Fast and robust cepstral techniques for digital image registration
MetadataShow full item record
The problem of fast and accurate translational registration of images using cepstral techniques is addressed in this thesis. Evaluation of the robustness and improvement in computational speed of cepstral registration, have been the focus of the work. Theoretical and empirical evaluation of the performance of the cepstral technique is presented. These tools are highly data-dependent However, somewhat general performance measures as a function of SNR and signal and noise bandwidths are worked out. The influence of signal and noise bandwidths on the mean reduction in the height of the cepstral peak, variance of the background noise, probability of correct peak detection and variance of peak location are analyzed and empirically justified. An extension of this work shall aid in the comparative assessment of the fast and robust cepstral tools with other area-based algorithms such as cross-correlation and phase-correlation, in the contexts of digital image registration and matching. A unifying treatment of the intensity-based approaches in terms of spectral whitening is also done. Cepstral methods using fast transforms such as the discrete sine, cosine and Hadamard transforms, which bring about a significant and critical reduction in the computation time, are developed. It is also shown how the complex cepstrum can be used to determine the direction of shift for translational correction.