Investigating the impact of CNN layers on dysgraphia handwriting image classification performance

The diagnostic and detection process for distinguishing dysgraphia handwriting is vital in the intervention procedure to personalise the severity level of children's handwriting at an early stage. Currently, many deep learning methods and developments focus on other different domains such as handwriting signals and computer-based screenings that offers several limitations in time consuming and intricate procedures. Therefore, the dysgraphia handwriting classification using convolutional neural networks (CNNs) has seen a lot of success in the domain of handwriting image-based data that offers clear structured topology in the regular lattice of pixels in the dysgraphia handwriting patterns. However, due to the diversity and different characteristics of handwriting patterns, the convolution operations such as local connectivity layers and architectural framework in CNN model are numerous and invariant to generalize the features map in learning model classification. Thus, the comparative study of CNN layers is presented to investigate the impact of the different number of layers based on automated feature extraction for classifying dysgraphia and non-dysgraphia handwriting images. Experimentally, five CNN models that differ in structural architecture layers namely CNN-1, CNN-2, CNN-3, CNN-4, and CNN-5 are trained and validated using synthetic letter images dataset to observe the performance of each model. The CNN models are evaluated based on the confusion matrix of predicted and actual classes. Overall, the experiment shows high accuracy gained when the number of feature extraction layers is increased, which is five convolution layers outperformed others with significant training accuracy of 97.2% and 95.86% validation accuracy. The testing accuracy of 87.44% has approved that the proposed model is relatively accurate.