A method based on deep learning for automatic segmentation and characterization of fruits defects

L. BUGNON, D.H. MILONE, G. STEGMAYER

2020-06-01

2020-12-01

Alimentos

Currently, significant food waste is generated in the world due to logistics costs and lostopportunities in low-profit markets. Artificial intelligence, and in particular machinelearning, can improve batch characterization processes, particularly of edible fruits.Through these techniques it is possible to automatically analyze fruit production fromimages, and thus reduce the cost and risks in quality management. With thedevelopment of smarter methods, it is possible to reduce the cost of adapting tools tonew markets and applications, and accelerate adaptation by new customers. The aimof this development is to present novel machine learning models based on deepnetworks to automatically characterize fruits based on a batch-based photographicregistration device. The developed system is composed of two stages. The first one isthe segmentation stage, which individualizes each fruit in the batch. The second one isa defect classifier, which is needed to identify the type of defect on each fruit, such asdecay, immaturity or insect damage. In addition, the proposed algorithms areself-explanatory, thus the person in charge of the production can verify which details inthe images are the ones that the deep neural network has focused on to make itsprediction. Finally, results on segmentation and classification, which correspond to thedelivered source code, are summarized.