Avaliação de métodos de otimização em redes neurais convolucionais para identificação de dígitos

Abstract

This work proposes the creation of software aimed at recognizing words in images, through a neural network, analyzing discrepancies between a network trained by classical methods, such as gradient descent, and probabilistic approaches. The central objective is to provide support to the academic community and those interested in neural networks, character identification, gradient descent and probabilistic methods. The focus is on the development of character identification software using Artificial Neural Networks (ANN) with different backpropagation methods. The methodology used involved the use of Eclipse 2023-06 and Scene Builder to create the graphical interface, with Java version 8 and JavaFX as the programming base. The software was designed with two main aspects: one aimed at users (“user screen” and “help”) and the other intended for internal testing (“dev screen” and “train CNN”). During the tests carried out, the execution of the algorithm using the gradient descent method presented results similar to those obtained with Particle Swarm Optimization (PSO), with the latter demonstrating slightly better performance, reaching an accuracy of 72.5%. However, the limitation of the testing machine, without the addition of a GPU for processing, made the classification of letters unfeasible, resulting in the classification of handwritten digits. It is important to highlight that these tests were conducted on small-scale image sets, revealing limitations of PSO when applied to larger image sets. A possible solution to deal with this issue would be to improve the hardware of the testing machine, through the addition of a GPU for processing.