Bachelor Thesis

Bachelor Thesis: A Weakly-Supervised Image Classification Framework Based on Graph Convolutional Networks

• A feasible framework based on graph convolutional networks (GCN) for weakly supervised image classification is built, including deep feature extraction, graph construction, and weakly supervised graph convolutional networks.
• Considering a single image as a vertex, autoencoder and SimCLR are employed to extract discriminative node features. Then two similarity-based strategies for adjacency matrix construction are designed.
• Initial residual connection and identity mapping methods are injected into SelfSAGCN, a model for weakly-supervised node classification, to ease the over-smoothing problem and improve its performance.
• The framework achieves better performance on CIFAR-10 and STL-10 datasets, compared with existing weakly-supervised image classification models based on GCNs.