Automatic Image Alignment Using Principal Component Analysis

Abstract

We present an automatic technique for image alignment using a principal component analysis (PCA) that broadly consists of two steps. The first step is the segmentation of the region of interest by thresholding. In the second step, PCA is applied on nonzero pixels in the binary image to determine the object rotation about the mean of the object pixels. Existing PCA-based techniques align the data in their principal spread; however, they have a critical problem of 180 degrees rotation in their principal axes. This paper provides an automatic solution to address this problem. The algorithm is based on the assignment rules inferred from the eigenvectors given by PCA. We applied the proposed algorithm to different datasets, including handwritten images of digits, a rotated fingerprint image dataset, and a dataset of magnetic resonance brain images, and confirmed that the proposed method aligns the data efficiently and accurately. In addition to alignment, the algorithm proposes two standard orientations for automatically assessing the true side (upside) of an object.