Implementation of a Fusion Classification Model for Efficient Pen-Holding Posture Detection

Abstract

Pen-holding postures (PHPs) can significantly affect the speed and quality of writing, and incorrect postures can lead to health problems. This paper presents and experimentally implements a methodology for quickly recognizing and correcting poor writing postures using a digital dot matrix pen. The method first extracts basic handwriting information, including page number, handwriting coordinates, movement trajectory, pen tip pressure, stroke sequence, and pen handling time. This information is then used to generate writing features that are fed into our proposed fusion classification model, which combines a simple parameter-free attention module for convolutional neural networks (CNNs) called NetworkSimAM, CNNs, and an extension of the well-known long short-term memory (LTSM) called Mogrifier LSTM or MLSTM. Finally, the method ends with a classification step (Softmax) to recognize the type of PHP. The implemented method achieves significant results through receiver operating characteristic (ROC) curves and loss functions, including a recognition accuracy of 72%, which is, for example, higher than that of the single-stroke model (i.e., TabNet incorporating SimAM). The obtained results show that a promising solution is provided for accurate and efficient PHP recognition and has the potential to improve writing speed and quality while reducing health problems induced by incorrect postures.