Research On Visual Relationship Detection Technology Based On Deep Learning

In recent years,with the re-emergence of deep learning,artificial intelligence has developed explosively.One of the development focuses is to combine image and semantics.This is an emerging frontier.It can be effectively applied in the fields of image subtitles,image question and answer,image reasoning,etc.,and the basis for realizing these is to understand the predicate relationship between objects in the image.Through these relationships,the image can be understood deeper and the image semantics can be enriched.It is only an era of simply describing what is in the image and not the deep meaning expressed by the entire image.In visual relationship detection,a simple expression(Object1,Predicate,Object2)is used to simplify the relationship between objects in a sentence.In this paper,a relationship detection model superior to the current mainstream is proposed,which is called a deep learning-based relationship detection model.Firstly,the destination detection network Faster R-CNN is used to select the target,and then the target feature map is input into the specified feature fusion network.The function of the feature fusion network is to fuse multiple feature inputs,including semantic features,spatial features,and image features.Among them,two spatial feature extraction methods are designed for spatial features.One is to calculate the corresponding relationship between objects by traditional methods.It is based on the network to predict the spatial coordinates between them.Later,it also made some improvements to the loss function currently used for multi-label classification,so that the model's learning ability is stronger.Finally the new model is used to test the relationship detection and zero-classification learning tasks on the visual relationship data set.Through a series of results analysis,we get a new model to dominate the relationship detection than the current mainstream model results especially in the Zero-shot learning task(Zero-shot learning)has a huge advantage.