A NEURAL NETWORK BASED MACHINE VISION SYSTEM FOR IDENTIFICATION OF MALTING BARLEY VARIETIES

Abstract

The main demand of the food industries for barley is as a source of malt and malt extract. The major use of malt is, of course, in the brewing of beers, whisky and in the production of distilled spirits of various kinds. The quality of the product is directly related with the variety type. Generally, the main reason for identifying variety is therefore to ensure appropriate quality type. At present, identification of variety ranges from visual inspection by experienced operator to complex biochemical and molecular tests. These methods, therefore either involve a human judgment which might be biased by a number of external factors leading to false identification or require a substantial amount of time which precludes them from "on-the-spot"use. In this context, a neural network based machine vision system could bean effective and reliable means for varietal identification. The underlying principle in this artificial intelligence system is to emulate some of the capabilities of human sensory organs and the cognitive faculty,the brain. A machine vision setup performs as the human eyes perceiving the properties of the objects, and a neural network makes decisions on the basis of the properties emulating,to certain extent, the decision making process involved in the human brain. The barley varieties under consideration are Manley, Tr118, Tankard and Creme. Two hundred of samples from each variety were selected randomly and used as training samples. Seventeen properties-Perimeter, Area, Major axis, Minor axis, Length, Width, Roundness, Minimum radius, Maximum radius, Average radius, Axis ratio, Perimeter invariant, Radius ratio, Length ratio, Box-area, Box-area ratio and Mean interior gray level were measured for each sample using machine vision setup. These quantitative measurements were used as the inputs to train the artificial neural network. The trained neural network was tested with four hundred new samples,one hundred samples from each variety.A total of 48 different neural networks were trained and tested to achieve optimum overall recognition accuracy. Based on single seed observation, the highest overall recognition accuracy of 84.75% is achieved using a neural network comprising of 17 inputs,6 nodes in the first hidden layer,25 nodes in the second hidden layer and single output node, and a statistical analysis of the network outputs shows that a variety can be identified presenting only 90 seeds to the network with a probability of 99.99%success. To minimize cost, space and time, redundancy on features was measured. Eight hundred samples from each of the four varieties were randomly selected and correlation among 17 features was calculated using Eigen system Analysis. The possible reduction in dimension of input space from 17 to 7 while maintaining most of the spatial information among the points was found. Probability of error (POE) and Average Correlation Coefficient Technique (ACCT)were then used to select the best 7 of 17 features. Those 7 features are Minor axis, Box-area ratio, Maximum radius, Mean interior gray level, Major axis, Radius ratio and Length. Based on single seed observation, the highest overall recognition accuracy of 81.25% for individual seeds is achieved with a neural network consisting of 7 inputs, 6 nodes in the first hidden layer, 13 nodes in the second hidden layer and single output node.