We study a lightweight hyperparameter-search layer for fine- grained flower classification on Oxford-102 (8,189 images, 102 classes). Our pipeline uses gradient descent for network weights.Still, it uses an Improved Harris Hawks Optimization(iHHO) [1] loop to select training configurations for a Recurrent neural network(RNN) [2] classifier fed by CNN features (feature dimension 1280). Each hawk encodes learn- ing rate, dropout, hidden units, [weight decay], [PCA dimension] within predefined bounds, and candidate settings are evaluated on the vali- dation split using a fitness function that combines validation loss with regularization terms (gap and complexity). Under the same data prepro- cessing and backbone, iHHO-RNN improves test accuracy from 65.21% (CNN baseline) and 67.45% (CNN+RNN) to 75.91%, while also yielding a higher F1-score (55.21%). The results suggest that metaheuristic se- lection of training configurations can reduce manual tuning and improve the consistency of convergence on fine-grained datasets.