Linear principal transformation: toward locating features in N-dimensional image space

Projection Functions have been widely used for facial feature extraction and optical/handwritten character recognition due to their simplicity and efficiency. Because these transformations are not one-to-one, they may result in mapping distinct points into one point, and consequently losing detailed information. Here, we solve this problem by defining an N-dimensional space to represent a single image. Then, we propose a one-to-one transformation in this new image space. The proposed method, which we referred to as Linear Principal Transformation (LPT), utilizes Eigen analysis to extract the vector with the highest Eigenvalue. Afterwards, extrema in this vector were analyzed to extract the features of interest. In order to evaluate the proposed method, we performed two sets of experiments on facial feature extraction and optical character recognition in three different data sets. The results show that the proposed algorithm outperforms the observed algorithms in the paper and achieves accuracy from 1.4 % up to 14 %, while it has a comparable time complexity and efficiency.