Color image analysis for cereal grain classification

Loading...
Thumbnail Image
Date
1997-08-01T00:00:00Z
Authors
Luo, Xiang Yang
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Images of individual kernels and bulk-grain samples for five grain types (Canada Western Red Spring (CWRS) wheat, Canada Western Amber Durum (CWAD) wheat, barley, rye, and oats) from 20 different growing regions across western Canada were acquired. Images of individual CWRS wheat kernels were also acquired for six damage types (broken, mildewed, grass-green/green-frosted, black-point/smudged, heated, and bin/fire-burnt). Morphological a d color features were extracted to identify different grain types and damage types (for CWRS wheat only) using statistical and neural network classification methods with different selected feature models (morphological, color, and combined). For the classification of different types of individual kernels, combining morphological and color features in the feature model improved the classification accuracies over using morphological or color features alone. A non-parametric (k-nearest neighbor) statistical classifier with a feature set of 15 morphological and 13 color featuresselected using SAS STEPDISC and DISCRIM procedures gave the best results. The average classification accuracies were 98.2, 96.9, 99.0, 98.2, and 99.0% for CWRS wheat, CWAD wheat, barley, rye, and oats, respectively, when using three different training and testing data sets. Similar classification accuracies were achieved using a neural network classifier with the same features. For the classification of damaged CWRS wheat kernels, color features were more efficient than morphological features, while combining morphological features with color features improved the classification accuracies over using color features alone. A non-parametric (k-nearest neighbor) statistical classifier with a selected feature set of 24 color and 4 morphological features gave the classification accuracies of 92.5 (healthy), 90.3 (broken), 98.6 (mildewed), 99.0 (grass-green/green-frosted), 99.1 (black-point/smudged), 97.5 (heated), and 100.0 (bin/fire-burnt)%, when using three different training and testing data sets. Similar classification results were obtained using a neural network classifier with the same features. For the classification of bulk-grain samples, a selected feature set of 8 color features was used with parametric and non-parametric statistical classifiers, and a neural network classifier. When tested on three different training and testing data sets, set1, set2, and set3, all the tested bulk sample images were correctly classified by the non-parametric classifier, while 5 out of 21 bulk images of CWAD wheat in set 2 were mis-classified as CWRS wheat by the parametric classifier and 3 out of 21 images of CWAD wheat in set 1 were mis-classified as barley by the neural network classifier. For the classification of bulk CWRS wheat samples from three grades (grade 1, 2, and 3), a selected feature set of 20 color features was used with parametric and non-parametric statistical classifiers, and a neural network classifier. When tested on three different training and testing data sets, the neural network classifier gave the best results with 81.0, 67.7, and 82.5% average classification accuracies for bulk CWRS wheat samples of grade 1, 2, and 3, respectively. However, the classification accuracies varied significantly (23.8% for grade 1, 36.5% for grade 2, and 47.6% for grade 3) with different training and testing data sets, indicating that the color features extracted from bulk-wheat images did not carry sufficient information for differentiating different wheat grades.
Description
Keywords
Citation