Our general purpose is twofold:
In a preliminary study
(Paturel et al. 1996)
our target was limited to
identification of galaxies already known in the LEDA database and to
extraction of some astrophysical parameters using our own digitization of the Palomar
Sky Survey. The resolution of the digitization was too poor
(6
)
to allow recognition of new galaxies.
We developed source extraction and automatic cross-identification
algorithms.
In a second study (Vauglin et al. 1998) the same source extraction algorithms were used but we applied an automatic galaxy recognition based on Discriminant Analysis method. The I-band CCD images were obtained with the 1-meter ESO telescope for the Deep Near Infrared Survey (DENIS). Because of the high dynamic of the CCD receiver the separation between stars and galaxies is relatively easy. Stars have a very high central intensity and a small surface area while galaxies do not.
In the present study we are aiming at the most difficult task of recognizing new
galaxies from the digitization of photographic plates (POSS1 and UK Schmidt plates).
Because of plate properties the center of a source (star or galaxy) is generally saturated and
only a few pieces of information can be derived from optical density of
pixels. Besides, the material is made of very inhomogeneous plates taken from very different
regions. It is simply not possible to imagine that the same method will work
in all conditions. This major conclusion of our preliminary analysis will force
us to imagine a method adapted to each individual plate.
The different methods of automatic recognition can presently be classified in 5 classes as follows:
The main characteristic of these methods is that they require a proper choice
of parameters describing each object. This choice is not obvious. It is guided
by the results obtained on a training sample for which each object has been
classified by an expert. This is the main difficulty in the present application.
The material (see Sect. 2) is so inhomogeneous
that it would be necessary to build a training
sample for each plate. This would mean classifying about
objects by eye.
So, we worked in another way.
In Sect. 3 we show that the diagram of the dispersion of pixel optical densities (i.e. standard error of
the pixel intensities) versus the inverse of the surface area of a given object performs this Star/Galaxy
separation well. Thus, we plotted these diagrams
for each plate and made the separation between stars and galaxies by adopting a
frontier function in an interactive manner. This constitutes the first step leading us to
a preliminary catalogue of 4.3 million galaxy candidates and 47.4 million star
candidates.
In Sect. 4, we built a large, general training sample of
stars
and
galaxies by cross-identifying our preliminary catalog
with well established star and galaxy catalogs. This
training sample was used to setup a neural network allowing us to filter the
star/galaxy candidates. After this filtering step we got an all-sky catalogue of
3.2 million galaxy candidates.
Finally, in Sect. 5 we made the internal cross-identification (what we
call the Auto-crossidentification) for galaxies seen several times
on different plates. Then, we made the cross-identification with LEDA galaxies
and cleaned the catalogue in order to remove contamination by known extended objects
(Planetary Nebulae, Globular clusters, Open clusters, Bright Nebulae,
Bright Stars) and by very faint galactic stars. This led
to the final catalogue of
galaxy candidates.
Copyright The European Southern Observatory (ESO)