Fig. 1. Example of 2D color fundus

Fig. 2. Examples of DR grading sample

Fig. 3. Original image and RGB channel output image

Fig. 4. Image preprocessing process and means. (a) Original image; (b) filter denoising; (c) contrast enhancement；(d) light equalization

Fig. 5. Architecture of improved DR－Net network model

Fig. 6. Dilated convolution under different expansion rates. (a) d＝1; (b) d＝2; (c) d＝3

Fig. 7. Architecture of aggregated residual

Fig. 8. Structure diagram of attention mechanism

Fig. 9. Examples of improved DR－Net test results

Fig. 10. Accuracy and error curve in training process of improved DR－Net. (a) Accuracy; (b) train loss

Fig. 11. Confusion matrix test results of improved DR－Net model

Fig. 12. Comparison of confusion matrix among mainstream models on DR classification. (a) Densenet121；(b) Mobilenetv2; (c) Efficient－net b5; (d) ResNext50

Fig. 13. Comparison of performance on DR－Net and improved DR－Net models