Enhancing Fine-Grained Visual Recognition in the Low-Data Regime Through Feature Magnitude Regularization