From

Since you'd expect a ridge/peak detection filter to have a positive response to a positive ridge/peak, you'd use the negated second derivative, and add that.

Look at these pictures:

1: 2: 3:

1. the [-1 5 -1] kernel, i.e. identity - laplacian = identity filter
2. the picture itself
3. [ 1 -3 1] kernel, i.e. identity laplacian = identity - filter

You see, #3 looks blurry because some high frequencies were subtracted.