Semantic class discriminant projection for image retrieval with relevance feedback

With the user’s feedback, projections are often used to reduce dimension and enhance class discrimination. The existing projections either use only the global euclidean structure or refer to the local manifold structure. However, global statistics such as variance (ie the method using the global euclidean structure) is difficult to estimate when there are not enough training samples. As for the methods that use the local manifold structure, the class discriminant is limited. In this paper, a Semantic Class Discriminant Projection (SCDP) is proposed for enhancing the performance of content-based image retrieval schemas with relevance feedback. SCDP can take advantage of the local geometry information of labeled and unlabeled images to learn a semantic subspace, and it obtains the most important properties of the subspaces to enhance classification. The experimental results performed on the two benchmark datasets have confirmed the superiority of the proposed method.