Image retrieval based on RST-invariant features

In the application of content-based image retrieval, the ideal characteristics should be invariance to geometrical transformations. That is, once the image undergoes geometrical transformations, we expect the features extracted from the image are invariant. Thus, in this paper, rotation, scaling and translation (RST) invariant features for image retrieval are investigated, and a new method is proposed to extract these features. This method performs log-polar transformations on images in order to convert the scaling and rotation transformations to translation transformations, and then utilizes the translation and rotation invariance property of the Burkhardt’s features to extract RST-invariant features. Moreover we take the structural information into account and combine it with the histogram descriptor. By combining these techniques ingeniously, we can retrieve both the RST transformed images and the similar images of the query image. The retrieval performance of the proposed method is illustrated in experiments and its advantages are shown by comparing with other methods.