FECT

Abstract: As the grade of green coffee beans largely depends on total number of defective beans in a given quantity of sample beans, removing defective beans is important for ensuring their quality and market price. On the other hand, the prevalent sorting method, i.e., manual handpicking, can be affected by human condition and is time-consuming. Studies have been proposed to classify defective fruits, vegetables, and beans by exploiting image processing and machine learning techniques so far. However, general single-label classification algorithms are not suitable for sorting coffee beans because some of them have multiple properties, e.g., broken and fade. In this study, we propose a deep neural network model with modifications to the EfficientNet-B1 by putting branches, which correspond to each defect after feature extraction layers, to classify coffee beans with multi-label. The proposed multi-branch EfficientNet-B1 model significantly improved overall performance, with an f1-score of 0.8229, compared to single EfficientNet-B1 model.

Keywords: Convolutional Neural Networks, coffee bean, EfficientNet-B1, f1-score, multi-label classification.