We present an application of the fast Independent Component Analysis FastICA method to the COBE -DMR 4yr data .
Although the signal-to-noise ratio in the COBE -DMR data is typically \sim 1 , the approach is able to extract the CMB signal with high confidence when working at high galactic latitudes .
However , the foreground emission components have too low a S / N ratio to be reconstructed by this method ( moreover , the number of components which can be reconstructed is directly limited by the number of input channels ) .

The reconstructed CMB map shows the expected frequency scaling of the CMB .
We fit the resulting CMB component for the rms quadrupole normalisation Q _ { rms - PS } and primordial spectral index n and find results in excellent agreement with those derived from the minimum-noise combination of the 90 and 53 GHz DMR channels without galactic emission correction .

We extend the analysis by including additional channels ( priors ) such as the Haslam map of radio emission at 408 MHz and the DIRBE 140 \mu m map of galactic infra-red emission .
Subsequently , the FastICA algorithm is able to both detect galactic foreground emission and separate it from the dominant CMB signal .
Fitting the resulting CMB component for Q _ { rms - PS } and n we find good agreement with the results from Górski et al .
( 1996 ) in which the galactic emission has been taken into account by subtracting that part of the DMR signal observed to be correlated with these galactic template maps . FastICA is therefore able to extract foreground emission from the DMR data and recover a “ clean ” CMB component .

We further investigate the ability of FastICA to evaluate the extent of foreground contamination in the COBE -DMR data .
We include an all-sky H \alpha survey ( Dickinson , Davies & Davis 2003 ) to determine a reliable free-free template of the diffuse interstellar medium which we use in conjunction with the previously described synchrotron and dust templates .
The derived frequency scalings of the recovered foregrounds is consistent with previous correlation studies ( Banday et al .
2003 ) .
In particular we find that , after subtraction of the thermal dust emission predicted by the Finkbeiner , Davis & Schlegel ( 1999 ) model 7 , this component is the dominant foreground emission at 31.5 GHz .
This indicates the presence of an anomalous dust correlated component which is well fitted by a power law spectral shape \nu ^ { - \beta } with \beta \sim 2.5 in agreement with Banday et al .
( 2003 ) .