Advanced classification system for biological products
Abstract
Color classification is an important method in grading agricultural and biological materials. The objective of this thesis was to develop color classification methods for biological products, with application to real-time grading and seed corn husk deduction. A nomenclature of classification systems was developed to formalize a review of color classification methods. The description and functional classifier types were introduced. To overcome the limitations of available classifiers, when applied to real-time hardware, two original classifiers were developed using a binary representation of class assignments in the color space. The binary classifier of type one (BC1) used pairwise discriminant functions, whereas the binary classifier of type two (BC2) used a more complex logic. The binary representations can be implemented with look-up tables or template matching neural networks. Three software packages and a number of tools, implementing the color classifiers and error evaluation methods, were developed. SPR implemented statistical pattern recognition classifiers, nSPR implemented neural network based classifiers, and Purclass implemented binary classifiers. Four methods were developed to evaluate classifier accuracy: (1) the global error measurements with resubstitution error, leave-one-out error, and hold-out error, (2) the confusion matrix analysis for individual classes, (3) the dimensionality analysis computing the resubstitution errors for all possible combinations of color bands and classifiers, and (4) a set of graphical representations of color classification problems. The software allowed further analysis of neural network classifier' behavior. It was found, for the problem studied, that the learning coefficient of the binary linear classifier of type one (BLC1) did not influence the convergence of the linear algorithms. The number of iterations necessary to reach the best resubstitution error was random. The BC1 and BC2 algorithms were successfully implemented on real-time image processing hardware. Classification rate was 6 images per second with the BLC2 classifier, for a three class problem, and 512 by 220 pixel color images. The developed real-time color classification system can accurately classify seed corn images and the color vision system improves the method of deduction, over the current manual method.
Degree
Ph.D.
Advisors
Krutz, Purdue University.
Subject Area
Agricultural engineering|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.