Pattern recognition and sensor fusion for quality sorting

Nissim Ozer, Purdue University

Abstract

A methodology was developed for automatic sorting of agricultural produce, using multiple sensors and was applied to cantaloupes as a case study. The methodology includes the examination of a multi-sensor fusion approach with regard to grading criteria, human classification, single sensors, multiple sensors and classification methods. Data acquired from eight sensors: vision, two firmness sensors, fluorescence, color sensor, electronic sniffer, refractometer and scale (weight) were analyzed and provided input for five classification models. The results indicated that fluorescence response and firmness measured by an impact sensor were the best indicators for fruit maturity among the sensors evaluated. A procedure was developed to estimate and minimize the training size for supervised classification. New criteria were developed to choose a training set such that a recurrent auto-associative memory neural network is stabilized. These methods developed for training size selection and the criteria for choosing training samples provides a prediction of classification error and permits rapid training to compensate for fruit variation from seasonal differences, cultivate differences, and growing environment conditions. An expert system was developed to measure variances in human grading, and procedures for the evaluation of classification performance were established. This procedure ensures that the evaluation of performance is not affected by the variability in human grading. Mathematical formalization for incorporating multiple sensors was presented, and parametric and non-parametric classifiers for grading were examined. It has been shown that multiple sensors significantly improve classification accuracy. Statistical methods were found to be as accurate as neural networks in grading. Classification models by voting didn't enhance the classification significantly. A hybrid model that incorporated heuristic rules and a numerical classifier was found to be superior in classification accuracy and with half the processing time used solely by the numerical classifier.

Degree

Ph.D.

Advisors

Simon, Purdue University.

Subject Area

Food science|Electrical engineering|Industrial engineering|Artificial intelligence

COinS