Visão computacional com redes neurais convolucionais: uma abordagem histórica, teórica e prática

Abstract

This paper presents a historical, theoretical, and practical review of convolutional neural networks applied to computer vision. It provides the main points in the timeline of the development of Artificial Intelligence and Neural Networks. Then, the architecture and description of the components of a convolutional neural network are introduced, going through artificial neurons, activation functions, convolutional layer, subsampling layer and fully connected layer. The most common regularization techniques are presented. Four high-performance convolutional networks are examined: LeNet-5, AlexNet, VGGNet and GoogleLeNet. Finally, the MNIST and Cifar10 databases are used to evaluate the performance of three of the mentioned networks. The LeNet-5 network presents excellent results, due to its simplicity and task specificity. The AlexNet network presents average results, affected by computational limitations. The VGGNet network presents unsatisfactory results, being a network with high computational cost and great losses with the necessary restrictions for its reproduction.