Computer Vision Object Relationship Detection Based On Deep Learning

The understanding of image semantics has become increasingly critical as the field of computer vision continues to evolve.However,it is not easy to learn the complete high-level semantics directly from the image.Therefore,the object relationship detection task is an image-semantic intermediate task,which is the bridge connecting object detection and image understanding.The purpose of object relationship recognition in the image is to learn and recognize the semantic relationship between objects based on the type of objects identified by the computer.Object relations are usually expressed in the form of triples of'"subject-predicate-object".The subject and object are both types of objects,and the predicate indicates the semantic relationship between the two.Through in-depth research on object-relationship triples,breakthroughs can be made in the bottlenecks of advanced image understanding tasks such as visual question answering,image captioning.However,the object relationship detection task needs to combine all the information to model a relatively subjective semantic relationship representation,which presents new and higher challenges for feature learning:·The reasoning logic between images and semantics is unclear,and it is impossible to find an accurate way to model the relationship vector between objects;·There are many types of objects and there are thousands of semantic relationships.The combination of the two is endless,so the identification method is more flexible;·The data set for object relationship detection is complicated,the frequencies of var-ious types of relationships appear uneven,and the bias is serious;·Semantic relations are initially subjectively set by humans,so they are inseparable from the learning of prior knowledge of natural language.In order to better apply prior knowledge,this paper designs the algorithm flow of generating prior knowl-edge graph structure.At the same time,in order to suppress the excessive influence of the prior knowledge on the original semantics of the image,improvements have been made in the network training method.To this end,this paper focuses on the core of object relationship detection task-modeling of semantic relationships,proposes a new network model design,completes experiments and verifies performance,including:1.The generation of input features used in relation detection and the way of context transfer between features;2.Design a multi-layer relational convolution network suitable for this task,so that the generated object features include relevant panoramic information in the image;3.For an image,grasp the effective description of the obj ect relationship to expand the complete semantic description.To this end,we design corresponding models,the purpose of which is to learn the existence probability of each object relationship in order to filter out the main information.In general,this paper proposes a multi-relational graph convolution encoding-decoding framework.The experimental process is to use multi-relational graph convolutional net-works to encode visual features and spatial features,and send the processed features into The decoder and the probability relationship network obtain the probability scores cor-responding to each relationship.Finally,the prediction of the effective relationship is obtained through the calculation of the probability score.The paper chooses the Visual Genome which is a challenging object relation data set for experiments.The experimental results show that our model has reached the current optimal performance under all three indicators.The model has theoretical and research value and can be applied to subsequent image understanding research work.