This site is devoted to mathematics and its applications. Created and run by Peter Saveliev.

Testing Pixcavator

From Mathematics Is A Science

Jump to: navigation, search

Pixcavator has been used numerous times in real life applications. The results were verified by the users, see Image analysis examples, especially Measuring and Counting.

The ground truth type testing is described below.

1 Area and contrast

Pixcavator was tested on simple images so that its results can be verified manually.

Dimensions of rectangles were recorded. Then their areas were computed by hand (pixel by pixel) and compared to Pixcavator's results.

Rectangles.jpg

One thing to keep in mind is the way we compute areas. Suppose we have a square and its corners are (5,5), (5,100), (100,5) and (100,100). Then one might think that the area should be (100-5)*(100-5). But since these pixels are tiles not just markers, all 4 corner pixels (and their rows and columns) are included in the square. So, the correct result is (100-5+1)*(100-5+1).

The results were identical.

As the image was created the gray levels of the squares were recorded. Then the contrasts were computed manually.

The results were identical.


2 Perimeter and roundness

The dimensions of rectangles and circles were recorded and their parameters were computed, then they were rotated and the roundnesses produced by Pixcavator were compared to the the original, expected values.

Rec-circles.JPG

The results were favorable for large enough images, see Roundness. For small, low resolution images and for small objects containing just a few pixels the the value of perimeter and roundness are off. The results are OK if we keep in mind that pixels are small.

3 Counting

Pixcavator's results (left) were verified manually and with ImageJ's mechanical cell counter (right).

The results are very close even except for the cases when some clusters aren't broken into actual particles.

Testing counting with imagej 1.jpg Testing counting with imagej 2.jpg

The ambiguity of the concept of objects in gray scale images has also to be taken into account.


DLPX.png


See also: