Multi-scale motivated neural network for image-text matching

Existing mainstream image-text matching methods usually measure the relevance of image-text pairs by capturing and aggregating the affinities between textual words and visual regions, while failing to consider the single-scale matching bias caused by the imbalance of image and text information. In this paper, we design a Multi-Scale Motivated Neural Network (MSMNN) model for image-text matching. In contrast to previous single-scale methods, MSMNN encourages neural networks to extract visual and textual features from three scales, including local features, global features and salient features, which can take full advantage of the complementarity of multi-scale matching to reduce the bias of single-scale matching. Also, we propose a cross-modal interaction module to realize the fusion of visual and textual features in local alignment, so as to discover the potential relationship between image-text pairs. Furthermore, we also propose a matching score fusion algorithm to fuse matching results from three different levels, which can be freely applied to other initial image-text matching results with a negligible overhead. Extensive experiments validate the effectiveness of our method, and the performance has achieved fairly competitive results on two well-known datasets, Flickr30K and MSCOCO, with a boost of 1.04% and 0.59% on evaluation metric mR compared with the advanced method.