Image similarity based on eigen-correspondences

Conventionally eigen-decompositions based on Principal Component Analysis and its variations have been used as a learning tool for capturing the pose and illumination changes in a large set of images, particularly faces. However, if this eigen-decomposition is performed on a single face image based on the row-column covariance statistics, the resulting dominant eigenvectors can be used for checking the statistical-synchronicity between any two images. This comparison can be done by determining the degree of alignment between the dominant eigenvectors which span the row or column spaces in the two images. This eigen-linking process has been found to be robust to several signal processing operations, scaling and noise insertion, despite remaining sufficiently discriminative across perceptually dissimilar images.