ABSTRACT In food industry, removing the nonedible parts causes reduction of production efficiency due to its heavily labor‐intensive process with a lot of loss of the edible parts. Challenges for automation of removing the cabbage stem end process, namely nonedible parts, is attributed to its varied shapes and sizes. To overcome, one‐stage YOLOv5 deep learning model with encoded depth class transform process was developed for accurately detect cabbage stem core region and classify encoded depth classes. In addition, PSO optimization was implemented to enhance performance of detection model by tuning hyperparameters. To evaluate detection performance, four variances of the YOLOv5 deep learning model and a two‐stage Faster R‐CNN model were implemented. The proposed approach demonstrated that the YOLOv5L variant achieved mean average precision (mAP) and intersection over union (IoU) scores of 0.99 and 0.89, respectively, whereas the YOLOv5S variant obtained mAP and IoU scores of 0.76 and 0.84, respectively. In addition, YOLOv5L variant showed highest F 1‐score for classifying encoded depth classes over 99% where YOLOv5S variant showed 82.5% F 1‐score which suggesting high detection and classification performance. This study provides important technical process to detect nonedible parts and depth to be removed during preprocessing of food manufacturing process. In future, the proposed optimal cabbage stem detection model can apply by embedded system to real‐time automated manufacturing process.
Kim et al. (Fri,) studied this question.