Accuracy improved image registration based on pre-estimation and compensation

• Propose to improve the registration accuracy by pre-estimation and compensation.
• Implement and verify the idea in the rotation-translation (RT) model.
• Evaluate the implementation using different typical images.
• Disscuss the limitations due to the Fourier–Mellin transform.

In this paper, we propose to improve the registration accuracy by pre-estimation and compensation. The idea is motivated by the observation of some registration algorithms that, for a given algorithm, the accuracies of the translation-only model are much higher than those of other complex models. Therefore, it seems that, if pre-estimation and compensation are performed and the residual model is close to translation only, the following estimation could achieve improved accuracy. To verify the idea, we implement two algorithms in the rotation-translation (RT) model. We use the Fourier–Mellin transform to isolate and convert the rotation into translation, then apply the classical Lucas–Kanade algorithm to obtain the high accuracy rotation estimation. In the following, the one takes account into the incomplete rotation compensation, and use the Keren algorithm for the residual model; the other assumes the rotation compensation is complete, and uses the second Lucas–Kanade algorithm. Finally, we perform simulations using typical test images, and the results confirm the accuracy improvement.