ICFSIFT: Improving Collection-Specific CBIR with ICF-Based Local Features

We present a new adaptive local feature, ICFSIFT, which utilises SIFT keypoints and Independent Component Analysis. The ICFSIFT feature combines the keypoint detection, and scale and orientation invariance, of SIFT with the collection-specific adaptive properties of Independent Component Filter (ICF) features. We evaluate the performance of this feature for image retrieval on two standard texture collections, comparing with SIFT features and previously published global ICF features. On both collections the ICFSIFT features perform best. We also show that combining these ICFSIFT features with the ICF-based global features further improves CBIR performance.