相关与模板匹配

Correlation is convolution without the flip

Finding a letter with a template

Normalizing the score

What matching survives

A kernel shaped like a rising ramp, small values on the left and large on the right, is applied to an image once by convolution and once by cross-correlation. How do the outputs differ?

Raw cross-correlation is used to find a dark letter template on a bright scanned page, and the response map peaks strongly on a plain bright margin. What went wrong?

A template has pixel values $(10, 10, 10, 10)$. After subtracting the template's mean, write the value of the first pixel.