Automatic Minirhizotron Root Image Analysis Using Two-Dimensional Matched Filtering and Local Entropy Thresholding Presented by Guang Zeng.

Automatic Minirhizotron Root Image Analysis Using Two-Dimensional

Matched Filtering and Local Entropy Thresholding

Presented by Guang Zeng

Importance of Studying Roots

Methods for Studying Roots

Soil Sampling Rhizotron Minirhizotron

Previous work on minirhizotron image analysis [Vamerali & Ganis 1999] Nonlinear contrast stretching technique is used to

enhance the local contrast of rootsLimitation: The minimum root length filter will eliminate some shorter roots.

[Natar & Baker 1992] An artificial neural system is developed to identify roots

Limitation: The accuracy will substantial decrease when applied to images that have not been trained.

[Dowdy & Smucker 1998]The length-to-diameter ratio is used to discriminate roots Limitation: Only works for a single type of root.

Preview of Experimental Results

Original image Extracted root Measured root

Approach Overview

Image Preprocessing1. Conversion to grayscale

2. Contrast stretching

3. Smoothing the image

.204,153

),(255

),(255

),(

),(0

),(

21

2

2112

1

rr

rjiIif

rjiIrifrr

jiI

rjiIif

jiJ

Red Green BlueColor

We set

Matched Filtering PrinciplesSimilarity between plant roots and blood vessels:

• Small curvature

• Parallel edges

• Young roots appear brighter

• Gaussian curve for gray level profile of

cross section:

)}2

dexp(k1{A)y,x(f

2

2

Motivation:

piecewise linear segments[Chaidhuri et. 1989]

Matched Filtering Procedure

• A number of cross sections of identical profiles are matched simultaneously. A kernel can be used which mathematically expressed as:

• Kernels, for which the mean value is positive, are forced to have slightly negative mean values in order to reduce the effect of background noise.

)2

xexp()y,x(K

2

2

for |y| ≤ L/2

where L is the length of the segment for which the root is assumed to have a fixed orientation.

Matched Filtering Procedure (cont.)• The kernel is rotated using an angular resolution of 15°

(12 kernels are needed to span all possible orientations).

• The kernel is applied at two scales (full image size and half image size, obtained by subsampling).

(a) 15° (b) 75° (c) 135° (d) 180°

Matched Filtering Output

(a) 75° (b) 90° (c) 135° (d) 180°

Local Entropy ThresholdingShannon’s entropy

)plog(pH i

n

1ii

and

.10,11

i

n

ii pp

M

l

N

kij klt

1 1

),(

otherwisekl

jklfandiklf

or

jklfandiklf

ifkl

0),(

),1(),(

)1,(),(

1),(

where,

Local Entropy Thresholding (cont.)

The probability of co-occurrence pij of gray levels i and j can

therefore be written as:

i jij

ijij t

tp

Divide co-occurrence matrix into quadrants, using threshold t (0 ≤ t ≤ L)

The local entropy is defined by the quadrants A and D.

Background-to-background entropy:

Aij2

t

0i

t

0j

AijA PlogP

2

1)t(H

Foreground-to-foreground entropy:

Dij

L

1ti

L

1tj2

DijD PlogP

2

1)t(H

Hence, the total second-order local entropy of the object and the background can be written as:

The gray level corresponding to the maximum of HT(t) gives the

optimal threshold for object-background classification.

)t(H)t(H)t(H DAT

Local Entropy Thresholding (cont.)

Local Entropy Thresholding Outputs

t = 122 t = 103t = 155 t = 130

(a) 75° (b) 90° (c) 135° (d) 180°

Selecting the Root

1. Connected component labeling

2. Root candidate selecting (Ai ≥ 0.8 Amax )

Comparison of Root Selection Methods

originalimage

combined MF output

[Chanwinmaluang and Fan 2003]

Our method

detected root detected root

...

separate MF outputs

Root Measurement

1. Object Skeletonization

2. Extracting medial line using Dijkstra’s Algorithm

Root Measurement (cont.)3. Estimating the length

Freeman formula

od NN2L Pythagorean theorem

2/12od

2d ])NN(N[L

Kimura’s method2/N])2/NN(N[L o

2/12od

2d

Root Measurement (cont.)

4. Estimating the average diameter

Step 1Select 10 nodes that equally divide the medial line into 11 parts.

Step 2Find the corresponding opposite boundary point pairs, calculate the distance between each opposite boundary point pairs.

Step 3Discard the two pairs that yield the maximum and the minimum distance

Root Discrimination

1. A bright extraneous object

2. Uneven diffusion of light through the minirhizotron wall

False positives are caused by

Root Discrimination: Five Methods

1. Eccentricity

e = c / a

2. Approximate line symmetry

3. Boundary parallelism

Root Discrimination: Five Methods (cont.)

4. Histogram Distribution 5. Edge Detection

Experimental Results

• We tested our method on a set of 45 minirhizotron images containing

• different sizes of roots

• different types of roots

• dead roots

• no roots

• The output of the algorithm is compared with hand-labeled

ground truth provided by the Clemson Root Biology Lab.

Experimental Results (cont.)

Original image Extracted root Measured root

Experimental Results (cont.)

Original image Extracted root Measured root

Experimental Results (cont.)

Original image Extracted root Measured root

Comparison of Root Length Measurement Methods

1. Measurement Deviation

2. Correlation

Measurement Devi ati on (%) Freeman Formul a Ki mura' s Method

Max 17.99 14.42

Mi n 0.296 0.074

Avg 7.99 4.56

Comparison of Root Discrimination Methods1. The optimal threshold point is the closest point to the perfect result. The closer the optimal threshold point to the point (0,1), the more accurate the method.2. The larger the area beneath an ROC curve, the more accurate the method.

Multiple root detection

• Works on some images, but the false positive rate is increased to 14% (more bright background objects are misclassified).

• Our technique is limited to zero or one root per image.• We tried detecting multiple roots by extracting the two largest components in the thresholded binary images, then running our algorithm.• Some results:

Conclusion

Fully automatic algorithm for detecting and measuring roots

• Works on multiple root types • Uses individual matched filters outputs, without first combining them.• Uses a robust thresholding method• Robust medial line detection using Dijkstra’s algorithm• Proposed five different methods for root / no-root discrimination

Future work

1. Accurate multi-root detection

2. Reducing the computation time