Melhorias na classificação de imagens contextuais através de grafos convolucionais e caminhada aleatória

Abstract

Humans can manage to assimilate the context between information from little perceived data, and this statement applies to the context in images. In machine learning, an image’s context can be represented and predicted using Graph Convolutional Networks (GCNs). In this context, graph convolutional networks need the graph structure to apply the convolutional method between nodes. Based on previous work, fully connected and threshold weights are edge policies that can be used to create image context graphs. However, the connection policies between objects generally used in the creation of convolutional graphs may not be the best in terms of classification or number of edges, which may confuse the classification models. Aiming to make context graphs more expressive and improve the accuracy of contextual image classifications, this paper proposes an efficient method based in random walks that can get better classification results compared to the literature. To increase the level of comparison of the proposed method, two other random methods were used; random cut and random edge. Some pre-trained neural network architectures were used to extract features of each object from the images. Different sizes of walks were also tested for the random walk, making the project more comprehensive. Were used different label contexts in four datasets to evaluate the graph edge connection policies. The proposed approach could perform better results in almost all the datasets and having edge sets smaller than the fully connected policy in all datasets. The proposed method could perform 56.20% of precision compared to 50.37% of the fully connected policie and 53.85% of the threshold weights method.