Adaptive Filter Based Image Registration
Benjamin Thomas Henson, Yuriy V. Zakharov · 2015
The work presented in this paper is the development of an image registration system based on adaptive filtering. The adaptive filter used has an additional penalty term to promote sparsity in the estimation of the convolution kernel. From the derived convolution kernel an estimate of the displacement field is made, from which an interpolated image is generated. This estimate can then be refined by iterating over the filtering process using the image generated as the new target image. Stability in the refining iterations is improved by using multiple space filling curves for the adaptive filter scan paths. This not only smoothes the changes in the displacement vectors but the multiple paths add diversity, which improves the evolution of the adaptive filter through more difficult portions of the image content. Due to this greater stability, the forgetting factor for the adaptive filter can be reduced allowing more detail in the displacement to be determined. The resultant system compares favourably with a standard intensity based image registration technique with the commercial Mat lab implementation. A selected set of tests were also performed with the Middlebury dataset [1], which shows the comparative strength and weaknesses of the approach.