Diabetic retinopathy is a serious complication of diabetes mellitus that can lead to vision loss if not diagnosed early. The aim of this study was to identify diabetic retinopathy using a robust model. The methodology employed was based on six phases: dataset acquisition, preprocessing (CLAHE, standardization and grayscale), feature extraction (SIFT, ORB, HOG), machine learning models (RF, KNN, SVM), deep learning models (BOTNet, CoAtNet, ViT, Swin Transformer, MobileViT and PiT), a hybrid approach and model evaluation (accuracy, precision, sensitivity and F1-Score). The best results demonstrate that the PiT model optimized with the hyperparameters (batch_size: 32, epochs: 15, learning_rate: 0.001 and optimize: SGD) obtained the best performance with an accuracy of 98.70%, precision of 99.11%, recall of 98.23%, and F1-score of 98.27%. In conclusion, the robust deep-learning-based model with hyperparameter adjustment significantly improved the classification of diabetic retinopathy.