NEXT II is a software package release obtained from the re-engineering of original NEXT package [Andreon et al. 2001], based on unsupervised neural network employment. It is designed by following the Object Oriented Programming (OOP) paradigm and implemented in C++. The final output consists of catalogues of objects extracted from astronomical images.

First step is the segmentation, a process in which the image is partioned into homogeneous regions. For an astronomical image, the segmentation consists of the recognition of internal objects, by filtering noise and revealing their own features. We decided to apply unsupervised neural models because of their intrinsic and well demonstrated high performance in terms of segmentation in noisy environments.

With these models, the segmentation is reduced to a classification problem for all pixels, as belonging to an object or to the background. Final result is a mask of input image, indicating which pixels belong to an object.

The segmentation phase inside NEXT II is based on two basic phases:

1. The clustering of image pixel set and the construction of the segmentation mask of the input image;

2. The ordering of image pixels based on the luminosity of belonging cluster;

Second step is the object evaluation for all elements coming out from the segmentation process. It is based on the measurement, for each object, of a set of photometric parameters. The input to this step is the image input and related segmentation mask. Its output is a file representing the raw catalogue of objects contained in the input image. In this file, each row identifies an object, while each column represents the measurement result of a specific photometric feature for that object.

The evaluation phase is based into four steps:

1. Identification of objects contained in the image mask, i.e. the result of the segmentation application to the input image;

2. separation of multiple objects: it potentially represents more than one object in a small region. These objects are generated when several objects are enclosed in a very small region, becoming very difficult to be clearly separated during segmentation;

3. refinement of revealed object contours: the segmentation often creates shapes not perfectly matching the original shape of objects inside the image. So far, basic role of this step is to minimize this shape shift, by trying to collimate the centroids and to cover the entire region;

4. extraction of the features and construction of raw catalogue of revealed objects;

Third step is the classification of objects as stars or galaxies. NEXT II makes use of supervised neural networks (MLP). Each revealed object, coming out from evaluation phase, is classified by performing a non-linear separation of input parameter space, between two categories (star/galaxy separation). Only after this procedure it will be possible to analyze objects inside the image.

The input of this phase is the raw catalogue, coming from the evaluation, while its output is the complete catalogue, obtained by adding a column to each object representation, reporting its type (S=star / G=galaxy).

The classification is based on two main steps:

1. selection of features to be used for the classification;

2. classification star/galaxy of each object and final refinement of object catalogue;

 

One of the main targets of NEXT II project was to perform:

• extension of process to multi-band images (simultaneous management of 3-band images);
• improvement of detection correctness;
• generation of more accurate classification (star/galaxy separation);

Fig. 4: NEXT II improvement by processing multi-band images

 

Main requirements achieved on the tool package were:

• configuration;
• modularity;
• flexibility;
• reliability;
• robstness;

Fig. 5: NEXT II configuration and robustness achievement

 

Below (Fig. 6) there is an example of segmentation during test of the package.