归一化互相关损失

LNCC

在医学影像配准论文《U-Net vs Transformer: Is U-Net Outdated in Medical Image Registration》中，使用了LNCC（local normalized
cross-correlation局部归一化相关损失）作为相似性度量。查阅文章引用的论文，描述如下：

LNCC代码：

class NCC(torch.nn.Module):
    """
    local (over window) normalized cross correlation
    """
    def __init__(self, win=9, eps=1e-5):
        super(NCC, self).__init__()
        self.win_raw = win
        self.eps = eps
        self.win = win

    def forward(self, I, J):
        ndims = 3
        win_size = self.win_raw
        self.win = [self.win_raw] * ndims

        weight_win_size = self.win_raw
        weight = torch.ones((1, 1, weight_win_size, weight_win_size, weight_win_size), device=I.device, requires_grad=False)
        conv_fn = F.conv3d

        # compute CC squares
        I2 = I*I
        J2 = J*J
        IJ = I*J

        # compute filters
        # compute local sums via convolution
        I_sum = conv_fn(I, weight, padding=int(win_size/2))
        J_sum = conv_fn(J, weight, padding=int(win_size/2))
        I2_sum = conv_fn(I2, weight, padding=int(win_size/2))
        J2_sum = conv_fn(J2, weight, padding=int(win_size/2))
        IJ_sum = conv_fn(IJ, weight, padding=int(win_size/2))

        # compute cross correlation
        win_size = np.prod(self.win)#返回给定轴上的数组元素的乘积
        u_I = I_sum/win_size
        u_J = J_sum/win_size
  
        #cross:计算协方差 I_var和J_var：计算方差部分
        cross = IJ_sum - u_J*I_sum - u_I*J_sum + u_I*u_J*win_size
        I_var = I2_sum - 2 * u_I * I_sum + u_I*u_I*win_size
        J_var = J2_sum - 2 * u_J * J_sum + u_J*u_J*win_size

        cc = cross * cross / (I_var * J_var + self.eps)

        # return negative cc.
        return -1.0 * torch.mean(cc)

补充

$Cov(x,y)=E[(X-E(X))(Y-E(Y))]=E[XY-XE(Y)-YE(X)+E(X)E(Y)]$
$D(X)=E[(X-E(X){)}^2]=E[(X^2-2XE(X)+E(X{)}^2]$

参考：NCC