3D Rhythmic Flow Modeling for Emotion Recognition from Handwriting

This study proposes a lightweight three-dimensional convolutional neural network (3D CNN) for emotion recognition from handwriting data. Unlike conventional approaches that rely on static image features or scalar statistics such as mean and maximum values, the proposed model reconstructs handwriting trajectories collected from digital tablets—comprising coordinates (x, y), pressure, pen status, and time—into three-dimensional spatiotemporal tensors. This volumetric representation captures both rhythmic and spatial variations in handwriting, enabling the network to learn emotional cues embedded in kinematic dynamics such as velocity, curvature, and pressure modulation. Experiments were conducted on the EMOTHAW dataset, which includes handwriting and drawing samples labeled across three emotional dimensions: anxiety, stress, and depression. The proposed model achieved consistent classification performance with an average variation of approximately ± 0.15 across tasks, reaching a maximum F1-score of 0.78 on the stress dimension. These results demonstrate stable convergence and effective learning of the spatiotemporal structure of handwriting, even under limited data conditions. Overall, this study highlights the potential of 3D CNNs as efficient and interpretable models for emotion recognition from handwriting in small-scale datasets.