Classificação contextual de imagens por meio de redes neurais convolucionais de grafos e poda por similaridade

Abstract

Computer vision techniques have become popular in the technology industry. This success is derived from advances in computational resources and the generation of complex data. In this way, the use of resources such as convolutional neural networks for classification tasks was widespread among companies. Although such networks present interesting results for image classification, their difficulty in dealing with data relationship aspects opened a possibility for the use of graph convolutional neural networks. However, these networks have a bottleneck related to the scalability of resources, which is not well explored by the state of the art for the context of computer vision. In this way, this research proposes an approach with the objective of reducing the consumption of computational resources of the graph convolutional neural networks. In this research, there were several tests executed, which revealed this work can be promissory in many scenarios, capable of reducing the training effort by reductions in the memory consumption of up to 60.24% and memory usage from a graphical card in 23.11%. The work can contribute to the reduction of computing resources in data centers, being one of the main offenders in terms of costs and carbon emissions.