One of the most challenging problems for image retrieval applications is to find the optimal mapping between high-level semantic concept and low-level features. Traditional approaches often assume that images with same semantic label share strong visual similarities and should be clustered together to facilitate modeling and classification. Our research indicates this assumption is inappropriate in many cases. Instead we model the images as lying on nonlinear image subspaces embedded in the high-dimensional feature space and find that multiple subspaces may correspond to one semantic concept. By intelligently utilizing the similarity and dissimilarity information in semantic and geometric (image) domains, we propose an optimal semantic subspace projection (SSP) that captures the most important properties of the subspaces with respect to classification. Theoretical analysis proves that the well-known linear discriminant analysis (LDA) could be formulated as a special case of SSP. To capture the semantic concept dynamically, SSP can integrate relevance feedback efficiently through incremental learning. Kernel SSP is further proposed to handle nonlinearly separable data. Extensive experiments have been designed and conducted to compare our proposed method to the state-of-the-art techniques such as LDA, locality preservation projection (LPP), local linear embedding (LLE), local discriminant embedding (LDE) and their variants. The results show the superior performance of SSP.