The Higher-Quality High-Level Semantics Based On Improved Visual Model For Vqa

Visual question answering is a new research direction that combines computer vision and natural language processing. Some recent research uses image interpretation to reason about answers. However, the accuracy of the answers is not satisfactory due to insufficient understanding of image content. This paper proposes an improved VQA method which applies higher-quality image high-level semantics based an improved visual model, thereby optimizing the performance of answer prediction. Firstly, we use the improved Faster-RCNN model to extract more precise attended objects from images. Then, the corresponding visual features are used to obtain high-level semantic knowledge (refined image captions and attended high-level attributes) in images. Finally, the reasoning module uses the joint features of two semantic knowledge and questions to predict the final answers. We conducted extensive experiments on the public datasets and the results of experiments were analyzed to illustrate effectiveness of the method.