Automated recognition of white blood cells (WBCs) is essential for rapid and reliable hematological diagnosis. This study presents a deep learning approach using the EfficientNetV2L architecture with a customized classification head for multiclass WBC classification. The model was trained in two stages: an initial phase where the classifier head was optimized with the backbone frozen, followed by fine-tuning of the deeper layers. To improve generalization, dropout, L2 regularization, and label smoothing were incorporated. The proposed framework achieved an overall accuracy of 99%, with class-wise F1-scores ranging between 0.95 and 0.99. Confusion matrix analysis confirmed minimal misclassification, and the ROC curves yielded near-perfect AUC values across all five classes. These findings demonstrate that the proposed EfficientNetV2L-based model is highly effective for robust and reliable WBC classification, indicating its potential application in clinical decision support systems.