Principais configurações na integração de visão computacional e aprendizagem profunda: algoritmos e técnicas

Abstract

Models based on Deep Learning have gain attention in several areas, mainly in Computer Vision. However, due to its extensive literature and the large number ofhyper parameters to be adjusted, it became a challenge for those who wish to develop solutions using this approach. Thus, the objective of this work is to synthesize, review and present the main concepts, configurations and applications that integrate Deep Learning and Computer Vision, through experiments and exploration of their configurations, methods and analyses. Families of Convolutional Neural Network, Autocoding Networks and architectures are explored, also Restricted Boltzmann Machines for pretraining, using the MNIST and CIFAR-10 image databases. The efficiency of the convolutional models in the classification task of the MNIST base and performance reduction when using the same configurations in the CIFAR-10 was observed. The Autocoder obtained satisfactory reconstruction error values and a good performance together with MobileNet in the classification task. The explored configurations can be used to help build solutions in Computer Vision tasks, and this work can be used as a guide for those who want to know or deepen their knowledge in Deep Learning applied to Computer Vision.