A Visual Memory Information Interaction Mechanism Inspired Graph Neural Network for Highly Imbalanced Image Classification

Human brains can recognize images regardless of the data distribution and not be affected by the sample size, since they have the ability of understanding and cognition based on visual memory. To better solve the highly imbalanced problems in image classification tasks, we research the theory of how visual memory works for information processing and image recognition in the temporal dimension. Inspired by the association of three kinds of visual memories and their interaction mechanism, this paper proposes a novel end-to-end framework, a Visual Memory Information inTeraction mechanism-inspired Graph Neural Network (VMIT_GNN). Three modules are sequentially designed and related to imitate information processing in different stages of visual memory, which helps better comprehend the minority based on incomplete and limited knowledge. Besides, to improve the cognition ability of the proposed model, we also explore the relation information among various samples by learning the function formulated based on the features of each pair. Extensive experiments on two highly imbalanced datasets are executed to validate the effectiveness of VMIT-GNN, and the results on multiple criteria verify the competitiveness of our method to the other comparisons.