Bag-of-Words Vector Quantization Based Face Identification

This paper investigates the possibility that uses Scale-Invariance Feature Transform (SIFT) feature for face identification. However, it is impossible to employ these SIFT keys,i.e. feature vectors, for identification directly, due to the space incompatible of such SIFT keys. To this end, the Bag-of-words (Bow) vector quantization introduced from scene or text classification is conducted for unifying them. And a novel distance, Cauchy-Schwartz Inequality Distance (CSID), is performed for determining which cluster each keypoint of image belongs to after quantization, in order to compose a histogram vector as our feature. To summarize, this approach solves the problem of space incompatible and avoids a high computation that gives rises to a "dimensional curse". The experiment shows a reasonable result using SVM classifier by ORL database.