2D-Convolution Based Feature Fusion for Cross-Modal Correlation Learning

Cross-modal information retrieval (CMIR) enables users to search for semantically relevant data of various modalities from a given query of one modality. The predominant challenge is to alleviate the "heterogeneous gap" between different modalities. For text-image retrieval, the typical solution is to project text features and image features into a common semantic space and measure the cross-modal similarity. However, semantically relevant data from different modalities usually contains imbalanced information. Aligning all the modalities in the same space will weaken modal-specific semantics and introduce unexpected noise. In this paper, we propose a novel CMIR framework based on multi-modal feature fusion. In this framework, the cross-modal similarity is measured by directly analyzing the fine-grained correlations between the text features and image features without common semantic space learning. Specifically, we preliminarily construct a cross-modal feature matrix to fuse the original visual and textural features. Then the 2D-convolutional networks are proposed to reason about inner-group relationships among features across modalities, resulting in fine-grained text-image representations. The cross-modal similarity is measured by a multi-layer perception based on the fused feature representations. We conduct extensive experiments on two representative CMIR datasets, i.e. English Wikipedia and TVGraz. Experimental results indicate that our model outperforms state-of-the-art methods significantly. Meanwhile, the proposed cross-modal feature fusion approach is more effective in the CMIR tasks compared with other feature fusion approaches.