ABSTRACT
Artificial vision was applied to the discrimination between 4 seed species (2 cultivated and 2 adventitious seed species). The main scope of this investigation was to study the strengths and the limitations of a fuzzy clustering technique, which was the fuzzy c-means algorithm (FCMA). Since the performances of FCMA depend on the initial configuration of the cluster centers, this study reports on a method for carrying out fuzzy classification with a nonrandom initialization of the cluster centers. A set of 58 size, shape, and texture features were extracted from color images in order to characterize each seed. The data was reduced by applying a principal component analysis and selecting the first 10 components. The first 10 principal components represented 93.16 percent of the total sum of squares. FCMA was applied, with the Euclidean distance, for the classification of learning and test data. For the present work, simulations showed that the classification performances depended on some parameters of FCMA. It was observed that there is no need to apply FCMA with a very high number of iterations. The best classification performances were 96.85 percent and 97.89 percent of the training and the test sets, respectively. FCMA may therefore be used in an on-line device for seed discrimination. However, further improvements are to be performed such as the use of a more sophisticated distance measure instead of the Euclidean distance.
