Stacked squeeze-and-excitation recurrent residual network for visual-semantic matching

Abstract

In recent years, visual-textual matching has been widely studied in the intersection of computer vision and natural language processing communities. A feasible scheme for learning discriminative representations is leveraging hierarchical features to align both modalities at multiple semantic levels. However, most existing approaches rely on pre-trained object detectors or semantic parsers to generate multi-level representations, whose performance is overly dependent on the extra supervision and thereby leads to its vulnerability. In this paper, we introduce a Stacked Squeeze-and-Excitation Recurrent Residual Network (SER2-Net) for visual-textual matching. Firstly, an efficient multi-level representation module is presented to produce a series of semantically discriminative features without the aid of extra supervision, which is built by stacking the squeeze-and-excitation recurrent residual (SER2) learning components. Specifically, SER2 incorporates the residual learning and inverse recurrent connection into the squeeze-and-excitation learning block, which allows for utilizing complementary current information and residual information to improve the modality-specific representation ability. Besides, to capture the implicit correlations contained among multi-level features, we propose a novel objective namely Cross-modal Semantic Discrepancy (CMSD) loss, which is characterized by exploiting the interdependency among different semantic levels to narrow the cross-modal distribution discrepancy. Extensive experiments on two benchmark datasets validate the superiority of our model, which compares favorably with the state-of-the-art approaches.