As per the studies, 35–65% of people with diabetes are affected by diabetic retinopathy. Diabetic Retinopathy is one of the primary reasons that affects the eyes and the main cause that leads to visual impairment. One of the significant problems with diabetic retinopathy is that the patients do not show any symptoms at the early stages. Patients may experience blurred or fluctuating vision in the early stages. If not treated early, it can damage the vision. Therefore, this research’s main target is to focus on developing automated techniques for detecting abnormalities in the patient’s eyes. Three deep learning-based convolutional neural network models are trained with different activation functions for their ability to capture complex patterns in medical images, essential for accurate diabetic retinopathy detection. We have demonstrated the comparison of our proposed approach with state-of-the-art ensemble algorithms on the following benchmark datasets, including DIARETDB1, APTOS, and Liverpool. We have applied a proposed ensemble method to improve the accuracy and lower the training costs to improve overall performance. Our proposed ensemble method has improved the classification accuracy by 8.13%, raising it from a baseline accuracy of 91.00%. The proposed method gives better results than other ensemble methods, with an accuracy of 99.13%, sensitivity of 91.39%, and specificity of 95.79%.