Fig. 1. Network framework diagram

Fig. 2. VGG-19 network feature extraction

Fig. 3. Representation of foreground and background of the image

Fig. 4. Linear combination of mask color channels

Fig. 5. Deep matting network structure

Fig. 6. Comparison of matting effects. (a1) (a2) Original images; (b1) (b2) Trimap; (c1) (c2) traditional matting; (d1) (d2) depth matting of this paper

Fig. 7. Transform network structure

Fig. 8. Comparison of style transfer texture of different convolution kernels. (a1)(a2) Content images; (b1)(b2) style images; (c1)(c2) 7×7 convolution kernel transfer texture; (d1)(d2) 3×3 convolution kernel transfer texture

Fig. 9. Graphs of loss functions. (a) Content image; (b) style image

Fig. 10. Texture comparison. (a1)(a2) Content images; (b1)(b2) style images; (c1)(c2) before normalization; (d1)(d2) after normalization

Fig. 11. Effect of style and content trade-off.(a) η=0; (b) η=0.25; (c)η=0.50; (d) η=0.75; (e) η=1.00; (f) style image

Fig. 12. Transfer results of this article

Fig. 13. Comparison of experimental results. (a1)--(a4) Content images; (b1)--(b4) style images; (c1)--(c4) images based on method proposed by Gatys et al^[5]; (d1)--(d4) images based on method proposed by Ulyanov et al^[10]; (e1)--(e4) images based on method proposed by Huang et al^[12]; (f1)--(f4) images based on method proposed by Chen et al^[26]; (g1)--(g4) images based on method proposed by Li et al^[27]; (h1)--(h4) images based on our method

Fig. 14. Objective evaluation. (a) MSSIM; (b) PSNR; (c) AvG; (d) Q^AB/F

Table 1. Stride length and receptive field parameter settings

Table 2. Transfer network parameters