Based On Graph Learn And Label Propagation Optimization Model For Image Co-saliency Object Detection

Co-saliency object detection is an important research topic in the field of computer vision and multimedia area.It has been widely used in many computer vision tasks.The aim of cosaliency object detection is to automatically locate the common saliency objects or regions from a group of related images.The main challenging issue of co-saliency object detection is how to use the relevant information between intra-image and inter-images to detect the regions with the same foreground.This thesis study the co-saliency object detection task in machine learning and deep learning.It explores graph construction and learning,multi-graph learning,graph convolutional networks and label propagation technologies to solve the problem of co-saliency object detection.First,this thesis proposes a unified energy optimization model for co-saliency object detection.The proposed model integrates image structural information,foreground background cues and linear features of the images to detect co-saliency object robustly and accurately.The main aspect of the proposed unified energy optimization model is that it can perform cosaliency object detection by exploiting label propagation,linear projection,background foreground prior regularization and related information within and between images simultaneously.An effective algorithm has been developed to find the global optimal solution for the proposed model.Extensive experiments on two widely used benchmark datasets demonstrate that the proposed unified energy optimization model can obtain promising performance.Second,this thesis proposes an adaptive graph learning optimization model for co-saliency target detection.The proposed model integrates background and foreground prior regularization,image structural information and linear integration of multiple image features together and also predicts the relationship between two super-pixel nodes based on image features to obtain an adaptive graph for co-saliency estimation.One main benefit of the proposed adaptive graph learning optimization model is that it conducts co-saliency propagation and prediction across different images while maintains the salient information of each image.Therefore,it ensures the consistency and communication across different images and thus can exploit the correlation information effectively for co-saliency object detection.An effective optimization algorithm has been designed to seek the optimal solution for the proposed optimization model.Experimental results show that the adaptive graph learning energy optimization model has promising performance on several standard benchmark datasets.Third,Graph Convolutional Networks(GCN)have been usually utilized for graph data representation in computer vision and machine learning area.However,existing graph GCNs generally use a single graph which can be not adapted for the data with multiple graphs.This thesis first proposes a Multiple Graph Convolutional Network(MGCN)for multiple graph data representation and learning.MGCN propagates information/knowledge across multiple graphs and obtains a consistent graph node representation and graph learning by integrating the information of multiple graphs information simultaneously.Based on the proposed MGCN,We propose a global-local multi-graph collaborative convolutional neural network model for image co-saliency detection problem.Promising experiments on several datasets demonstrate the effectiveness of the proposed model based co-saliency detection approach.Finally,for the above MGCN,it only considers the fixed graph structure which may has some limitations on the labeling of graph nodes.It is believed that the graph structure can be dynamic changed according to the learned node features.This thesis proposes a multiple graph combination learning and collaborative convolutional neural network model for image cosaliency object detection.The model conducts graph convolutional learning and labeling on both inter-graph and intra-graph cooperatively.Moreover,The model employs a graph learning mechanism to learn both inter-graph and intra-graph adaptively.Experimental results on several benchmark datasets demonstrate the effectiveness of multiple graph combination learning and collaborative convolutional neural network model on co-saliency object detection tasks.