Reconhecimento de objetos em uma cena utilizando redes neurais convolucionais

Abstract

Identifying and classifying objects in an image is a process that has attracted attention in recent years due to the evolution of artificial neural networks. Today there are artificial neural networks capable of being trained with N classes and still maintain a high accuracy for predicting objects. The convolutional neural networks are the main ones when it comes to images because they are designed with a focus on finding objects in scenes. Thus, both the classification and location of cans and bottles in which the proposed objects are made possible through convolutional neural networks. It is important to stress that the training base of a convolutional neural network is summarized in large-scale computation, which in turn requires a lot of processing, so the use of tools that bring benefits in relation to the cost of processing such as GPUs and of CUDA technology was an indisputable criterion, since without them it would not be possible to use the ideal network configurations and also negatively impact the training time of the network. Finally, the results obtained with convolutional neural network training to detect cans and bottles reached expectations, considering that the network reached an average accuracy of 87.55%, in a time interval of approximately four days of training.