Accurate classification of variable stars is fundamental to measuring astronomical distances, understanding stellar evolution, and refining models of galactic structure. This work presents a hybrid deep learning framework that improves variable star classification by combining image based light curve analysis with astrophysical features derived from Fourier decomposition and skewness analysis. The proposed approach addresses key challenges such as class imbalance, phase misalignment, and difficulty in distinguishing closely related subtypes.

Minima Phase Standardization is introduced to align phase folded light curves, ensuring consistency across samples. A Fourier best fit model extracts physically meaningful coefficients that capture light curve morphology, while a Variable Star Light Curve Simulator generates synthetic data to augment underrepresented classes, particularly Anomalous Cepheids and Type II Cepheids. This results in a more balanced and representative training dataset.

The hybrid Multiple Input Neural Network achieves an overall classification accuracy of 89.8 percent within ten training epochs, with substantial performance gains for rare classes. For common variable star types, a standalone convolutional neural network reaches 98.1 percent accuracy. These results highlight the importance of integrating domain specific astrophysical knowledge with deep learning to enable reliable, large scale automated classification of variable stars and advance our understanding of the universe.