Research And Application Of Several Important Issues In Attribute Learning

As a type of high-level object description, attribute learning has recently attracted more attention in machine learning and pattern recognition domains. Compared with conventional low-level features based on statistical information(e.g., histogram of RGB colors), attribute representation owns the advantage of high flexibility, good interpretability and strong generalization. In literature, many existing attribute-based studies have shown that the use of attributes can simplify data representation, reduce model complexity, and promote learning performances as well as robustness of corresponding models. Despits the success in many practical applications, existing attribute-based learning methods usually suffer from the following limitations: 1) the(correlated) relationship among attributes is not considered; 2) most existing methods cannot deal with the problem of high-dimensional features; 3) the class imbalanced problem cannot be efficiently overcome; and 4) images owning apparent structure characteristics cannot be well described by attributes.This dissertation is a research of several important issues in attribute learning, especially for attribute representation, attribute relationship learning, attribute feature selection, attribute classification model design, and corresponding practical applications in object recognition and neuroimaging analysis. The main contributions can be summarized as follows.(1) As multiple attributes are used to describe the same subject in attribute learning, there exists some relationship among these attributes. To explore such relationship from data, this dissertation proposes an automatic attribute relationship learning(ARL) method to model the relationship among attibutes explicitly. In this method, the attribute relationship is captured by using an inverse covariance matrix that reflects the correlated relation among different attributes in a joint attribute classifier learning model. Furthermore, the learned attribute relationship is incorporated into traditional attibute classifiers that ignore such relationship. Experiments on four benchmark attribute databases show that our methods can not only precisely model the correlated relation among attributes, but also promote attribute classification performance especially when the labeled data are severely limited.(2) Existing attribute learning methods usually adopt low-level features that are of high-dimensional, which is time-consuming and may also bring poor generalization performance. To address this problem, this dissertation proposes a pairwise attribute constraint guided sparse(CGS) feature selection method for attribute learning, where two pairwise constraint-guided regularization terms are used to guide the feature selection process. In addition, we further develop two variants of CGS, i.e., semi-supervised CGS(SCGS) and ensemble CGS(ECGS). Experimental results on thirteen benchmark data sets demonstrate that our methods can effectively not only reduce the feature dimension, but also promote the attribute classification performances in both supervised and semi-supervised learning scenarios.(3) In current attribute learning methods, different object categories usually share the same set of attributes. Therefore, only a small portion of categories have responses in some attributes, which may lead to a class-imbalance problem. To this end, this dissertation presents a two-stage cost-sensitive learning method for attribute classification, which utilizes cost information in both the feature selection stage and the classification stage. In the first stage, we design three cost-sensitive feature selection algorithms, which employ cost information to select the most discriminative feature subset that can minimize the total misclassification cost. In the second stage, cost-sensitive classifiers are used to conduct classification, to avoid the problem that correponding classifiers are dominated by the majority category, e.g., the negative category. Experimental results show that, compared with conventional methods, the proposed method can efficiently reduce the overall misclassification cost and achieve high classification accuracy.(4) Existing attribute learning methods mainly focus on using semantic attributes, visual attributes and discriminative attributes to describe objects, and there is no specific attribute representation for images that owns apparent structure characteristics. To address this problem, this dissertation first proposes an attribute representation method, called structural attributes, and then develops a clustering algorithm to determine the required structural attributes from data. Afterwards, we design a unified framework for structure attribute-based neuroimaging classification. Specifically, there are three stages in this framework, including 1) structural attributes representation for neuroimaging, through which multiple sets of features can be obtained for a brain image; 2) feature selection by using the proposed relationship-induced sparse feature selection algorithm; and 3) an ensemble classification method based on multiple structural attributes for neuroimaging classification. Experimental results on the public database validate the efficacy of our method.