Comparative Study of Deep Learning Model for Transient Image Classification

In this research, we explore the efficacy of various deep learning models, including DenseNet, Xception, VGG19, VGG16, ResNet50, MobileNet, MobileNetV2, InceptionV3, and ResNet101, in classifying images within Transient Image datasets, which are pivotal for discerning celestial changes. The datasets, comprising real and bogus images, were converted from FITS to JPG format to circumvent potential modeling inconveniences. Addressing the data imbalance issue, where real images are outnumbered by bogus ones at a 1:7 ratio, we employed oversampling on the real images, utilizing image rotation for the task. Our experiments yielded accuracies up to 98% across several models. However, scrutiny of the performance graphs indicated potential overfitting, suggesting that the training dataset may lack requisite complexity. Notably, while ResNet50, ResNet101, and VGG16 models exhibited commendable performance graphs, their accuracy was not paramount, contrasting with models that demonstrated high accuracy but suboptimal performance graphs due to overfitting. This underscores the intricate interplay between model structure and outcome, suggesting that overfitting may also stem from the inherent complexity of the models themselves. This study illuminates the nuanced challenges and considerations in employing deep learning for astronomical image classification.