Diabetic retinopathy (DR) is a frequent eye disorder mostly affecting diabetics. It affects millions of individuals worldwide and is the leading cause of blindness and visual impairment in diabetics. DR occurs when excessive blood sugar levels damage the retina's small blood vessels. In this paper a novel deep learning-based approach for classifying retinal images into three categories: normal retina, good retina, and bad retina. The proposed system utilizes convolutional neural networks (CNNs) for automated feature extraction and classification. The pipeline begins with image preprocessing, including rescaling, train-test splitting, and data augmentation, to enhance model performance and generalization. The preprocessed images are then fed into a CNN model, which extracts features using several convolutional layers, ReLU activation, and pooling layers to minimize spatial dimensionality. The collected features are flattened and processed through fully linked layers, resulting in a SoftMax activation function that produces probabilistic classification results. The accuracy of the suggested method can reach 99.04%, compared to 83.1%, 83%, and 92.1% for conventional models like the Structured learning, High speed detection, and Fuzzy broad learning. In comparison to the existing approaches, the accuracy of the suggested methodology increased by 16.09%, 13.8%, and 3.75%, respectively.