![](bg7.png)
Figure 9: Example root trace using a mixed state model
consisting of two states, normal growth (white) and
gravitropic (black).
5 CONCLUSIONS
5.1 Discussion of Results
The results comparing the software root length
measures to the manual measurements show the new
technique to produce results to about 2% of the
actual measures most of the time. There was a
larger error when comparing the new software with
the expert user (2.2%) compared to the non-expert
(1.7%), however the images in Section 3.2 are more
challenging than those in Section 3.1, which may
account for some of the increased error also.
Something to be wary of with these kinds of
comparisons is using manually marked-up ground
truths to compare with the automated measurements.
There is an inherent subjectivity in determining the
length of the roots, dependant on, for example, the
accuracy with which the curves in the roots are
traced. The more finely the shape of the root is
followed, the longer the measurement. There is
similarity here with the coastline measuring
problem. Some structures can be thought of as
fractal in composition, such as a coastline
(Mandelbrot, 1967) or complete root systems (Eshel,
1998). When trying to measure such systems, the
scale (or accuracy) with which the waves and
perturbations are traced has a bearing on the overall
length calculated. This software can be thought of as
producing the finest scale estimate of length
available at the image resolution, and so is likely to
overestimate length compared to a manual
measurement. This may be reflected in the results
reported in Section 3, with most errors indicating an
overestimate of line length.
Even if a user and the new software were to use
the same scales of measurement, there is still human
error present in the measuring process, which can be
quantified by the standard deviation of the manually
measured data. The manual measurements in section
3.1 give an average standard deviation of ~2 pixels.
Therefore most (99%) of the manually measured
lengths can be expected to fall within about 6 pixels
(three standard deviations) of the true value for roots
of around the length seen in section 3.1. The new
software used on these roots has an average relative
error of 1.7% which translates to a error of 2.1 pixels
on average for these roots, and therefore this
software error falls within the expected error bounds
of manually entered data.
The time to use the new software was less than
the time to take the measurements manually. This
should be improved upon still when implementation
of the root tip finding algorithm is completed. The
system should be less fatiguing to the user as less
high-accuracy input is required. This will help to
lower the number of mistakes made over the course
of measuring many roots.
Labelling of the different growth modes of the
root as illustrated in Section 4.2 is also ongoing
work, but early results indicate the system can be
used for identifying different ways in which a root
trace line is produced, as long as trace motion
models exist to sufficiently differentiate the modes
of production of the line.
5.2 Improving the Reliability
As it stands, the software is still in trial stages and
reliability is still being improved. There are a
number of possible ways to decrease the number of
errors that can occur. One problem is as the particle
filter tracks the root towards the tip, it is liable to
trace lateral roots if they are long enough and
provide a high enough quality measurement, as
shown in Figure 5. A simple way to remove this
problem is to simply trace the root from the end tip
upwards. Due to the geometry of the lateral roots the
tracing algorithm is then not presented with viable
alternative routes until the lateral roots join and
terminate. Therefore, the only way they can be
followed is if they lie parallel to the main root for
long enough, and are close enough for the particles
on the tracing algorithm to ‘jump’ across to the other
track. The difficulty with this approach, however, is
that the tracker would have to be started on the
thinnest, least visible section of the root, which may
be hard to detect, and automatic termination of the
tracking becomes harder as the delineation at the top
of the root is less clear.
Other general improvements include increasing
the resolution of the images, as during testing at
least some of the mis-tracing of the roots was due to
poor representation of the roots in the image.
Improving the measurement model may lead to less
problems with the system tracking lateral roots.
Finally, increasing the number of samples may be
beneficial, especially in combination with greater
image resolution. However, in such a case speed of
BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing
114