Image retrieval by local evaluation of nonlinear kernel functions around salient points

Feature histograms based on the evaluation of Haar integrals with nonlinear kernel functions were used successfully for the purpose of invariant content based image retrieval. In addition to being invariant to rotation and translation, the features have the advantage of preserving structural information of the image. The work presented here concentrates on the idea of calculating these features by evaluating the kernel functions around a small set of preselected points. These points are called the salient points and represent, together with their neighborhood, the most important visual information in an image. The use of these salient points leads to a better representation of the image. Compared to previous work, experiments show that this method gives better retrieval results without introducing extra computational overhead.