An Improved Graph Convolution Network for Robust Image Retrieval

Image retrieval is one of the most critical foundations for many content-based search applications. However, the image retrieval methods have to balance demands on both training accuracy and generalization effectiveness. In this paper, we propose a graph convolution network (GCN) to improve retrieval robustness by integrating the constructs of normalized residual network (NRN) model and feature dropout (FD) operations. The normalized residual networks use skip connection and normalize vectors in each layer to enhance the learning and strengthen the generalization ability. The feature dropout step randomly discards a portion of features in the network to prevent the model from overfitting. We tested our proposed model on several benchmark datasets and the experiment results showed an improvement of 1-3 mAP in comparison with the state-of-the-art Guided Similarity Separation (GSS) algorithm.