https://echo360.org.au/lesson/1dff8680-04a8-4cd7-ac45-0822b6f936e9/classroom#sortDirection=desc
总结:
1.Basic Segmentation methods
1) Thresholding
(1) 低于Threshold为background 高于为Foreground ; binary-segmentation
(2) regions有重叠的灰度大小,就难以分割了,要么background变得noisy,要么foreground不整齐
2)K-means clustering
(1)• Could work if the number of clusters is known a priori
(2)有时候regions多,threshold就不好用了,可以用 k-means (例1中用了3-means)
(3) 对于另一些图 k 不好确定,k-means就不太能用
3)Feature extraction and classification
(1)将Image分成patches,然后sliding window提取features 做classification
2.More sophisticated segmentation methods
1) Region splitting and merging**
(1) 计算 Connected Components : 4-connected , 8-connected两种方法计算出的数量是不一样的
(2)Connected Components Algorithm:
1. 从上到下,从左到右,两层遍历所有pixel
2. 如果pixel是object_pixel,检查其Neighbor,neighbour种有label就标上最小的那个label,没有就标上新的
3. 对于相连而不同labels的pixels,记录等价的labels: Equivalence sets {1,2,6} {3,4,5}
4. 再遍历一遍,对于pixel赋值 Equivalence sets中最小的值,背景赋值为0
(3)Region splitting
1. 在直方图中找到最好的分割开peaks的threshold——t (峰谷),重复直到regions are either fairly uniform or below a certain size
2. 递归版本的分割,从左下的图开始,原始图片入栈,分割图片成几个clusters,几个子图入栈,然后pop出来继续。
(4)Reigion Merging
1.Heuristics-based region merging
2. Graph-based region merging
3.Merging by region growing
2) – Superpixel segmentation
3) – Watershed segmentation
(1) 选几个markers开始,一般是自动选取Local minimum
(2)intensity从下到上,intensity越小,priority越大,入栈
(3)pop出priority最大的pixel,如果它的neighbour全是同一个label,则assign它同样的label,如果neighbour不是全一样,则ignore it
(4)注意:
1.有时候会oversegment,一个object分成好多个,这个时候要么prepossesing要么postprocessing
2.object是白色时,需要intensity inversion 来找local minimum
4)– Conditional random field
5)– Maximally stable extremal regions
6)– Active contour segmentation
7)– Mean-shifting algorithm
1.在特征空间中,随机选择pixel作为start point,计算中心点(带权重的),并把start point迁移到中心点;不断迭代直到收敛
2. 用高斯公式计算权重,离中心点越近的权重越大
3. 优点:
(1)不需要获知 k , 可以自己找到 number of clusters
(2)只有一个变量
4.缺点:
(1)计算量大,因为要计算所有的k
(2)变量window size很难解释
(3)高维空间难以解释
8)– Level-set segmentation
3.How to evaluate segmentation methods
1) Quantitative evaluation metrics
2)– Receiver operating characteristic