Research On Visual Relationship Detection Based On Deep Learning

Deep learning is a research hotspot in recent years.It's performance in many fields such as image,speech and text is excellent.The research of deep learning mainly refers to deep neural network.Visual relationship refers to the relationship between objects in an image,such as "man driving a car","cup on the table" and so on.Visual relationship detection is one of the important tasks in computer vision.It is important to understand image content and connect images and text.This paper focuses on the task of visual relationship detection.The main research work is as follows:1)Analyze and research the relevant research basis of visual relationship detection tasks,the latest research methods and progress at home and abroad2)A visual relationship detection method based on multiple features is proposed.The method is based on the deep learning model framework,which can comprehensively utilize the local information and global information of the input image to effectively predict the relationship between objects in the image.Using the statistical features in the relevant data sets,the accuracy of the model for visual relationship prediction is further improved.3)A method of introducing graph method into visual relationship detection task is proposed.The graph method is used to extract the feature containing common sense information from the external knowledge map.Experiments show that the method further improves the accuracy of visual relationship detection.This paper comprehensively utilizes a variety of features for the task of visual relationship detection.A variety of features represent information from observations of different aspects of the picture,making full use of the information provided by the input picture.In addition,this paper also proposes to use the word embedding or graph method to introduce external information,which can play a complementary role and improve the prediction accuracy.