Recognition and Age Prediction with Digital Images of Missing Children

Principal Components Analysis is a dimensionality reduction technique that determines eigenvectors (principal components) and corresponding eigenvalues from the covariance matrix of a data set. The eigenvectors that do not contribute much to scatter are truncated without excessive data loss. The remaining eigenvectors represent a new coordinate system in lower dimensional space, allowing a more compact and efficient representation of the original data. In this project, we age-progress digital face images of children by applying PCA to training images that maps young to aged faces. Once we compute the principal components from the training images, we project an input image onto the principal components to obtain a weight vector. The coefficients of the weight vector represent proportions of each corresponding principal component needed to approximately reconstruct the original input image by a weighted summation. We want the weight vector to be in close proximity to a cluster of projected training images. As a result, the reconstructed image should capture the aged features by resembling a weighted average result. In this report, we explain the mathematical derivation of PCA, and the eigenface approach that applies PCA to image data. We then cover the design and development of our age progression program that applies PCA to grayscale images of whole faces. Next, we discuss the extension of the program to support color images, and a feature-based approach to age progression. Finally, we show results of tests performed to evaluate the performance of the age progression program using various training image sets of young and aged face images.