Rough sets theory as a pattern classification tool for quality assessment of edible beans

Quality assessment of edible beans is a critical step before commercializat ion. The quality of edible beans was evaluated by computer vision using rou gh sets theory as a pattern classification tool. The rough sets theory is a new concept for automatic knowledge induction. This knowledge is expressed in terms of "if... then" rules, and is used for the identification of unkn own test patterns. A rough sets classifier was proposed with two different discretization approaches, the uniform and the equal-frequency intervals, f or the discrimination between acceptable, broken, damaged, small, foreign b eans, and stones. The classification results were compared to a statistical classification method based on discriminant analysis. Each object was desc ribed by 35 morphometrical features extracted from color images. The rough sets classifier provided 99.60% and 90.39% correct recognition of the train ing and test sets, respectively. The discriminant analysis provided 95.19% and 84.70% for the same data sets. Simulations showed that the performance of the rough sets classifier greatly depended on the number of discretizati on intervals. It was found that the rough sets classifier like other patter n recognition approaches, was vulnerable to the "over-fitting" problem when too many discretization intervals were introduced.