similarity measure

********************

SAD: Sum of Absolute Differences

SAD=Σiⁿ|X_i-Y_i|

(similar lighting)some distinct points work well others don’t such as: edges, areas of uniformity

(some lighting changes and slight (2 degrees)rotation)SAD adds up constant lighting changes.

If two images are identical, but offset(偏移) in greyscale intensity by D, then the SAD score will differ by nD

**********************

SSD: Sum of Squared Differences

SSD=Σiⁿ (X_i-Y_i)²

Enhances effect of very different pixels

Derive how SSD is the same as NCC if patches are normalized

(done on the board, show normalized SSD=2-2NCC : the bigger NCC is, the lower SSD is

Correlation =Σ_iⁿX_iY_i

•NCC =(ΣX_iY_i)/(||X|| ||Y||) -turn each patch into unit vector

•CSM = Σ(X_i-Xu)(Y_i-Y_u) -subtract mean for lighting level

•VNC =(Σ(X_i-Xu)(Y_i-Y_u) /Nsd(X)sd(Y)

        =(Σ(X_i-Xu)(Y_i-Y_u)/Nsqrt[Σ(X_i-X_u)(X_i-X_u)/N]sqrt[Σ(Y_i-Y_u)(Y_i-Y_u)/N]-compensate for both lighting level and brightnes