PATTERN RECOGNITION OF BARLEY SEEDS USING FOURIER DESCRIPTORS AND NEURAL NETWORKS
Date
1994-08
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
Type
Degree Level
Masters
Abstract
Emulation of human visual inspection using automated systems is a growing field of study especially in agricultural applications where quality inspection and minimal product handling are required. Seed inspection and grading are areas where artificial intelligence could replace the human grain inspector. Machine vision systems using appropriate pattern recognition techniques are an objective alternative to subjective human interpretation. The mathematical modelling of biological components in the human visual and cognitive systems is the underlying principle when creating automated grading and inspection devices. In this thesis, Zahn and Roskies Fourier descriptors are used to quantify the shape of barley seed contours and act as feature inputs to a multi-layer neural network in two pattern recognition experiments - variety discrimination and contour grading.
The first experiment involved attempts at distinguishing between three different varieties of barley - CDC Guardian, Harrington, and TR118. Eighty seed samples from each variety are used for pattern recognition training and testing. Twenty Zahn and Roskies Fourier descriptor harmonic amplitudes from each seed type are used for training. Simulation results and classification accuracies for all the simulations investigated are presented. The best overall recognition accuracy among the test samples was 80.4% using a neural structure of 20 inputs, 25 neurons in the two hidden layers, and a single output neuron. Statistical analysis showed that only the second, fourth, and sixth harmonic amplitudes were stable. The simulations were repeated using only these three features. The best overall recognition accuracy fell to 77.1% using a neural structure of 3 inputs, 25 neurons in the two hidden layers, and a single output neuron. The problems associated with the low recognition accuracy are discussed.
The second experiment involved using the sensitivity of the Fourier descriptors to distinguish between "good" and "poor" barley contours. A summed square error distance measure between the Fourier spectra of an ideal template and a thousand TR118 seed samples assisted in creating training files for the neural shape discriminator. Fifty harmonic amplitudes are used as shape features. The training simulations were successful using a neural structure of 50 inputs, 25 neurons in the two hidden layers, and a single output neuron. Grading accuracy is 87.0%. Assigned grades are presented and discussed. Problem contours are also examined in detail.
Description
Keywords
Citation
Degree
Master of Science (M.Sc.)
Department
Electrical Engineering