Seam Carving 算法

如何能改变图像大小的同时尽可能保留图片主体？ Seam Carving 算法可以解决这个问题。

Transcript

1. Seam Carving Algorithm @Rick Sun
2. Aspect Ratio Changed Crop Original Scale
3. Content-aware Resizing Seam Carving Original
4. Example
5. Humans Vision 落款 富⼠⼭ 波浪
6. Computer Vision
7. Pixel Colors
8. Energy e(x) 192 Energy: 值越⼤，表示该像素与周围像素对比越⼤。 Energy function
9. Energy Map Energy function
10. Energy Map
11. ? How to resize? 哪些像素需要被删除？
12. 1. Optimal 依次删除 energy 最小的像素
13. 2. Pixel 每⾏移除等量 energy 最小的像素
14. 3. Seam 每次移除⼀条 energy 之和最小的 seam
15. 3. Seam 什么是连续？ 8 连通路径(8-connected path )
16. Seam Carving Recipe 1. 计算每个像素的 energy， ⽣成 energy map。 2. 找出 seam 并删除。 3. 未达到终⽌条件，回到第 1 步。
17. Implements Energy Function 01 Find Seam 02
18. Energy Function 1. 2.
19. Energy Function 3. Image Kernels RGB -> Grayscale Grayscale = (R + G + B ) / 3. Grayscale = 0.299R + 0.587G + 0.114B Kernels
20. Refer: https://setosa.io/ev/image-kernels/
21. Implements Energy Function 01 Find Seam 02
22. Find Seam
23. Find Seam Brute-force search O(w * 3^h) O(w + h) The greedy approach
24. Dynamic Programming Those who can not remember the past are condemned to repeat it.
25. Dynamic Programming Those who can not remember the past are condemned to repeat it. Stairs Problem Go up one step or two steps at a time House Robber Problem • You are a robber who has found a block of houses to rob. • You cannot rob two adjacent houses.
26. Energy Map Find Seam Seam Energies ? 80+45 seam[i][j] = EnergyMap[i][j] + Min(seam[i-1][j-1], seam[i-1][j], seam[i-1][j+1]) 45 405 35 275 1o5 25 100 25 40 5 20
27. Find Seam
28. Seam Energies 100 25 130 275 Find Seam interface SeamEnergyCell { i: number; j: number; energy: number; prev: SeamEnergyCell | null; } 45 35 50 10
29. Other Applications Object Label 01 Image Enlarging 02
30. Object Removal Set energy to -Infinity
31. Object Protect Set energy to +Infinity
32. Image Enlarging Repeat Seam Repeat K Seams
33. Top K Problem Find Top K largest numbers O(n log n) 1. Sort 2. Heap sort
34. Heap • Max Heap Build Heap Get root Heapify O( n) + kO(log n) • Min Heap 8 9 5 Build Heap K Add Heap if lt root 8 9 7 Add Skip O( k) + (n - k)O(log k)
35. 3. Quick select Top K Problem O( n) K th?
36. Thanks!