Deep graph fusion for graph based label propagation

A graph which represents the similarity between samples plays an important role in graph-based semi-supervised learning. Though for this case one of the most important keys is the feature descriptor used for similarity calculation. Recently deep neural networks have been expressed as powerful methods which simultaneously perform feature extraction and classification. Moreover, feature fusion at different levels can improve the performance. The contributions in this article are as follows; first, we use deep features that are extracted from the last layers of Convolutional Neural Network for graph construction. Second, we adopt feature fusion to improve discriminability. Thirdly, we propose deep graph fusion as a way to combine information while preserving the inlying structure of each descriptor. The experiments are conducted on two face databases and the proposed method is compared with three hand-crafted features. The results obtained using three label propagation techniques show that the use of deep features can improve the performance compared to hand-crafted features. Moreover, the fusion in graph level can improve the results compared to the use of each deep feature alone.