Fig. 1. Network framework

Fig. 2. Algorithm flow

Fig. 3. Cluster graph comparison. (a) Input image; (b) VGG19 space; (c) RGB space

Fig. 4. Algorithm flow

Fig. 5. Color space comparison. (a) Input image; (b) RGB space; (c) LAB space

Fig. 6. RSB structure

Fig. 7. Soft threshold function and its derivative. (a) Soft threshold function; (b) soft threshold function derivative

Fig. 8. Image representation of different network levels. (a) Original image; (b) Conv_1_1; (c) Conv_2_1; (d) Conv_3_1; (e) Conv_4_1

Fig. 9. RSB validity verification. (a) Content image; (b) without RSB; (c) with RSB

Fig. 10. The effect of style transfer with different cluster numbers. (a) Input image; (b) K=1; (c) K=3; (d) K=5; (e) K=10

Fig. 11. The effect of style transfer with different fusion coefficients. (a) Input image; (b) αk= 0.2; (c) αk= 0.6; (d) αk=1.0

Fig. 12. The effect of style transfer with different coverage ratios. (a) Input image; (b) coverage ratio is 0.25; (c) coverage ratio is 0.5; (d) coverage ratio is 1.0

Fig. 13. Comparison of different style transfer methods. (a) Content images; (b) style images; (c) method of reference [5]; (d) AdaIN; (e) SANet; (f) MST; (g) proposed method

Fig. 14. Comparison results of different data augmentation methods. (a) Original image; (b) flip horizontal; (c) rotate 90°; (d) rotate 180°; (e) rotate 270°; (f) contrast; (g) color enhancement; (h) brightness enhancement

Fig. 15. Effect of different dataset generation. (a) Original murals; (b) masked images; (c) real; (d) real+s.a.; (e) synthetic; (f) real+synthetic