Vision transformer-based model for early detection of dysgraphia among school students

Abstract

Learning disorders, an umbrella term for a range of learning difficulties, impair a person's capacity to learn new skills. Dysgraphia is one of the prevalent learning disorders among children all over the globe. It is defined as a child's functional restriction in establishing accurate letter or word construction, inadequate speed, and readability of written text. Lack of availability of experts who can diagnose and high diagnostic cost, makes it important to discover a diagnostic approach for dysgraphia that is accurate, accessible and simple to use. Analyzing handwriting is the most common technique to detect dysgraphia which can be automated through image processing techniques. The use of deep learning algorithms has become increasingly widespread in image processing over the course of the last few decades and analysis. However, the effective classification of these handwritten images presents a number of challenges like low accuracy, inadequate availability of labelled data for training purposes. Considering the notable efficacy demonstrated by the Vision Transformer (ViT) in image classification, we proposed a ViT-based classification model in this paper. This model splits handwriting images into patches and then process those through a transformer. Just like word embedding, these input image patches are passed in a sequence to the transformer. To compare performance of proposed model, we also applied transfer learning techniques VGG16, VGG19, ResNet50 and InceptionV3. After comparing the results, it was found that Vision Transformers are best suitable for the classification. Vision transformer has outperformed with macro average F1 score value of 0.92 for the classification. Out of all pretrained models VGG16 performed best with the macro average F1 score value of 0.90. The findings of this study indicate that ViT-based model has the potential to assist experts in the early detection of dysgraphia.