Diabetic retinopathy is the most severe complication of diabetes that recently became a major cause of vision impairment and blindness worldwide. Early detection along with proper classification of DR can avoid severe complications. However, the diagnosis made by ophthalmologists manually is cumbersome and observer-dependent. The paper presents an automated deep learning approach towards the detection and classification of diabetic retinopathy from real-time retinal pictures. The method uses EfficientNet B0 for feature extraction and DeepNet V3 for DR severity classification. The approach starts with picture segmentation, which divides useful regions in retinas using U-Net architecture. Then, it uses EfficientNet B0 to extract discriminative features from the segmented images that is classified into five grades of severity ranging from “No DR” to “Proliferative DR” with DeepNet V3. This detection and classification technique proposed here is very sensitive for DR, hence making it a very crucial tool for clinical decision-making purposes. Early diagnostic help to healthcare providers is the aim of the method so that the timely intervention could avoid vision loss to the patient.