Rating Pre-writing Skills in Ecuadorian Children: A Preliminary Study Based on Transfer Learning, Hyperparameter Tuning, and Deep Learning

Pre-writing skills are a set of skills that have a great impact on children’s self-esteem and academic achievement. This is a big problem in the development of children. Educators and therapists don’t have much time to detect children lacking these skills in an early stage. To deal with this problem, we present a system that helps evaluate drawings of geometric shapes (circle, triangle, and square) drawn by children. To reach this goal, the system is divided into 4 phases: first the images are divided into 2 categories. Second, the dataset is loaded and a data augmentation is applied. Third, the following Convolutional Neural Networks are applied: CNN with hyperparameter tuning, VGG16, ResNet50 and InceptionV3. To feed these models, we used the “UPSWriting-Skills” dataset. Fourth, results are shown and the best models are selected to present their final hyperparameters. Finally, this system could be the base for new experiments with other Convolutional Neural Networks and other techniques to improve the accuracy of the system.