Hi Seyyed,
To my understanding, castor-norm is based on a direct normalization, unlike the paper from Pepin et al. which is based on a component-based normalization.
The principle of a direct normalization is simple : the normalization factor is given by the ratio between the expected projection and the measured projection.
The expected projection is the forward projection of the normalization phantom. This is why you have to provide the true image of the normalization phantom to castor-norm, rather than the reconstructed (hence, approximative) image of the normalization phantom.
Experimentally, it is better to use an annular phantom to neglect attenuation and scattered coincidences. If not, you have first to correct the measured data for attenuation and scattered coincidences. You shall also remove the random coincidences.
The downside of direct normalization is that you need a large number of coincidences per line-of-response. Direct normalization is very sensitive to the noise in the measured data.
Best,