Mutual information registration based on pixel intensity has been widely used in recent years. However

its application to subimages with small samples is questionable

because many localmaximums may happen or the global maximum may be away from the actualmaximum value which causes unnecessary registration error. A new approach

mutual information registration based on feature label(MIF)

is proposed to solve such problem. This method first uses image’s intensity and gradient features to train the selforganized mapping(SOM) neural network

and then builds up the feature classifier for each modal image. Using such a classifier

images are project into a feature space with decreased dimensions. Finallymutual information is evaluated in the feature space tomatch images. Our results demonstrate that this method increases the success rate of the subimage registration

and is optimal for thewhole images(eitherwith or without noise) elastic registration.