I = αF + (1 − α)B. (2)
We call α the alpha channel model of image I (see
Fig. 2(a)(b) for an example). If we assume both F and B
are locally smooth1, then
I ∗ h = αF ∗ h+(1−α)B ∗ h = αbF +(1−αb)B, (3)
where αb = α ∗ h is α blurred by kernel h. From Eqn. 3,
we know that αb is the alpha channel model of Ib (see
Fig. 2(c)(d) for an example). So the LHS of Eqn. 1 is
∇(αbF + (1 − αb)B) · b = (F − B)∇αb · b. (4)
来自文献 Motion from Blur cvpr2008
公式（3）我一直看不明白 尤其是那个ab 卷积好像算得有问题吧 我晕脑壳了
回复此楼
