DIAT-DSCNN-ECA-Net: separable convolutional neural network-based classification of galaxy morphology

There will be an unprecedented increase in the number of galaxies observed as a result of the current and upcoming surveys. Consequently, data-driven approaches have become the main tools for deciphering and evaluating this massive volume of data. Computer vision combined with deep learning has proven most effective for recognizing galaxy morphology but most of the conventional deep learning models are large in terms of parameters due to which computational cost, risk of overfitting increases. In this paper, we proposed a lightweight convolutional neural network (CNN) model using separable convolution which helps to reduce trainable parameters of the model. Further, Efficient Channel Attention (ECA) mechanism is used to focus on important features. ECA focuses on features channel wise without dimensionality reduction which reduces the computational overhead. Performance of proposed model named as “DIAT-DSCNN-ECA-Net” is evaluated on two datasets such as Galaxy Zoo 2, Galaxy Zoo DECaLS, each having seven different types of galaxies, achieved an accuracy of 90.81% and 94.17% respectively at the cost of 1.8 Mega-Byte model size, 0.13 million parameters, 1.04 Floating Point Operations (FLOPs). The outcomes of the experiments demonstrate that the proposed approach can outperform the existing CNN models.