Reconhecimento de palavras manuscritas infantis utilizando deep learning

Abstract

The recognition of children's handwriting is a significant challenge in the field of computer vision due to the high variability, inconsistency, and imprecision found in the writing of children in the early stages of literacy. This work aims to develop a recognition system for handwritten children's words using Convolutional Neural Networks (CNNs) with deep learning techniques. To achieve this, a dataset consisting of 3,476 images of words written by children aged 6 to 10 was used. The images were processed and classified into 84 classes corresponding to an educational lexicon. Preprocessing techniques were applied, and VGG-16 and ResNet-50 models were trained, testing the application of data augmentation and transfer learning using models pre-trained on the IAM dataset. The models were evaluated using metrics such as accuracy, precision and Word Error Rate (WER). The results showed accuracy rates of up to 96.55% in word recognition, validating the effectiveness of the proposed approach. The system has potential applications in educational environments, particularly in the form of digital games, contributing to the literacy process in an inclusive and interactive way.