Graph Convolutional Network (GCN) is a graph-based learning framework that goes beyond the limitations of regular grid-sampled convolution operated by traditional CNN. By incorporating graph topological structure to the network, GCN is capable of encoding and aggregating features from neighboring nodes and exploiting the data nature in a non-linear way. For HSI data of cultural heritage artifacts, the lack of ground-truth labels imposes extra difficulty on the hyperspectral image classification task using a learning framework. For differentiating subtle spectral differences between varied pigments, we develop a two-layer semi-supervised GCN model for pigment classification, which can be directly applied to an adjacent region of interest (ROI) for automatically clustering unseen pixels without ground-truth labels. We first explore graph construction approaches in depth, and focus on two advanced graph construction strategies: Density-weighted kNN and Natural Nearest Neighbor (NNN). By assigning adaptive number of neighbors based on local data density, a local-scale adaptive graph structure can be built and infused into the GCN model. The experiments on the Selden Map and the Gough Map prove the validity of our method. Under semi-supervised learning framework, SAD+DWkNN GCN model can achieve near 90% classification accuracy for each class based on less than 30% labeled training samples. Compared to the ground-truth labels produced by the Graph Modularity algorithm, the GCN classification map matches even better to the actual spectral characteristics of each material via spectral analysis. The learning and generalization ability of the proposed GCN model shows a promising prospect for applying Graph Convolutional Network for a large-scale HSI clustering task. This research can alleviate the dilemma of deficient labeled samples and aid historians for analyzing varied pigments composition in a more intelligent manner.
|