Application of fuzzy C-Means clustering for seed discrimination by artificial vision

The Fuzzy C-Means Algorithm (FCMA) was applied for the discrimination between seed species by artificial vision. Colour images of seeds belonging to 4 species were acquired with a CCD camera. In order to characterise the morphology of the seeds, a set of quantitative features were extracted from the images. The aim of this study was to cluster, by FCMA, the measured learning and test data into 4 groups. The FCMA process was improved by the introduction of a non-random initialisation of the cluster centres. In addition, the Mahalanobis distance was used, instead of the Euclidean distance, as a measure of the proximity of a pattern to a cluster. Furthermore, a classification approach with a reject option was investigated for increasing the classification performances. This was achieved by assigning the seeds which were lying in the fuzzy boundaries between the available classes to a reject class. The proposed initialisation method outperformed the random initialisation both in terms of the computation time and the percentage of correct recognition. With the Mahalanobis distance, the error of classification was 5.32% and 6% for the training and test sets. In comparison to the Euclidean distance, the Mahalanobis distance allowed a decrease of the classification errors by 1.12% and 1% for the training and test sets, respectively. The main advantage of the Mahalanobis distance is that it takes into account the real underlying shapes of the clusters. The application of FCMA with a reject option showed that the classification errors dropped to 3.38% and 5% when 7.43% of the learning seeds were rejected.