Layer Layer Relationship Inference Of Monocular Image Based On Dictionary Learning

Layer realationship inference for monocular image is the foundation of image analysis and deep image information extraction.The survey has a wide variety of application,including image understanding(target recognition,pose estimation),robot vision,motion analysis and visual tracking,etc)and three dimensional reconstruction,etc.Thus,Layer realationship inference of monocular image is very important in the survey of computer vision.The work of this paper can be divided into two big problems:(1)partial order relationship inference of image areas.(2)global order relationship inference of image according to the obtained partial order relationship.In this thesis,the dictionary learning method is used to determine the partial order relationship.For learning classification problem,the select of features for classification is important.In this thesis,we use the strength of the edge feature and occlusion edge information while occlusion edge information includes the regional differences of color and texture information around the occlusion edges.In this thesis,lots of edge feature is used to ensure the richness of feature extraction.The sparse classifier based on Fisher index method is introduced to the classification problem to get the exact occlusion edge.First of all,the edge features extracted from the training set are used to train a FDDL(Fisher Discriminative Dictionary Learning)occlusion edge classifier.Later,the obtained edge classifier can be used to classify the edge features extracted from the test set.In this way,occlusion edges of test set can be obtained.With the T junction detection and convexity detected around the occlusion edges,partial order relationship inference of image area can be otained.In this thesis,the theory of Markov Random Field(MRF)is used to solve the problem of global order relationship inference.According to the obtained partial order relationship,the graph model can be formed.The vertices in the graph represent occlusion regions and the edges represent the local depth relationship.The markov random field is used to remove the loop in the graph.In this case,the global order relationship inference of image area can be turned to search of valid path.In this thesis,the NYU V2 depth dataset and the Cornell depth dataset are used as our dataset.Half of them are used as the training set and the remaining the test set.The experiment results show that the proposed algorithm framework has pretty good performance.