In this paper, an improved DR-Net recognition algorithm based on DR-Net model is proposed to solve the problems of unbalanced diabetic retina image dataset, insufficient feature extraction of tissue morphology, and low classification accuracy of diabetic retinopathy. The Kaggle APTOS 2019 contest dataset is selected, which is expanded with various data enhancement strategies, and the Eye-PACS dataset is introduced for unbiased correction. Moreover, morphological methods such as Gaussian filtering are used to intensify the fundus image characteristics. Then the aggregated residual structure of ResNext50 is pre-trained and the parameters and structure of the baseline model are fine-tuned through transfer learning. In addition, the cavity convolution is introduced to replace the ordinary convolution, and the attention mechanism is also involved to further optimize the model performance. The test results show that the improved DR-Net model greatly improves the accuracy of diabetic retinopathy classification: the positive and negative predictive values reach 97.9% and 98.03%, respectively, with the accuracy being up to 98.04%, which is much higher than those of similar algorithms. In short, the screening of retinopathy with the assistance of deep learning technology is of guiding significance for the research of early automatic screening for retinopathy.