Table 2.

Computational efficiency including the image representation calculation and registration time consumption.

Registration methods	Intensity-based	MACMI	MSERg	SIFT
Representation computation	NA	5 min	10 min	7 sec
Registration	0.5 min	3.5 min	160 min for learning; 2.5 min for testing	1.5 min

Although MSERg has a higher computational cost during the learning step, this burden can be greatly reduced via parallel computing because all the single-scale representation registrations are essentially independent of each other, with no dependencies.