Deep Emotion Recognition Using Angle-Amplitude Graph Images and EfficientNet-B0 with Evaluation on KTU-MEDAFE and DEAP Datasets

Özkellekci B., Hatipoğlu Yılmaz B.

2026 6th International Conference on Electrical, Computer and Energy Technologies (ICECET), 6 - 09 Temmuz 2026, ss.1-5, (Tam Metin Bildiri)

Özet

Electroencephalography (EEG) signals provide an objective measure of neural activity associated with emotional states and constitute a key component of EEG-based emotion recognition. In this study, a signal-to-image transformation method is applied to EEG data from two multimodal emotion datasets: the benchmark DEAP dataset and KTU-MEDAFE, a multimodal emotion dataset developed by our research group. The resulting image representations of EEG signals are subsequently classified using EfficientNet, a convolutional neural network (CNN) architecture known for its efficiency and robust performance in image-based classification tasks. A unified processing pipeline is employed across both datasets to evaluate the effectiveness of signal-to-image representations for EEGbased emotion recognition and to investigate the applicability of EfficientNet to EEG-derived images. While the DEAP dataset enables direct comparison with existing studies, KTU-MEDAFE provides novel multimodal emotional data for further exploration. Experimental results demonstrate that, on the KTUMEDAFE dataset, the best average performance was achieved for the 1vs2 emotion pair, with an accuracy of 72.31% and an F1- score of 71.93%. On the DEAP dataset, the proposed approach achieved average accuracies of 92.55% for Valence (VAL) and 91.04% for Arousal (ARO), demonstrating strong performance on this benchmark dataset.