Publication

Abstract : The classification of galaxies is a critical task in astrophysics, providing insights into their formation and evolution. Traditional manual methods are slow and error-prone, motivating the use of deep learning for automation. To address the challenges of automated galaxy classification, various neural network architectures such as Vision Transformer, EfficientNetB0, ResNet18, MobileNetV2, DenseNet121, Deep CNN, LSTM-RNN, and Simple CNN were applied to the galaxy image dataset. MobileNetV2 achieved the best performance with 98.4% accuracy, 87% precision, 67.6% recall, and an F1-score of 73.7%. Regularization techniques were also tested, with Label Smoothing yielding superior results (98.7% accuracy, 96.4% precision, 98.2% recall, F1-score 97.3%). Optimizer comparisons showed that AdamW provided the lowest loss (8.84%) while maintaining high accuracy. These results demonstrate that deep learning, particularly MobileNetV2 with Label Smoothing and AdamW, offers a reliable and efficient solution for galaxy classification.