Diabetic retinopathy (DR) is a diabetes complication that causes vision loss and blindness due to retinal damage. It’s a global health problem. DR grading is hindered by challenges like subtle visual differences, small lesions, and imbalanced data distribution. The struggle to identify these nuances, particularly with imbalanced data, poses a risk of compromising the precision of DR grading. This research proposes a DR grading solution with the Convolutional neural network (CNN) with enhanced attention mechanisms called DRNet to address the above problems. This model uses a transfer learning method and maps fundus images to labels across a 5-grade scale from No DR to Proliferative DR. The methodology applied for DRNet follows data processing, data augmentation, feature extraction and the attention mechanism at last classification. First process involves data processing which standardizes the dataset images and applies augmentation techniques like resizing, affine transformation, color jitter and horizontal flip which are applied to make the model robust. Features are extracted using a fine-tuned pre-trained deep learning model called ResNet. After that, in the subsequent stages, these features are passed to the Attention block which focuses on the regions of interest in the images and then to the fully connected neural network layer which is used for classification. The experimental results of this approach gave 97.1% accuracy; this helps ophthalmologists discover problems early; the device may be able to reduce the workload on hospital units.