Computer Vision summaries

Chapter 7. Image sharpness.

7.1. INTRODUCTION.

In previous Chapters it has generally been assumed that the images presented for spatially
interactive analysis are 'optimally' blurred (i.e. just sufficient blur to pass maximum information
without introducing significant sampling noise) (e.g. Chapters 2.2 and 3.3). This in turn implies one
of two things. Either

i) The original input material is very sharp relative to the sampling employed (with edges

approximating to true step functions) and the input is subjected to a Gaussian blurring

with a standard deviation (S.D.) estimated to be ideally of about 1.3 sampling intervals

(see Chapter 3.3). or

ii) Small additional blurring has been carried out on an image which is already blurred, but

such that the composite blur is roughly Gaussian, with an S.D approximately equivalent

to the estimated optimum Gaussian blur function of S.D. = 1.3 sampling intervals.

What does this mean in practice?

It is a fact that the majority of physical edges in natural scenes are sharp, but there are significant
occasions when edges in such natural scenes are unsharp. This most particularly applies to shadow
edges cast by local light sources which are other than point sources. For instance, shadow edges in
a natural outdoor scene will always be unsharp (since the solar source subtends a significant angle
nearly equal to 0.5 degrees). It is my firm opinion that one should not try to compensate for such
unsharp shadow edges, but rather that, as will become clear later, one should attempt to sense the