Semantic Discriminative Projections For Image Retrieval

With the development of digital imaging technology and the popularity of World Wide Web, Gigabytes of images are generated every day. It is a challenge that manage effectively images visual content. Content Based Image Retrieval (CBIR) is receiving research interest for this purpose. However, there are still many open issues to be solved.Firstly, the visual content such as color, shape, texture, is extracted from an image as feature vectors. The dimensionality of feature space is usually very high. It ranges from tens to hundreds of thousands in most cases. Traditional machine learning approaches fail to learn in such a high-dimensional feature space. This is the well-known curse of dimensionality.Secondly, the low-level image features used in CBIR are often visual characterized, but it doesn't exist the directly connection with high-level semantic concepts, i.e. so-called semantic gap.To bridge low-level visual feature to the high-level semantic is a great challenge in CBIR. We use Laplacian to learn the images semantic subspace in order to achieve more discriminative image representation for CBIR.In our work, both visual similarity and semantic dissimilarity are applied to construct neighborhood graph since not only they contain the descriptive information of the unlabeled images but also the discriminative information of the labeled images is utilized in learning. We introduce a penalty 7 to formulate a constrained optimization problem in the difference form, so that the optimal projection can be found by eigenvalue decomposition. Information of conjunctive graphs is represented by a affinity matrix, and it is much more computationally efficient in time and storage than LPP and LDE. On the other hand, the learnt subspace can preserve both local geometry and relevance information. Previous works often neglect the singularity problem and the optimal dimensionality, but we will determine the optimal dimensionality and avoid the singularity problem simultaneously. Subspace learning has attracted much attention in image retrieval. In this thesis, we present a novel subspace learning approach, referred to "Semantic Discriminative Projection" (SDP), which learns the semantic subspace through integrating the descriptive information and discriminative information. We first construct one graph to characterize the similarity of contented-based features, and another to describe the semantic dissimilarity. Then we formulate the constrained optimization problem with a penalized difference form. Therefore, we can avoid the singularity problem and get the optimal dimensionality while learning a semantic subspace. Furthermore, SDP may be conducted in the original space or in the reproducing kernel Hilbert space into which images are mapped. This gives rise to kernel SDP. Learning with tensor representation is further introduced. To capture the semantic dynamically, SDP can integrate relevance feedback efficiently through incremental learning. We investigate extensive experiments to verify the effectiveness of our approach. Experimental results show that our approach achieves better retrieval performance than state-of-art methods. Finally, we summarize our findings and discussion extensions to the current work.