Abstract
Image retrieval is a challenging task of searching images similar to the query one from a database. Previous learning-based methods adopt various ingenious designs to increase the representatively positive and negative sample pairs in training. Still, these methods are performance immanently limited by the size of the mini-batch. To this end, we here introduce the learnable descriptor graph convolutional network (LDGC-Net), which effectively enhances the hard mining ability of the model and clears the boundary between different categories. We present an analysis of why our LDGC-Net can aggregate relationships between original descriptors in a constrained size of the mini-batch. Also, we propose an innovative end-to-end training framework with the LDGC-Net for image retrieval to accelerate model convergence. In particular, our LDGC-Net can be conveniently integrated into other current methods as a plug-and-play module with inappreciable computational cost. Experimental results in three benchmark datasets show that the proposed LDGC-Net can improve performance compared with several state-of-the-art approaches.