Cross-modal recipe retrieval via parallel- and cross-attention networks learning

摘要

Cross-modal recipe retrieval refers to the problem of retrieving a food image from a list of image candidates given a textual recipe as the query, or the reverse side. However, existing cross-modal recipe retrieval approaches mostly focus on learning the representations of images and recipes independently and sewing them up by projecting them into a common space. Such methods overlook the interplay between images and recipes, resulting in the suboptimal retrieval performance.Toward this end, we study the problem of cross-modal recipe retrieval from the viewpoint of parallel- and cross-attention networks learning. Specifically, we first exploit a parallel-attention network to independently learn the attention weights of components in images and recipes. Thereafter, a cross-attention network is proposed to explicitly learn the interplay between images and recipes, which simultaneously considers word-guided image attention and image-guided word attention. Lastly, the learnt representations of images and recipes stemming from parallel- and cross-attention networks are elaborately connected and optimized using a pairwise ranking loss. By experimenting on two datasets, we demonstrate the effectiveness and rationality of our proposed solution on the scope of both overall performance comparison and micro-level analyses.