Research On Image Detection And Recognition Algorithm Based On Graph Neural Networks

Graph Neural Network(GNN)is a deep neural network that worked on graph data structures.Essentially,a graph neural network captures the edge features from the graph through the information-transfer between the graph nodes.In which each node aggregates features from neighboring nodes in its convolutional layer.In this thesis,GNN is appiled to a image detection task,which has been a primary task in computer vision field.This task requires locating the position of an object in a picture and giving the class label of the item by a recognition algorithm.Currently,commonly used target detection models are based on Convolutional Neural Network(CNN),which mostly recognize the pixel information in a images for object localization and classification.The advantage of GNN-based methods is that by constructing the graph data structures(e.g.,spatial graphs,knowledge graphs,or super-pixel graphs),the models are able to obtain additional features and use them to analysis the objects,thus improving the detection performance.Therefore,several frameworks are proposed in this thesis to discuss the application of graph neural networks on image detection task.First,this thesis proposes an image spatial based network to detect the object.This network makes a reasonable determination of the label of objects as well as their specific locations by relying on the spatial informations from the images.The model adopts a two-channel structure with U-NET to extract the image pixel information and a graph convolutional network to extract the image spatial informations.The two features are fused with each other by a special gating mechanism to obtain the final output.Secondly,this thesis proposes a superpixel-based residual graph neural network algorithm.This algorithm achieves task-scale compression by transforming large-scale pixel data into tens of superpixels.This model constructs a superpixel graph by the relationship between the position of the superpixels.And designed a residual-based graph neural networkto solve the problem of oversmoothing.This thesis also proposes a gated graph neural network algorithm based on the knowledge graph.Semantic-level inference is performed on the knowledge graph to achieve accurate discrimination of objects that are difficult to be recognized.The model also proposes a semantic consistency model by studying the similarity of points on the knowledge graph,which can more accurately determine whether an image is suitable for knowledge inference.Finally,several comparison experiments are designed based on the COCO dataset and the VG dataset.The data show that the image detection model based on image spatial information and knowledge graph has a significant improvement in recognition performance compared with the baseline model.And the super-pixel-based image detection model can effectively reduce the amount of floating-point operations of the model,and reduce the model complexity.