Image sequence registration and image estimation techniques for super-resolution reconstruction

Abstract

This dissertation proposes two novel algorithms: the highly accurate sub-pixel image registration and the robust norm for SRR algorithm. The proposed registration assumes the affine motion as the relationship between blocked images (the current frame and the reference frame). It is applicable to not only the standard sequences but also real sequences with complex motion. Therefore, it can be implemented in the previous SRR algorithms. Moreover, it can be implemented in motion estimation algorithm. To realize the implementation of the proposed sub-pixel image registration, the fast algorithm is designed to reduce the computational load for the proposed sub-pixel registration. This dissertation considers the use of a regularized maximum likelihood estimator in the image estimation process due to its high performance and low complexity. This dissertation also studies the effect of norm estimation in SRR algorithm. The L1 or L2 norms with different regularized functions are interested in this work. The novel robust norms (Huber norm, Lorentzian norm and Tukey’s Biweigth norm) are proposed into the model of the SRR framework using the proposed registration. To evaluate the effectiveness of the proposed image registration and the robust norm for SRR algorithm, various noise model and image sequences used in SRR algorithm have been investigated. Five experiments have been carried out to demonstrate the performance of the proposed methods: 1) Experimental on fast affine block-based registration, 2) Experimental on the SRR algorithm using fast affine block-based registration, 3) Experimental on robust estimation technique for SRR, 4) Experimental on the SRR algorithm using robust estimation technique with classical registration and 5) Experimental on robust estimation technique using affine block-based registration for SRR. Experimental results show than the affine block-based registration algorithm clearly gives a higher accuracy than the classical algorithm both objectively and subjectively. By using this proposed registration algorithm, the super-resolution algorithm can be applied on the general sequence such as Foreman and Susie sequence. Moreover, the proposed robust SRR can be effectively applied on the images that are corrupted by various noise models. Experimental results clearly demonstrated that the proposed robust algorithm is applicable on the several noise models such as Noiseless, AWGN, Poisson Noise and Salt&Pepper Noise and Speckle Noise and the proposed algorithm can obviously improve the result both subjectively and objectively.