We record the images of the small particles in the flow (currently at 500 fps by using four 1280 x 1024 pixel CMOS cameras with Full Camera-Link). The image processing we're looking for is:

a) high-pass filter (to remove background uneven illumination)

b) contrast-enhancement

c) threshold binarization

d) for each particle detected we want to receive 4 values: 1,2) center-of-mass (x,y), 3) pixel size (or diameter/area in pixels) and 4) sum of grey levels for each particle (4 numbers per object) extraction.

Therefore, we expect to reduce the bandwidth for each camera from 650 Mb/s (1.3MB image x 500fps = 650MB/s) to 20KB/s, e.g. 2000 particle objects in a single frame, each particle includes four numbers that could be something like 20 bytes per object, . In the future, we plan to work with faster cameras, e.g. 1000 fps or maybe even faster.

We would like to have an option to control (in live) the following parameters of the image processing algorithm:
a) grey level threshold per camera (each camera can see the flow from a different angle)
b) size of the particle objects: min, max area in pixels, min/max size in x,y directions (horizontal/vertical number of pixels) separately
c) sum of grey level values (before binarization)
and optional:
d) minimal grey values in the object (helps treating of the overlapping particle images)


In the original e-mail we mentioned the so-called (sorry for jargon) "Sobel algorithm" - this one is used by Mikrotron GmbH in their MC1324 camera which we test. Unfortunately, the details of the algorithm are not accessible for us and therefore, we cannot provide more details. The algorithm works for us (we manage to measure things using this camera) but we cannot explain you (and ourselves) how. Only if you have an access to that algorithm, we would be glad to consider it too as the second option.

As more general statement, the algorithm described above is probably only one of many possible ways and we would not like to restrict ourselves to the one that we know only. If there are algorithms that you can suggest us that will provide a similar output from the images, we would gratefully appreciate your proposals.