Contextual Image Classification Through Fine-Tuned Graph Neural Networks

Nowadays, computer vision techniques have become popular in several domains (e.g., agriculture, industry, medicine, and others). Their success derives from the advances in computational resources and the large volume of complex data (i.e., images). These factors led to an increase in the use of convolutional neural networks. However, such deep learning architectures do not appropriately explore the relationships between the data (e.g., images) and their respective structure. To better gather and encodes these affinity connections into a deep neural network, we can use the so-called graph neural networks (GNNs). These graph-based networks also present drawbacks. The high number of relationships in a graph can be considered a bottleneck regarding the available resources and scalability. Hence, to mitigate this issue, we propose to use GNNs automatically tuned, defining their well-suited connections according to a given image context, which improves their efficiency and efficacy. We performed experiments considering different types of state-of-the-art deep features aggregated with the GNNs. The results demonstrate that our proposed method can achieve equal accuracy (statistically) to GNNs with complete and random connections. Moreover, we decreased the number of edges to a great extent (up to 96%), testifying to our method’s effectiveness.