Study On Layer Relationship Decision Algorithm Of Monocular Image Based On Hybird Features

With the development of computer vision applications, the requirements of image analysis and understanding are also increasing. Obtaining the depth information and layer relationship from a monocular image is an important issue for image analysis, and it is also the premise of completing higher level computer vision tasks. Therefore it’s of great significance to analyse the depth information of image exactly.The description of layer relationship between the image regions is called2.1D sketch or image layering in computer vision. Computing the2.1D sketch is an important step to resolve higher level tasks of computer vision, such as foreground-background separation, depth estimation,2.5D sketch, image and video encoding, object recognition, motion tracking and analysis.Edge intensity, region color, texture and gradient information are important clues to resolve2.1D problem, although each clue can infer parts of depth information, there still exists several limitations. In this thesis, we survey and analyse many state-of-the-arts approaches and present a novel algorithm framework of for inferring the2.1D sketch. Firstly, image is segmented into several small regions by super-pixel segmentation technology in which the internal structure is consistent. Secondly, an occlusion edge classifier is trained by integrating multiple local features, and we use this classifier to remove noise edges to obtain a number of meaningful image regions.In addition,T-junction and convexity features are both employed to infer the figure-ground relationship, and color and texture features are utilized together to infer the same-level relationship. Considering the existence of conflicts between local features, we constrain the global consistency by a cost function. Finally we obtain the2.1D sketch.In this thesis,D-order data set is utilized for evalution, in which300pictures are selected as a training set, and the remaining787pictures are used for testing. The experimental results show that the proposed algorithm framework has an excellent performance.