An efficient algorithm to compute eigenimages in PCA-based vision systems

In traditional PCA-based vision systems, it is assumed that the object can be easily segmented from the environment. This is only true in simple scenes. One method to get around the segmentation problem is to apply multi-scale methods such as the pyramid method or the scale space method. In multi-scale methods, the computation of eigenimages in different scales is computationally intensive. Hence it poses a main problem concerning its usage. In this paper, an efficient method to compute eigenimages in different scales is presented. This method is exactly true only when the similarity condition holds. In general, it trades accuracy for efficiency. A theoretical error analysis is given for the general situation. Thorough experiments are conducted to test the proposed method. It is found that this algorithm indeed gives good representations in different scales.