Fig. 1. Detail comparison of algorithm content

Fig. 2. Comparison curves for different parameters of Geman-McClure function

Fig. 3. Grouping index. (a) Image cell index groups; (b) locally extended region

Fig. 4. Cross windows. (a) Expansion regions for different groups; (b) local image region

Fig. 5. Error matching rates of cross window and original method on test images

Fig. 6. Multi-category graphs. (a) Plastic; (b) Meanshift; (c) SNIC; (d) texture structure graph

Fig. 7. Improved local expansion movement algorithm

Fig. 8. Comparison of error rates between improved generation mechanism and the original method in one iteration

Fig. 9. Filling methods. (a) Texture region; (b) texture-less region. Red regions represent stable points, black regions represent unstable points, and white regions represent culling points

Fig. 10. Iterative optimization process. (a) LocalExp; (b) proposed algorithm

Fig. 11. Flow chart of proposed algorithm

Fig. 12. Optimization procedure of our algorithm

Fig. 13. Disparities of some global PatchMatch algorithms (points with error matching rate greater than 1 pixel are shown in red). (a) Image; (b) PMBP; (c) SPM-BP; (d) GCLSL; (e) PMSC; (f) LocalExp; (g) proposed