Combining PCA-based datasets without retraining of the basis vector set

A method of combining multiple PCA datasets together with re-projecting the dataset into the new PCA space is presented which does not require preservation of the original datasets from which the PCA descriptors were derived. Practical applications based on face recognition are described where (i) multiple PCA datasets can be combined and (ii) an existing PCA dataset can be augmented with a new set of original data samples. Test results performed on a database of 560 facial regions indicate that this method yields practically identical results with the classical approach of retraining over the original dataset.