A Robust Machine Learning Algorithm for Cosmic Galaxy Images Classification Using Neutrosophic Score Features.

Abstract

The development of galaxy images classification automated schemes is necessary to
identify, classify, and study the evolution and formation of galaxies in our universe as it is one of the
main challenges faced by astronomers today. Scientists can also build a deeper understanding of
galaxies evolution and formation by classifying them into various classes. This paper proposed a
robust novel hybrid automated intelligent algorithm based on neutrosophic techniques (NTs) and
machine learning techniques for classifying the galaxy morphological astronomical images into
various types of galaxies images (Hubble types) based on its features into three main classes;
Elliptical, Spiral and Irregular. A nine classifiers performance was assessed based on the machine
learning (ML) techniques by using a combination of a sets of morphic features (MFs); obtained from
image analysis and principal component analysis (PCA) features. The results indicated that; the
classifier which called, multilayer perceptron (MLP) gives the better results for the features set
consisting of nine MFs and 24 PCs features among all tested cases; Mean squared error (MSE) = 0.0021;
Normalized mean squared error (NMSE)= 0.0371; Correlation coefficient (r) = 0.9889, and the Error =
0.7751 with an accuracy 99.2249 %. Then, to improve the system efficiency; the neutrosophic
techniques were applied in combination with the classifier that gave the best results in the previous
step on the same extracted features to get a three robust component namely; membership,
indeterminacy and non-membership components to fed to the neural network. The results showed
that; the combination between the NTs and MLP classifier for (MFs with 4PCs) gives the best results;
MSE = 0.0001; NMSE = 0.0009; r = 0.9997, and Error = 0.4212 with an accuracy about 99.5788 % in total
for all chosen sets of features. The results showed the high performance of the proposed method
comparing with other methods. The experimental results are performed based on a sample from
(EFIGI) catalog