Normalization Methods of SIFT Vector for Object Recognition

In the clustering of key points, the ratio of inter-class distance and intra-class distance is enlarged using Normalized SIFT vectors. In order to significantly make the ratio larger, a new normalization method is proposed, which is termed the ´normalization method based on seedpoints´. In image preprocessing, for the removal of the object background, a SIFT-based entropy method is given, while the removal of the background, effectively retaining the SIFT key points of the object. Before clustering the key points, SIFT feature vectors are normalized using both Lowe´s normalization method and the normalization method based on seed points. Experimental result of clustering shows that two kinds of normalized SIFT vectors, compared with the non-normalized SIFT vector, make the ratio of inter-class distance and intra-class distance larger. Especially, the SIFT vector normalized by the normalization method based on seed points more significantly enhances the distinctiveness between different classes, and makes the clustering better than Lowe´s normalized SIFT vector.