Graph Based Object Detection And Segmentation

Due to the exponential growth of multimedia information, such as images and videos on a daily basis, effectively and e?ciently processing a huge amount of multimedia data has become highly demanded. Image segmentation is a key technique for processing images and videos. Due to the complexity and diversity of natural or medical images, it is hard to design a universal segmentation algorithm. The dissertation will focus on two issues existing in the ?eld of image segmentation: objects representation and segmentation model design.Firstly, the use of multiple features and contextual information can alleviate the in?uence of complexity and diversity of natural or medical images on segmentation performance.Thanks to the advantage of modeling the underlying connection between data using graph structure, we explore to design effective segmentation models based on graph structure in this dissertation. Hence several models are proposed, including multiple features conditional random ?led(CRF) model for interactive segmentation, graph based semi-supervised learning model for segmentation, adaptive clustering model for segmentation and graph based clusters aggregation model for segmentation.Secondly, object detection and location is an effective way of extracting low-level or mid-level information in images. Algorithms should locate objects and excavate information automatically in many application such as image retrieval, video analysis etc.Object detection is a technique for locating region of interest(ROI) automatically in images. Hence, another research topic in our dissertation will focus on designing effective object detection algorithms for accurate segmentation. In a word, this dissertation mainly investigate image segmentation and saliency object detection methods based on graph structure. The contributions are summarized as following:In the research on interactive segmentation, this dissertation develops methods to utilize multiple features and contextual information contained in user input for CRF.An piecewise and feature-speci?c learning strategy is proposed, in which o?ine training is not needed, few samples are demanded and the learning model is of stronger generalization ability. Moreover, we design an active learning model based iterative segmentation framework for accuracy improvement. In order to design methods for segmenting medical volume data based on interactive segmentation framework, a localized clustering method is introduced for initial object location. The location strategy facilitates the propagation of information such as shape or location by utilizing the temporal and spatial relationships of volume data effectively.In the research on clustering based image segmentation, a model based on adaptive clustering is presented in this dissertation. Compared with the classical models,the proposed method can reduce the computational burden while still maintain satisfactory segmentation accuracy by utilizing image information. Furthermore, the cluster numbers are decided adaptively by histogram analysis.It is observed that object detection can improve the segmentation accuracy and e?ciency. Hence in the research on salient object detection and segmentation, three frameworks are introduced in this dissertation. Firstly, this dissertation proposes a multi-layer graph based non-parametric learning model for object detection and segmentation.Compared with the classical detection method, the proposed framework has higher level of performance for images with complex background, weak boundary or texture by constructing multi-layer graph structure to model contextual information of images and incorporating high-order label consistency constraints into the detection energy function. Secondly, a Bayesian saliency detection framework is presented,in which three new regional saliency measurements are given and initial salient regions are located by region ranking. Hence, the Bayesian detection model performs better in highlighting salient objects and suppressing clutter background. Based on three regional saliency measurements, a salient object segmentation method based on automatic labeling and CRF is proposed, in which automatic labeling improves the accuracy of object locating and multi-feature CRF increases the segmentation accuracy.Thirdly, we propose a diffusion based saliency model for multiple objects segmentation.It is observed in the experiment that multi-scale and multi-layer image information can be extracted from saliency maps. Hence, the role of saliency map should not be limited as locating single objects. Motivated by this observation, a multi-label image segmentation model based on multiple object location is introduced. Different from classic methods in which saliency models are use for locating a single object, the role of saliency is extended for multiple objects location in the proposed method. Finally,the multi-region segmentation problem is solved using a generative multi-label image segmentation model. Furthermore, a segmentation model by constructing informative bipartite graph for salient regions aggregation is presented to compose diverse and multi-scale visual patterns of a natural image, so as to achieve better performance in term of detail preserving.