Exploring the EmoBone Dataset with Bi-Directional LSTM for Emotion Recognition via Bone Conducted Speech

Facial expression recognition (FER) serves as a vi- tal interface for bridging human emotions and ma- chine understanding, enabling applications across psy- chology, healthcare, and human-computer interaction. This study explores the performance of machine learn- ing classifiers—SVM, Random Forest, KNN, and Deci- sion Tree—on the CK+ dataset, a benchmark for FER research. Preprocessing techniques, such as grayscale conversion and histogram equalization, were employed to enhance feature clarity. Features extracted via His- togram of Oriented Gradients (HOG) were evaluated us- ing k-fold cross-validation. SVM emerged as the most ac- curate classifier, achieving a 100% recognition rate with a linear kernel, while Random Forest demonstrated ro- bust but slightly inferior performance. Decision Tree and KNN exhibited lower accuracies, highlighting the trade- offs between interpretability and performance. These findings underline the potential of SVM for designing reliable and efficient FER systems suitable for practical applications.