Graph Network Representation And Semi-supervised Learning Model And Its Application In Object Tracking

Graph represenentation and learning is an important research task in computer visual and has been extensively studied in recent years.Using it,one can gain a deeper understanding of the structural data information.Traditional machine learning methods often use the heuristic methods like degree statistics or heuristics of kernel functions to extract structural information from graphs.With the development of deep learning and non-linear dimensionality redudction methods,more and more methods have been proposed to encode graph structures into low-dimensional embedding space with self-learning.These methods can make better use of the graph structure and explore structural data information more deeply.In many real applications,due to the high cost of manually labeling data,the number of labeled data is much smaller than that of unlabeled data.In these tasks,traditional supervised learning may be less-learned due to lacking of samples.At the same time,the labeled data and the unlabeled data are both independent identically distributed sampling from the entity,so the data information contained in the unlabeled data is also useful for the training process.To deal with this problem,semi-supervised learning has been studied which aims to use unlabeled data to improve the performance of unsupervised learning.In particular,graph-based semi-supervised learning algorithms is widely used,such as handwritten digit classification,medical image segmentation,image retrieval,etc.Comparing with other semi-supervised learning methods,graph-based semi-supervised learning methods can better utilize data distribution to reveal unlabeled data.Recently,graph convolutional neural networks have been widely studied for graph-structured data representation and learning.In this thesis,we present Graph Diffusion-Embedding networks(GDENs),a new model for graph-structured data representation and learning.GDENs are motivated by the development of graph based feature diffusion.GDENs integrate both feature diffusion and graph node(low-dimensional)embedding simultaneously into a unified network by employing a novel diffusion-embedding model.GDENs have two main advantages.First,the equilibrium representation of the diffusion-embedding operation in GDENs can be obtained via a simple closed-form solution,which thus guarantees the compactivity and computational efficiency of GDENs.Second,the proposed GDENs can be naturally extended to address the data with multiple graph structures.Experiments on various semi-supervised learning tasks on several benchmark datasets demonstrate that the proposed GDENs significantly outperform traditional graph convolutional networks.This thesis also applied the proposed semi-supervised learning model to the visual object tracking.Specicially,considering the difficulty of eliminating the influence of background information in visual object tracking,this thesis proposed a weighted patch method for object representation and tracking.Firstly,the target bounding box is divided into non-overlapping patches,and then a weight value is assigned to each patch by exploiting an effect graph node ranking model.The patches with large weights are more likely to belong to the target object,and the patches with small weights are more likely be belong to the background.Finally,the obtained weighted patch descriptor is combined into a structured SVM tracking framework to achieve object tracking.This thesis propose three rankings methods for patch weight calculation.First,this paper adopt a flexible graph manifold ranking for patch weight computation which explores both unary feature and structure relationship in a unified manner and thus performs more discriminatively than existing models which generally only explore structure relationship in patch representation.Then,learn an adaptive and robust graph to better capture the intrinsic relationship among patches and thus can help to obtain a more robust patch representation.Second,on the basis of the previous method,the adaptive graph learning is further considered and combined into the graph ranking based patch weigh calculation.In this process,they can promote each other and thus can boost their respective performance to obtain a more robust weighted patch representation.Third,this paper propose a unified temporal coherence and graph optimized ranking model for weighted patch representation in visual tracking problem.The proposed model simultaneously explores the unary feature,structure relationship and temporal coherence of patches in a unified manner and thus performs more robustly and discriminatively than existing models which generally only explore structure relationship for patch ranking.Object tracking experiments on several benchmark datasets verified the robustness and effectiveness of the proposedtracking algorithms.