Fast object-based image registration using principal component analysis for super-resolution imaging

In this paper, an object-based image registration method with real-time performance and no constraints on the three registration parameters (i.e., translation, rotation, and scaling) involved is proposed. The coordinate values of the translation parameters can be quickly estimated based on the center of mass of the binary mask which corresponds to the segmented region of interest. For computing the values of rotation and scaling parameters, the principal component analysis (PCA) is conducted on a 2 × 2 symmetric covariance matrix, which is established from the same binary mask. The formulas derived from the eigenvalues and eigenvectors of the covariance matrix provide accurate estimation of the amount of rotation and scaling. The proposed image registration method is further compared with a frequency-domain image registration approach in terms of their performance achieved in super-resolution imaging application using both real and synthetic video sequences. Experimental results clearly show that the proposed method achieves superior performance on the aspects of reconstructed high-resolution images (due to its accurate registration) and its real-time delivery (due to its low computational complexity).