|
Currently, inspection of wheat in the United States for grade and class is performed by human visual analysis. This is a
time consuming operation typically taking several minutes for each sample. Digital imaging research has addressed this
issue over the past two decades, with success in recognition of differing wheat classes, and distinguishing wheat from
non-wheat species. Detection of wheat kernel defects, either by damage or disease, has been a greater challenge. A study
has been undertaken that uses high-speed black and white imaging at 10-bit photometric resolution to detect damaged
kernels one kernel at a time. The system, composed of hardware (camera, lighting, power supplies, and data acquisition
card), software (LabVIEW and MATLAB), and analytical (MATLAB and SAS) components, is designed to a) capture
images of free-falling kernels at opposing angles through the use of optical grade mirrors, b) parameterize the images
and, c) perform classification. The system operates with a 1/30,000 second exposure time though with restrictions on
image transfer rate (60 Hz) and image processing routines for feature extraction (currently conducted offline). Fifty
samples of hard red and white wheat subjected to weather related damage during plant development were used in this
study. Parametric (linear discriminant analysis) and non-parametric (k-nearest neighbor) classification models were
tested to determine the image features that best foster recognition of the damage conditions of mold, sprout, and black
tip. The morphological features used in classification included area, projected volume, perimeter, elliptical eccentricity,
and major and minor axis lengths. Textural features from calculated gray level co-occurrence matrices (including
contrast, correlation, energy, and homogeneity), are also under consideration though not reported herein. So far, our
results indicate that with as few as three image parameters, classification (damaged vs. sound) levels approach 85 to 90
percent accuracy. Information learned from this study is intended to lead to the streamlining of feature extraction in
image-based high speed sorting.
|
