| Abstract: | This study describes the use of colour image analysis to identify four seed varieties. A wide range of kernel measurements was obtained from digitised colour images of whole seed samples of rumex, wild oat, lucerne and vetch. The combination size, shape (including kernel seven invariant moments) and texture parameters is the major element in this investigation. Two pattern recognition approaches were attempted in the classification: stepwise discriminant analysis, which is part of statistical pattern recognition techniques, and artificial neural network. The artificial neural network was found to outperform discriminant analysis. With only three inputs, a simple three-layer perception network exhibited performances exceeding 99% both in learning and test sets. It is shown that a mixture of features improved classification from 92% for size and shape parameters to 99% for size, shape and texture parameters. Two species, totally overlapped in the morphometrical space, were well separated by texture. The best characteristics are extracted from the red channel images. Limitations of neural computing concepts are discussed with respect to seed classification. |
