ASM: Active Shape Models--Their Training and Application

时间:2023-03-10 03:27:37
ASM: Active Shape Models--Their Training and Application

这篇论文的前半部分基本就是论文《Training Models of Shape from Sets of Examples》的全部内容,只不过多两个应用示例,后半部分在PDM模型的基础上提出了ASM模型。

对于ASM模型的介绍在博客中有详细的说明,出于时间的考虑这里仅列出我对ASM模型的部分理解,如下:

1> 在训练ASM模型时,为了确保由于ASM: Active Shape Models--Their Training and Application的变化产生的形状与训练集中的形状类似,需要对ASM: Active Shape Models--Their Training and Application的值进行一些限制(就是所谓的形状约束),即

ASM: Active Shape Models--Their Training and Application

其中ASM: Active Shape Models--Their Training and Application通常为3。如果ASM: Active Shape Models--Their Training and Application在更新过程中ASM: Active Shape Models--Their Training and Application,则使用下式对ASM: Active Shape Models--Their Training and Application加以约束。

ASM: Active Shape Models--Their Training and Application

2> 在训练集上为每个特征点计算局部特征,用于根据相似度量确定每个特征点的下一个位置,即确定ASM: Active Shape Models--Their Training and Application

3> ASM的搜索策略非常重要,具体步骤如下:

a) 对平均形状进行仿射变换得到一个初始模型

ASM: Active Shape Models--Their Training and Application

其中,ASM: Active Shape Models--Their Training and Application,是根据平均形状和PDM的参数ASM: Active Shape Models--Their Training and Application计算得到的;ASM: Active Shape Models--Their Training and Application是中心点坐标。

b) 根据第二点中计算得到的局部特征计算出每个特征点的新位置,可得位移向量

ASM: Active Shape Models--Their Training and Application

c) 此时已经知道初始模型和经过仿射变换后的模型,那么就可以计算出仿射变换参数的变化量ASM: Active Shape Models--Their Training and ApplicationASM: Active Shape Models--Their Training and ApplicationASM: Active Shape Models--Their Training and ApplicationASM: Active Shape Models--Their Training and Application。如下:

ASM: Active Shape Models--Their Training and Application

代入a)中表达式后可得

ASM: Active Shape Models--Their Training and Application

同时

ASM: Active Shape Models--Their Training and Application

所以有

ASM: Active Shape Models--Their Training and Application

再者,由PDM模型可知

ASM: Active Shape Models--Their Training and Application

上式减去ASM: Active Shape Models--Their Training and Application后可得

ASM: Active Shape Models--Their Training and Application

ASM: Active Shape Models--Their Training and Application

至此,可以求出ASM: Active Shape Models--Their Training and Application

参数的更新顺序为ASM: Active Shape Models--Their Training and Application,所以可以对仿射变换参数和b做如下更新:

ASM: Active Shape Models--Their Training and Application        式中ASM: Active Shape Models--Their Training and ApplicationASM: Active Shape Models--Their Training and ApplicationASM: Active Shape Models--Their Training and ApplicationASM: Active Shape Models--Their Training and Application是控制参数变化的权值。

此时,一次搜索过程就结束了。根据ASM: Active Shape Models--Their Training and ApplicationASM: Active Shape Models--Their Training and Application即可计算出新的形状。

d) 重复执行该搜索过程

e) 当ASM模型的参数(仿射变换的参数和b)变化不是很大或迭代次数达到指定的阈值时就结束该搜索过程。

参考文献:

1> ASM(Active Shape Model)算法介绍

2>人脸识别之ASM模型-Active Shape Models - 'Smart Snakes'

3> ASM(Active Shape Model)算法介绍