The early detection of glaucoma is vital to preventing irreversible blindness, which affects millions worldwide. This study presents a novel deep learning-based approach for glaucoma detection using the InceptionV3 model, focusing on multi-scale feature extraction from retinal fundus images. A comprehensive evaluation was conducted using 7,815 fundus images (5,002 normal and 2,813 glaucoma) from six diverse public datasets: RIM-ONE, ACRIMA, DRISHTI-GS, ORIGA, G1020, and LAG, chosen for their variations in demographics, imaging conditions, and disease severity. To enhance model performance and generalization, we incorporated advanced techniques such as data augmentation, class weighting, learning rate scheduling, and stratified K-fold cross-validation. The proposed model achieved outstanding results, including 100% accuracy, precision, recall, and F1-score on DRISHTI-GS, LAG, RIM-ONE, and ACRIMA datasets, and maintained high accuracy on ORIGA (99%) and G1020 (94%), outperforming state-of-the-art methods. This work demonstrates the potential of InceptionV3 for reliable glaucoma detection across diverse imaging scenarios, offering a significant step toward early intervention and improved clinical outcomes for patients at risk of vision impairment. © 2024 IEEE.