On-line Hierarchical Classifier for Agricultural Sorting Systems

This paper presents an on-line hierarchical classifier for agricultural products. The classifier contains two levels: a low level, for population identification and a high level for feature selection and grade classification. The low level contains an on-line clustering algorithm that clusters the produce into distinct populations. Each population defines a different location in the feature space. When a new population appears, the system adapts to it. The high level contains n similar classification algorithms that differ by the number of features. A fuzzy system analyzes these algorithms results, outputs a decision regarding the fruit grade and changes each of the algorithms reliability level in order to determine the current best-fit classifier. Feature selection is conducted on-line according to the classified population. Using a synthetic data set, comparison of the proposed hierarchical algorithm to a kNN classifier that incorporates all features indicated a 11% improvement in classification accuracy.