| Summary: | Visual Question Answering (VQA) is a challenging multi-modal learning task since it requires an understanding of both visual and textual modalities simultaneously. Therefore, the approaches used to represent the images and questions in a fine-grained manner play key roles in the performance. In order to obtain the fine-grained image and question representations, we develop a co-attention mechanism using an end-to-end deep network architecture to jointly learn both the image and the question features. Specifically, textual attention implemented by a self-attention model will reduce unrelated information and extract more discriminative features for question-level representations, which is in turn used to guide visual attention. We also note that a lot of finished works use complex models to extract feature representations but neglect to use high-level information summary such as question types in learning. Hence, we introduce the question type in our work by directly concatenating it with the multi-modal joint representation to narrow down the candidate answer space. A new network architecture combining the proposed co-attention mechanism and question type provides a unified model for VQA. The extensive experiments on two public datasets demonstrate the effectiveness of our model as compared with several state-of-the-art approaches.
|
|---|