Comparing CNN-based Architectures for Dysgraphia Handwriting Classification Performance

Abstract

Deep learning algorithms are increasingly being used to diagnose dysgraphia by concentrating on the issue of uneven handwriting characteristics, which is common among children in the early stage of basic learning of reading and writing skills. Convolutional Neural Network (CNN) is a deep learning model popular for classification tasks, including the dysgraphia detection process in assisting traditional diagnosis procedures. The CNN-based model is usually constructed by combining layers in the extraction network to capture the features of offline handwriting images before the classification network. However, concerns have been expressed regarding the limited study comparing the performance of the Directed Acyclic Graph (DAG) and Sequential Networks in handwriting-related studies in identifying dysgraphia. The proposed method was employed in this study to compare the two network structures utilized for feature extraction in classifying dysgraphia handwriting To eliminate this gap. Therefore, a new layer structure design in the Sequential and DAG networks was proposed to compare the performance of two feature extraction layers. The findings demonstrated that the DAG network outperforms the Sequential network with 1.75% higher accuracy in classification testing based on confusion matrix analysis. The study provides valuable insights into the efficiency of various network structures in recognizing inconsistencies identified in dysgraphia handwriting, underlining the need for additional research and improvement in this field. Subsequently, these findings highlight the necessity of deep learning approaches to advance dysgraphia identification and establish the framework for future research.