Diabetic retinopathy is one of the leading causes of blindness worldwide, making timely treatment essential to minimize visual impairment. Automated detection systems for DR diagnosis have emerged because they are competent in the diagnosis of DR by analyzing retinal images. Nevertheless, distinguishing between the various stages of recognition remains an issue due to variations in image quality and inherent data set imbalances. Even though existing models such as MobileNetv2, CNN, and ResNet50 show good success in the detection of DR, they struggle to get a good accuracy level while dealing with data set imbalance issues. Additionally, there is also the overhead distraction of having data augmentation to improve an already established model. The proposed DenseNet-121 model aims to mitigate the challenges of the other models in diagnosing DR. Since DenseNet-121 uses dense connections between layers, this leads to better feature reuse and enhanced flow of the gradient to further boost the learning of the model, which we believe is the blue sky for the given field. In contrast with the other models, DenseNet-121 is entirely novel: A fresh model concentrating on maximizing classification innovation without resorting to complicated preprocessing techniques. The DenseNet-121 model has been tested extensively on an indirectly balanced dataset for DR. The proposed outcomes indicate that DenseNet-121 vastly outperforms MobileNetv2, CNN, and ResNet50 with respect to classification accuracy. The model achieves state-of-the-art results, having higher precision and recall in all stages of DR, thus showing its usage in clinical settings.