Application of Image SIFT Features to the Context of CBIR

This paper is mainly concerned with the application of a kind of distinctive local invariant feature i.e. Lowe´s SIFT feature for the purpose of CBIR, instead of the usually used global feature and local statistical feature based on image segmentation. In our CBIR system, the visual contents of the query image and the database images are extracted and described by the 128-dimensional SIFT feature vectors. The KD-tree with the best bin first(BBF), an approximate nearest neighbors(ANN) search algorithm, is used to index and match those SIFT features. As our contribution, a modified voting scheme called nearest neighbor distance ratio scoring (NNDRS) is put forward to calculate the aggregate scores of the corresponding candidate images in the database respectively. By sorting the database images according to their aggregate scores in descending order, the top few similar images are shown to users as the retrieval results. Additionally, RANSAC can be adopted as a geometry verification method to re-check the results and remove the false matches. Experiments show that our approach can obtain high recall and high precision in the context of CBIR on the famous image databases ZuBud.