Redes neurais: uma aplicação de visão computacional.

Abstract

Artificial neural networks have become essential in various fields due to their ability to solve complex problems. Analyzing different architectures is important to optimize model performance in specific tasks. This undergraduate thesis explores the application of artifi- cial neural networks, focusing on fully connected and convolutional neuron architectures, in computer vision problems. The main objective is to compare the performance of these architectures using the Fashion MNIST dataset, composed of images of fashion articles. Sequential and convolutional network models were implemented and trained with regula- rization techniques such as Batch Normalization and Dropout to analyze the impact of these approaches on image classification performance. The results indicate that convolu- tional networks outperform sequential networks in terms of accuracy and generalization capacity, especially when combined with normalization and regularization techniques. It is concluded that the convolutional architecture, although more complex, offers better results in computer vision tasks.