In this paper we address a challenge of the problem of the dimensionality curse and the semantic gap reduction for content based image retrieval in large and heterogeneous databases. The strength of our idea resides in building an effective multidimensional indexing method based on kernel principal component analysis (KPCA) which supports efficiently similarity search of the heterogeneous vectors (color, texture, shape) and maps data vectors on a low feature space that is partitioned into regions. An efficient approach to approximate feature space regions is proposed with the corresponding upper and lower distance bounds. Finally, relevance feedback mechanism is exploited to create a flexible retrieval metric in order to reduce the semantic gap between the user need and the data representation. Experimental evaluations show that the use of region approximation approach with relevance feedback can significantly improve both the quality and the CPU time of the results