Efficient eye location using the Accuracy-Weighted Principal Component Analysis

Automatic facial feature location is an important problem in the field of computer vision and automatic face recognition. In this paper, the algorithm of eye location with the Accuracy-Weighted Principal Component Analysis (AWPCA) is proposed grounded on the idea of machine learning. Firstly, the appropriate eigenvectors of the covariance matrix of the set of eyes images are selected by comparing the value of the classification accuracy. Secondly, the threshold for the classifier is determined by using the selected eigenvectors and the accuracy. Lastly, the unknown image region is projected into the selected eigenvectors having the largest accuracy, and the absolute values of the projection coefficients and the corresponding accuracy can be expressed as the total sum of products, which is compared with the threshold to determine whether the unknown region contains the human eye. The algorithm is called the AWPCA because the projection coefficient is multiplied by the accuracy. The performance of our automatic eye location technique is subsequently validated by using the CAS-PEAL database. The experiment results show that the AWPCA algorithm may locate eye more effectively than the original PCA algorithm based on the reconstruction error, especially for the face images with glasses.