Normalization is the SPM term for inter-subject registration - transforming a brain to better match a brain of another subject or template.
There are several possible ways to do this.
The methods split into those that use the structural image for warping, and those that use the EPI.
Using the structural image for warping
Methods that use the structural image for warping rely on having good matching (ProcessingCoregistration) between the structural and functional images, and practically this means you will have to use [wiki:FmBackground EPI undistortion.
Our current preferred option is to use the method that is now standard in SPM5, which is to segement the subject's structural image to get an estimate of the position and probability of gray matter voxels and then warp this gray matter image to match an estimate of the gray matter for the template. In SPM2, this involves a less automated process than SPM5 - following what used to be known as the optimized VBM protocol.
Our previous standard practice was to use [wiki:NormalizeSkullStripped the skull-stipped structural image for normalization].
Using the EPI for warping
Amother method is to use the functional mean image and warp it directly to the SPM EPI template. This works reasonably well, if you use cost function masking to remove the effects of the susceptibility artefact holes in the EPI images; see the [wiki:MaskedEpiNormalization masked EPI to EPI normalization] page.