Block-Coordinate Gauss-Newton/regression Method for Image Registration with Efficient Outlier Detection

In this paper, the block-coordinate Gauss-Newton/regression method is proposed to jointly optimize the spatial registration and the intensity compensation. Here, the intensity compensation is conducted constructing a polynomial regression model, which enables the detection of occluded regions as outliers. Based on the block-coordinate method, we separate the parameter update into two steps for registration and compensation, respectively. Hence, we can perform a joint optimization with low computational complexities, and can apply an appropriate scaling technique to the parameters to be updated for a stable and fast convergence of the algorithm. Excluding outliers, we can successfully align images compensating the intensity differences.