The features from the fundus pictures' optic disc must be precisely located before landmark features included in the fundus images can be identified. According to the severity of the DR, existing research employed a variety of Artificial Intelligence (AI) strategies for screening and diagnosing DR sooner to protect diabetic patients from going blind. In contrast to the real-world optimization strategy, the current models, while effective, had limitations related to time consumption and premature convergence. As a result, the Multi-class Transfer Learning that has been proposed is updated to improve performance based on its fundamental design. The model was trained to find solutions more effectively thanks to the higher convergence. The suggested approach gets around constraint problems and improves accuracy with better feature selection. In comparison to the current transfer model, which achieved 98.94% accuracy for DIARETDB1, the suggested method achieved an accuracy of 97.46%. The e-ophtha dataset, however, achieved accuracy of 98.91%. The predicted algorithm can be integrated with any device at home-centric environment.