Deep Learning-Based Classification of Dysgraphia Severity in Children from Handwriting Images

Abstract

This study focused on the design and testing of a deep learning model which could be used to distinguish between 'Low Potential Dysgraphia' and 'High Potential Dysgraphia', based upon using these handwriting image samples for children at school age. One of the methodologies was to create a Convolutional Neural Network (CNN) by using augmented samples of handwriting image data for training. These samples were divided into separate training and testing batches. All the handwriting samples were resized to 150x150 pixels and pre-processed. The CNN was made up of different convolutional and max-pooling layers, with dropout regularization included. At the end came a dense layer with sigmoid activation for binary classification. Training lasted for 10 epochs, with an emphasis placed on tracking accuracy at both the training and testing phases. Testing accuracy was about 84%; the training score also increased steadily, so the model is learning well without overfitting because its testing and training accuracies are similar. The model's performance in identifying 'High Potential Dysgraphia', was reflected by a precision of 0.76, a recall rate of 0.96, and an F1-score (the geometric mean between them)of respectively: On the contrary, 'Low Potential Dysgraphia' was identified, with a precision of 0.95 and falling recall at 0.74 giving an F1-score of 83 %.