Robust registration of calcium images by learned contrast synthesis
We describe an automatic method for fast registration of images with very different appearances. The images are jointly segmented into a small number of classes, the segmented images are registered, and the process is repeated. The segmentation calculates feature vectors on superpixels and then it finds a softmax classifier maximizing mutual information between class labels in the two images. For speed, the registration considers a sparse set of rectangular neighborhoods on the interfaces between classes. A triangulation is created with spatial regularization handled by pairwise spring-like terms on the edges. The optimal transformation is found globally using loopy belief propagation. Multiresolution helps to improve speed and robustness. Our main application is registering stained histological slices, which are large and differ both in the local and global appearance. We show that our method has comparable accuracy to standard pixel-based registration, while being faster and more general.