Identificador de letras manuscritas usando redes neurais convolucionais

Abstract

Optical character recognition (OCR) has been widely used in document scanning, facilitating the conversion of printed texts into digital formats. However, handwritten letter recognition still represents a significant challenge due to the wide variation in writing styles. This work aims to develop a handwritten character identification algorithm using computer vision, machine learning and convolutional neural networks techniques. The main objective is to create a model capable of recognizing different writing styles, trained with a large set of data to achieve satisfactory performance in different scenarios. The methodology employed included data collection and preprocessing, definition of the neural network architecture, application of regularization techniques and evaluation of the model both on internal data and in an external environment. To validate the robustness of the model, a completely new data set was collected in collaboration with the UTFPR community, allowing its effectiveness to be tested in different scenarios. The model was evaluated on both previously consolidated external data and on newly collected data, ensuring a comprehensive analysis of its performance. The results obtained demonstrated that the applied approach is promising, achieving Precision (90%), Recall (91%), F1-Score (90%) and Accuracy (91%) rates. Among the advantages of the developed method, the ability to adapt to different writing styles stands out, while the main limitations involve the need for a large volume of data for effective training. It is concluded that the objective of the work was achieved, contributing to the area of pattern recognition and paving the way for future improvements, such as the incorporation of new handwritten text processing techniques to increase the accuracy of the model.