图像的联合概率分布

如何计算两幅图像的联合概率分布

https://www.medphysics.wisc.educourses/mp573/2010/mutualinformation_revised.pdf

中给出的定义是：

The spatial information obviously required for a registration method is provided by the definition of a joint probability distribution that depends simultaneously on image A and B. The conventional expression for it is $p_{A,B}(a,b)$. It is calculated as the number of times out of the total number of pixels N that a pixel in A contains the value a and the same pixel that is, the pixel in the same image position, in B contains the value b; this number of pixels is then divided by the total number of pixels to give the joint probability of a,b.

个人的理解是：

• 图像配准需要的空间信息取决于图像A和B的联合概率分布，它的常规表达式是 $p_{A,B}(a,b)$ 。
• 统计图像A中像素为a的位置，计算B中处于和a相同位置的像素b的个数n；
• 将n除以像素总数N，得到a，b的联合概率。

假如图像分别为
$$A=\left(\begin{array}{lll}1& 1& 2\\ 2& 7& 1\end{array}\right)$$
$$B=\left(\begin{array}{lll}1& 2& 2\\ 2& 1& 2\end{array}\right)$$
则图像的联合概率分布为
$$p_{A,B}(1,1)=\frac{1}{6}$$
$$p_{A,B}(1,2)=\frac{2}{6}$$
$$p_{A,B}(1,7)=\frac{0}{6}$$
$$p_{A,B}(2,1)=\frac{0}{6}$$
$$p_{A,B}(2,2)=\frac{2}{6}$$
$$p_{A,B}(2,7)=\frac{0}{6}$$
$$p_{A,B}(7,1)=\frac{1}{6}$$
$$p_{A,B}(7,2)=\frac{0}{6}$$
$$p_{A,B}(7,7)=\frac{0}{6}$$

• 虽然有$3^2$个可能的有序像素值对需要考虑，但总匹配数将只有6个。至于我们是否应该称之为“两幅图像的联合概率分布”——这只是作者们想要使用的术语。有很多种方法来获取图像的属性并定义联合分布。

https://www.physicsforums.com/threads/joint-probability-distribution-of-two-images.589741/

参考两幅图像联合熵的概念，对上述理解用Matlab进行验证

%A=rgb2gray(imread('img1.png'))
%B=rgb2gray(imread('img2.png'))

% 给出A,B矩阵
img1=[1,1,2;2,7,1];
img2=[1,2,2;2,1,2];
A=double(img1)+1;
B=double(img2)+1;
[M,N]=size(A);
Large=max(max(max(A),max(B)))+1;
temp=zeros(Large,Large);
for m=1:M
	for n=1:N
		i=A(m,n)
		j=B(m,n);
		temp(i,j)=temp(i,j)+1;
	end
end
temp=temp./(M*N);

验证temp为
$$\left[\begin{array}{cccc}1/6& 2/6& \cdots & 0\\ 0& 2/6& \cdots & 0\\ ⋮& ⋮& \ddots & ⋮\\ 1/6& 0& \cdots & 0\end{array}\right]$$
与上述计算一致。