Research On Image Co-saliency Detection Algorithm Based On Graph Neural Network

Co-saliency detection aims to capture the salient and common areas in a collection of related images,it has been widely used in image/video compression,image 3D reconstruction and so on.Graph neural network(GNN)introduces the core idea of convolutional neural network(CNN)into graph signal processing.GNN provides a promising solution for optimizing traditional graph-based co-sal algorithms.The difficulty of co-saliency detection lies in how to extract features with sufficient power among intra-image and inter-image for different scenes,and fully mine and utilize the clues of co-sal.In this dissertation,with the help of the single-frame progressive feature extraction module,using the characteristics of the gradual expansion of the receptive field during the graph neural convolution,a novel graph neural network detection framework are proposed.It mines the saliency clues among intra-and inter-image in a unified way.At the same time,this dissertation introduces the self-attention mechanism into the graph neural network to make full use of the intra-and inter-image cues of the image,wich can adaptively measure the Euclidean distance of the global vertex pair in the graph embedding space,and learn the high-level semantic cues.Aiming at the problems of co-saliency detection,the research content of this dissertation is as follows:(1)Aiming at the problems of traditional graph-based detection methods that rely heavily on artificially designed pattern,and poor robustness in complex scenes,this dissertation introduces GNN into co-saliency detection.The core of the proposed detection framework is to model the detection of the co-salient pattern as the classification of the graph vertices.By extracting single-frame hierarchical salient features and image superpixel segmentation,the detection framework constructs an implicit-edge graph model.The stacking of the multi-layer dynamic graph convolution realizes the hierarchical feature extraction of the multi-graph model,and forms a unified hierarchical feature system with the single-frame hierarchical feature extraction,which greatly improves the expressive ability of the network.(2)Aiming at the difficulty of exploring effective cues among intra-and inter-image,this dissertation introduces the self-attention mechanism into the graph neural network which is suitable for the implicit-edge graph model.The core principle of the module is to measure the Euclidean distance of the vertex pair in the graph embedding space to explore their relationship,and to weight the co-sal vertices.The module realizes the learning and expression of the semantic cues,and enhances the detection ability in complex scenes.(3)The advanced nature of the proposed algorithm is verified through fair and sufficient experiments.The proposed algorithm is compared with many state-of-the-art works based on both traditional and deep learning methods,and both qualitative and quantitative experiments are implemented on different benchmark datasets.Qualitative experiments have proved that the graph neural network detection framework has the advantages of accurate positioning and uniform highlighting of the target area.Quantitative experiments show that the algorithm has achieved the best or very competitive detection effect compared to others.