Some domains (remotely sensing, medical imaging, etc) require subpixel level accuracy of image registration. Here you can find the following algorithms:

• mi_reg.py Mutual Information (MI)

Zhang, Boyang, et al. "A mutual information based sub-pixel registration method for image super resolution." Intelligent Information Hiding and Multimedia Signal Processing, 2009. IIH-MSP'09. Fifth International Conference on. IEEE, 2009.

• poc_reg.py Phase-Only Correlation (POC)

Nagashima, Sei, et al. "A subpixel image matching technique using phase-only correlation." Intelligent Signal Processing and Communications, 2006. ISPACS'06. International Symposium on. IEEE, 2006.

Rule of thumb:

• MI is slow, but better at registering multi-modal images (i.e. X-ray vs CT scan vs MRI)
• POC is fast, but might have issues if images have significant differences

Original image

Shifted image (5.3p in X, 6.7p in Y)

from mi_reg import main_mi_reg
from poc_reg import main_poc_reg

dx, dy, match_height = main_mi_reg(im_orig, im_shifted)
print(dx, dy, match_height)
> -5.300313983855842 -6.699973244887778 0.9967009221534583

dx, dy, match_height = main_poc_reg(im_orig, im_shifted)
print(dx, dy, match_height)
> -5.299159785340295 -6.700930722501576 0.9797807076193092