Research On Semi-supervised Sparse Feature Selection For Image Annotation In Web Space

In recent years, with the rapid development of computer technology, internet technology, multimedia information technology, and extensive use of digital products, images in web space have shown continuously explosive growth. Confronted with these image data in web space, how to effectively retrieve and manage them has become a critical research issue in multimedia understanding and computer vision fields.Automatic image annotation technique, which connects the keywords or related document description to images, has become an important tool to effectively index, retrieve, organize, and manage large-scale images in web space. However, in the face of exponentially increasing image data in web space, automatic image annotation technology confronts with two key problems:one is how to improve the annotation efficiency of large-scale images, and the other is how to improve the annotation accuracy by utilizing the large number of unlabeled images in web space.As an important means, feature selection plays an important role in images annotation in web space. In recent years, semi-supervised sparse feature selection has becomes a hot research topic since it can improve the performance of image annotation. This paper has conducted a research into the existing semi-supervised sparse feature selection methods thoroughly and has proposed several new methods from three aspects, i.e. sparse representation theory, semi-supervised learning methods and multi-view learning. The main research results and contribution of this paper include:(1) Semi-supervised sparse feature selection algorithm based on l2,1/2matrix norm has been proposed.We have deeply studied the l2,p(0<p≤1) matrix norm proposed recently and have proposed a new semi-supervise sparse feature selection algorithm based on l2,1/2matrix norm, which has the best performance when p belongs to (0,1]. l2,1/2matrix norm model not only considers the correlation between different features, but also has the best sparsity. Therefore it makes the selected features more sparse and more discriminative. In addition, it can reduce the computational complexity and improve the efficiency. This paper has proposed a semi-supervised sparse feature selection framework based on l2,1/2matrix norm, namely sparse Feature Selection based on Graph Laplacian (FSLG). An efficient iterative algorithm is designed to solve the objective function. The proposed algorithm FSLG has been applied into image annotation in web space and the results illustrate FSLG can improve the performance and efficiency of image annotation. (2) Semi-supervised sparse feature selection algorithm based on Hessian regularization has been proposed.Among different semi-supervised learning methods, graph Laplacian based semi-supervised learning is the representative work. However, Laplacian regularization is short of extrapolating power and it cannot utilize the geometric structure of unlabeled data well. Contrast to Laplacian regularization, Hessian regularization favors functions whose values vary linearly with respect to geodesic distance and preserves the local manifold structure well, leading to good extrapolating power to boost semi-supervised learning. This paper has applied Hessian regularization into semi-supervised sparse feature selection framework and proposed the Hessian sparse Feature Selection based on L2,1/2-matrix norm (HFSL). An efficient iterative algorithm is designed to solve the objective function of HFSL. We apply HFSL into image annotation in web space and the results demonstrate it is capable of improving the image annotation performance.(3) Two semi-supervised sparse feature selection algorithms based on multi-view learning have been proposed.Most existing semi-supervised feature selection algorithms are developed for single-view data and they often directly concatenate multi-view data into a long vector once these methods confront with multi-view data. This concatenation strategy can’t explore the complementary of different view data efficiently. In addition, this concatenation strategy ignores the physical interpretations of different views. In recent years, multi-view learning has received wide attention and research since it can make full use of the complementary and consensus properties between different view data. In light of the complementary property, this paper presents a multi-view Hessian semi-supervised sparse feature selection framework (MHSFS) based on multi-view learning. In light of the consensus property, this paper presents a semi-supervised sparse feature selection based on l2,1/2-matix norm with shared subspace learning (SFSLS). Two simple yet efficient iterative methods are proposed to solve the objective functions accordingly. MHSFS and SFSLS are applied into image annotation in web space and the results show that the proposed methods significantly outperform the state-of-the-art sparse feature selection methods to improve image annotation performance.