Diabetic retinopathy (DR) is a retinal condition that occurs due to diabetes mellitus and might lead to blindness. Early identification and treatment are crucial to slow down or prevent vision loss and degeneration. However, categorizing DR into several levels of severity remains a challenging problem due to the complexity of the disease. The Diabetic Retinopathy Grading System divides retinal pictures into five severity categories: No DR, Mild Non-Proliferative Diabetic Retinopathy (NPDR), Moderate NPDR, Severe NPDR, and Proliferative Diabetic Retinopathy. In this study, three deep learning models, namely ResNet50, Densenet201, and InceptionV3, were utilized for the classification of the APTOS 2019 diabetic retinopathy image dataset. For the individual experiments of the models, transfer learning with finetuning and layer freezing was applied. Additionally, a decision-level fusion idea using soft voting was implemented across the three pre-trained models. The maximum accuracy achieved for the classification of the original imbalanced dataset was 85% with the fusion idea. To further improve the classification performance, a balancing technique based on oversampling with augmentation operations was applied to the original APTOS 2019 dataset. The proposed approach, which involves the idea of soft voting-based fusion across models along with data balancing, improved the classification performance and achieved an accuracy of 90%.