Image Dense Captioning of Irregular Regions Based on Visual Saliency

Traditional Dense Captioning intends to describe local details of image with natural language. It usually uses target detection first and then describes the contents in the detected bounding box, which will make the description content rich. But captioning based on target detection often lacks the attention to the association between objects and the environment, or between the objects. And for now, there is no dense captioning method has the ability to deal with irregular areas. To solve these problems, we propose a visual-saliency based region division method. It focuses more on areas than just on objects. Based on the division, the local description of the irregular region is carried out. For each area, we combine the image with the target area to generate features, which are put into the caption model. We used the Visual Genome dataset for training and testing. Through experiments, our model is comparable to the baseline under the traditional bounding box. And the description of irregular region generated by our method is equally good. Our model performs well in image retrieval experiments and has less information redundancy. In the application, we support to manually select the region of interest on the image for description, for assist in expanding the dataset.