The automation of post-harvest fruit grading in the industry is a problem that is receiving considerable attention in the realm of computer vision and machine learning. Classification accuracy with automated systems in this domain is a challenge given the inherent variability in the visual appearance of fruit and its quality-determining features. While the accuracy of automated systems is of paramount importance, the usability and the interpretability of machine learning solutions to the operators are also crucial since many sophisticated algorithms involve numerous tunable parameters and are often “black-boxes”. This research presents a generalizable machine learning solution that balances the need for high accuracy and usability by decomposing the problem into sub-tasks. A powerful boosting algorithm (AdaBoost.ECC) with low interpretability is employed for learning fruit-surface characteristics. The classification outputs of boosting then become inputs for rule-induction algorithms (RIPPER and FURIA), generating human-interpretable rule sets that are amenable for review and revisions by operators. Using seven datasets of different fruit varieties, the performance of the proposed method was compared against a manually calibrated commercial fruit-grading system. The results showed that the proposed system is able to match the accuracy of machines calibrated by domain experts having many years of experience, while providing simpler rule sets possessing high interpretability and usability while yielding knowledge discovery.
|Title of host publication||Transactions on Engineering Technologies|
|Subtitle of host publication||Special Issue of the World Congress on Engineering and Computer Science 2013|
|Editors||Haeng Kon Kim, Sio-Iong Ao, Mahyar A. Amouzegar|
|Number of pages||16|
|Publication status||Published - 1 Jan 2014|