Notes about error tests with nih-image
--------------------------------------

The .hqx file should decompress to a folder containing two folders (pre-
and post- version 1.53).  Each of these folders has 3 CricketGraph 3 files
- 2 data files and 1 graph file + the odd format file.

The tests were done with two versions of nih-image to test the suitability
and accuracy of two different perimeter measurement routines, under the
'AnalyzeParticles' command.

Particles were orginally generated and rotated in SuperPaint 3.0 as draw
(vector-based) objects to prevent the sort of degradation of the edges 
of a particle you get when rotating a bit-map object.

The top graphs show the effect of size on measured perimeter - the noise
at smaller particle sizes reflects the problem of representing a curve
by small blocks, and therefore the effect of measuring 'steps' in particle
edges.  The lower graphs are to look at the effect of perimeter as a function 
of particle orientation.  The perimeters are expressed as a function of 
the true perimeter, which was calculated from the particles' measured  area.
This was because the rotation in SuperPaint inevitably involves a small change
in area.  As a consequence the measured perimeters were standardised to a
value around 2500 (this is why the columns 'corrected_perimeter' and 
'predicted_perimeter' appear).

The 3-pixel moving average method (post- 1.53) in effect assess the shape 
of the pixel grouping and weights its results accordingly.  The shape can 
only be one of five (I think) -

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

For aspect ratios of 5:1 and below errors on the perimeter measurement are
below 4%.  For very elongate objects they may be larger.  For
spiky and dendritic objects, they are likely to be worse, but there is 
a problem in having a measurement control for these sorts of things.  It 
depends on what you're doing and how much discrimination in perimeter 
measurement you require whether or not this is OK - they are fine for 
my purposes where the particles tend to be low aspect ratio.

A 4- or 5-pixel moving average may cope with spiky objects better, but
the number of possible arrangements of this many pixels increases quite 
dramatically and will produce a speed penalty.

Richard Herd - 7th April 1995

Richard Herd
Volcanology Group
Environmental Sciences Dept.
Lancaster University
Bailrigg
Lancaster, LA1 4YQ
U.K.