Optimal Image Feature Ranking and Fusion for Visual Question Answering

摘要

Visual Question Answering (VQA) is a moderately new and challenging multi-modal task, which endeavors to discover an answer for a given pair of an image and a relating question. This AI-complete task gains attraction from numerous researchers from the areas computer vision (CV) and natural language processing (NLP) due to its various potential applications. The general flow of VQA algorithms consists of image feature extraction, question feature extraction and joint comprehension of these two to generate an appropriate answer. Existing VQA systems did not pay attention to input feature extraction, but only celebrated different ways of multimodal embedding. This paper proposes to improve the task of VQA by feature-level fusion of visual information. The goal of feature fusion is to consolidate relevant information from two or more feature vectors into a solitary one with additional discriminative power. Unlike simple concatenation, this paper uses discriminative correlation analysis (DCA) for fusion, which is the only method that incorporates the class structure into the feature-level fusion. Since the VQA systems are generally modeled as classification systems by treating the correct answers as classes, class-specific DCA suits well here. The newly created fused feature vectors are close to the right answers and thus raise the role of image understanding in VQA. The experimental results show the effectiveness of the new approach on DAQUAR dataset with mutual information (MI) as an evaluation metric.