Husk Deduction
Husk deduction is the correction process of a seed corn load for husk weight. The application was developed to measure the proportion of husk and corn on a conveyor belt as corn is unloaded. A color classifier was set up to measure the husk/corn ratio in real time. The system is using Matrox (tm) boards and perform classification into four color classes at 6 image/second.
In this case the color classes are:
- Background
- Green husk
- White husk
- Yellow kernel
Example Image
The sample points (data base) can be represented as an image.
Data Base Image
Each band in the image represents a color class sampled from the test image. This data base comprised of 6,305 points. This image can be used to visually estimate a color classifier capabilities. The following legend will be used to show the classified images.
Background.......Green Husk.......White Husk.......Yellow Kernel.....Error
Different color classifiers are available in the software packages developed in the Agricultural Engineering Department. These software, SPR, nSPR, and Purclas, implement statistical pattern recognition classifiers (SPR), and neural network based classifiers (nSPR and Purclas).
Our lab system set-up is represented in the following image.
Lab System Set-up
3D data base animation :
We found out that 3D graphs were important to understand the workings of a color classifier. However, many people have problems seeing in 3D, and that animation of 3D graphs allows a better understanding. An animation (~500Kb) of the 3D plot of the data base can be viewed.
The following images present a visual evaluation of some algorithms. First the data base is classified showing the resubstitution error of the classifier, then the example image is presented.
Linear Bayes classifier:
Data Base Classification
Example Image Classification
Quadratic Bayes classifier:
Data Base Classification
Example Image Classification
MultiLayer Perceptron classifier (Neural Network):
Data Base Classification
Example Image Classification
Binary linear classifier of type I (Neural Network):
Data Base Classification
Example Image Classification
Binary linear classifier of type II (Neural Network):
Data Base Classification
Example Image Classification
Many other color classifiers are implemented within our software. Other sections of the color classification document will describe the classifiers in more details.
These examples show the variation existing between color classifiers. For the husk deduction application it was found that the binary linear classifier of type II was the best classifier in term of resubstitution error. Husk deduction can now be measured 6 times per second.