Fig. 1. Point clusters after K-L transformation in color coordinate space

Fig. 2. 3×3 normal convolution and deformable convolution diagram

Fig. 3. Deformable convolution feature extraction process

Fig. 4. Internal structure of dense deformable convolution model

Fig. 5. Multiscale dilated convolution

Fig. 6. Internal structure of AGs

Fig. 7. Computation of compatibility tensor dimension

Fig. 8. U-shaped network architecture model of adaptive scale information

Fig. 9. Preprocessing images in each stage. (a) Original image; (b) image of green channel; (c) principal component I1; (d) principal component I2; (e) principal component I3; (f) morphological transformation

Fig. 10. Local vessel block informations. (a) Gold standard block information; (b) DRIVE data block information

Fig. 11. Retinal vascular segmentation results using different algorithms. (a) Origin image; (b) gold standard of image; (c) proposed algorithm; (d) results in Ref. [9]; (e) results in Ref. [28]; (f) results in Ref. [29]

Fig. 12. Performance comparison of deep learning segmentation algorithms. (a) Origin images; (b) details of original images; (c) details of gold standard; (d) details of proposed algorithm; (e) details of algorithm in Ref. [9]; (f) details of algorithm in Ref. [29]

Fig. 13. ROC curves of proposed algorithm. (a) ROC curve of DRIVE dataset; (b) ROC curve of STARE dataset

Fig. 14. Fmeasure of different algorithms. (a) Fmeasure of DRIVE dataset; (b) Fmeasure of STARE dataset