (KNN) classifiers are used to classify expressions. Experimental validation is performed on standard facial expression datasets, such as the CK+ and FERG DB. The suggested framework achieves 98.74% accuracy with SVM and 95.60% accuracy with KNN, outperforming recent FER methods that use handcrafted features. These results show that using adaptive weighting on complementary descriptors is an effective and computationally efficient way to recognize facial expressions.