Object Segmentation Model Based On Supervised Hierarcical Dirichlet Process

With the development of computer vision technology, category-level ob-ject segmentation, also known as object segmentation, becomes one of the hot topics of image analysis and computer vision field. It brings vigorous pro-gresses in several industrial fields such as web searching, intelligent appliances,intelligent transportation systems, autonomous vehicles and etc. In this disser-tation, based on the Bayes' theorem and random field models, we propose a novel supervised non-parametric model and the related methods to solve the object segmentation problem as follows.1. A new non-parametric discriminative model relying on hierarchical Dirich-let process (HDP) is proposed for object segmentation.Firstly, considering the benefits of HDP and hidden conditional random field (HCRF), we model the correlation between the low level features of images and semantic categories by using HCRF, of which the hid-den variables are then used in a low order potentials to formulate spa-tial dependencies of samples, and we model the correlations between dif-ferent input data (such as images) with HDP. Secondly, we combine the HDP with HCRF through shared hidden variables, and propose so called HDP-HCRF model as a novel discriminative non-parametric random field model to segment objects from a single or a set of images.2. Methods of maximal conditional likelihood estimation and inference are proposed for supervised non-parametric model with hidden variables, such as HDP-HCRF.Since HDP-HCRF cannot be directly trained by using EM and CD algo-rithm, we propose an online conditional MCEM algorithm to perform the above MCL estimation according to theory of MCEM. Comparing with other algorithms, the CMCEM algorithm is effective and flexible because the couplings between the components in HDP-HCRF model are weak.Since there are infinite statuses of hidden variables in HDP, the number of relative parameters also is infinite. So we try to approximate the ob-ject function by using pseudo likelihood and tree based expression of the structure of MRF to reduce calculations. To avoid over-fitting, we also introduce sparsity into HDP-HCRF by using L1 regularization.3. A high-order potential is proposed to improve HDP-HCRF model and a modified BMRM algorithm is proposed for training.Since there are often many instances of more than one kind objects in one image, we propose a novel high-order potential with respect to hidden variables to model the correlations of different instances in the same cate-gory. To avoid calculating the partition functions, we transfer the problem into its max posterior margin form which is solved by using BMRM al-gorithm.Experiments on several widely used datasets demonstrate superior perfor-mance of our proposed HPD-HCRF models and related algorithms over state-of-art methods for object segmentation with super-pixel features.