Erasure Detection and Recognition on Children's Manuscript Essays

In this paper, we present a machine learning approach to classify erasures in children's manuscript essays. We aim at highlighting challenges faced by children through their learning process, specifically when it comes to their subjective ability of converting thoughts into written text. The image processing software receives an entire scanned essay as input, then it segments each word. Afterwards, a classifier decides whether each word is "clean" or if it corresponds to an "erasure". We evaluate several Machine Learning (ML) techniques, including Support Vector Machine (SVM), boosted and bagged decision trees, k-nearest neighbor, Naive Bayes, discriminant analysis,logistic regression, and deep neural networks. To this end, manuscript essays containing erasures were scanned, each word(with and without erasure) was segmented, and were separated into two classes, providing knowledge to the system for the identification of new samples. Due to the difference between image dimensions, it was necessary to use features capable of describing the image independently of it corresponding to a letter or a word, or character dimension used by different children during writing. To solve this problem, we identify strokes directions, usually consisting of parallel lines. The obtained results show that the image descriptors when used alone achieve maximum accuracy limits of the order of 90%, but combined with the others, they exceed 98% accuracy. Compared to other used classifiers, SVM with Gaussian kernel produced higher accuracy.