Semantic Compositional Networks for Visual Captioning

TLDR

The authors develop a Semantic Compositional Network that generates image and video captions by detecting semantic tags and using them to compose LSTM parameters. SCN replaces each LSTM weight matrix with a set of tag‑dependent matrices, weighting each by the image‑derived probability of its tag to compose the network parameters. The method significantly outperforms previous state‑of‑the‑art captioning models on COCO, Flickr30k, and Youtube2Text according to multiple evaluation metrics.

Abstract

A Semantic Compositional Network (SCN) is developed for image captioning, in which semantic concepts (i.e., tags) are detected from the image, and the probability of each tag is used to compose the parameters in a long short-term memory (LSTM) network. The SCN extends each weight matrix of the LSTM to an ensemble of tag-dependent weight matrices. The degree to which each member of the ensemble is used to generate an image caption is tied to the image-dependent probability of the corresponding tag. In addition to captioning images, we also extend the SCN to generate captions for video clips. We qualitatively analyze semantic composition in SCNs, and quantitatively evaluate the algorithm on three benchmark datasets: COCO, Flickr30k, and Youtube2Text. Experimental results show that the proposed method significantly outperforms prior state-of-the-art approaches, across multiple evaluation metrics.