Object-based Image Discrimination Relationship Recognition

Abstract

With the development of self-media, we can publish images freely on the Internet, and there will be some discriminatory images more or less. Discrimination is an emotional relationship. However, there is no open data set to help us complete the task of discriminatory image detection. In this article, we created a simple data set that can be used to detect discriminatory relationships. Then, on this data set, we propose a baseline for discriminatory relationship detection. This baseline is a multi-task network based on translation embedding model, which can output both the action categories of objects in the graph and the relationship categories between the two. Experiments show that our multi-task network can effectively detect relationships, and it is better than the simple translation embedding model.