|
Original Image |
Detection Results (New) |
Detection Results (Old) |
|
|
|
 |
|
|
|
 |
|
|
|
 |
|
|
|
 |
|
|
|
 |
|
|
|
 |
Minirhizotrons are small transparent tubes buried at an angle in the ground to observe the underground roots of neighboring plants. At regular intervals, researchers collect data by sliding a miniature camera into each tube and collecting still images of the roots seen through the tubes. The result is an overwhelming amount of data that must be analyzed, traditionally by hand. The goal of this research is to automate the procedure of extracting, identifying, and measuring roots in minirhizotron images.
System Description
In previous work [2], we developed an algorithm that applies two-dimensional matching filtering followed by local entropy thresholding to produce binarized images, from which roots are detected. After applying a root classifier to discriminate fine roots from unwanted background objects, a root labeling method is implemented to identify each root in the image. Once a root is identified, its length and diameter are measured using Dijkstra's algorithm for obtaining the central curve and using the Kimura-Kikuchi-Yamasaki method for measuring the length of the digitized path.
The previous algorithm was computationally demanding, requiring tens of seconds of processing per image -- thus limiting its usefulness. To overcome this problem, we have developed an approach for rapid, automatic detection of plant roots in minirhizotron images. The problem is modeled as a Gibbs point process with a modified Candy model, whose energy functional is minimized using a greedy algorithm whose parameters are determined in a data-driven manner. The speed of the algorithm is due in part to the selection of seed points, which discards more than 90% of the data from consideration in the first step. Root segments are formed by grouping seed points into piecewise linear structures, which are further combined and validated using geometric techniques. Once root centerlines are found, root regions are then detected using a recursively bottom-up region growing method. The new algorithms is not only faster than the old method by orders of magnitude, it is also more accurate.
Results
The following table compares the results of the two algorithms on several images. The colors used to outline the detected roots are arbitrary, with different colors indicating different roots.
|
Original Image |
Detection Results (New) |
Detection Results (Old) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Database
A database containing 1190 minirhizotron images (640 x 480) from three different plant species (peach, maple, and magnolia) is built for developing and testing our algorithms. Click here (61 MB) to download the images and results.
Publications
[1]
G. Zeng,
S. T. Birchfield, and C. E. Wells. Rapid Automated Detection of Roots
in Minirhizotron Images.
Machine Vision and Applications, 2008 (under review).
[2]
G. Zeng,
S. T. Birchfield, and C. E. Wells.
Detecting and Measuring Fine Roots in
Minirhizotron Images Using Matched Filtering and Local Entropy Thresholding.
Machine Vision and Applications, 17(4):265-278, 2006.
[3]
G. Zeng. Automatic Minirhizotron Root Image Analysis Using Two-Dimensional Matched Filtering And
Local Entropy Thresholding. Master's thesis, Dept. of Electrical and Computer Engineering, Clemson University,
May 2005. (slides)
This work was supported by the National Science Foundation under grant DBI-0455221.
