M-GCN: Brain-inspired memory graph convolutional network for multi-label image recognition

Traditional single-label classification methods can not be effectively applied in multi-label classification due to the semantic correlation. Conventional methods using the attention mechanism or prior knowledge, lacks deep semantic correlations, resulting in degradation for detection performance. Considering the hippocampal circuit and memory mechanism of human brain, a brain-inspired Memory Graph Convolutional Network (M-GCN) is proposed. M-GCN presents crucial short-term and long-term memory modules to interact attention and prior knowledge, learning complex semantic enhancement, and suppression. We evaluate the effectiveness of our method on public benchmarks (Microsoft COCO and PASCAL VOC). Extensive experiments demonstrate that M-GCN outperforms general state-of-the-art methods and shows the advantages in semantic correlation and complexity comparing with traditional memory models.