本研究使用Solomon AccuPick 3D機器視覺軟體建置青花菜成熟度辨識模型,演算法以Mark R-CNN作為訓練基底,以不同圖資數量(600張、300張、150張)及不同訓練方式(一般訓練、資料擴增)建置6種模型,使用模型評估指標比較模型間之差異。試驗結果顯示,6種模型具有相近的辨識表現,其中600張圖資訓練之模型皆有較高的召回率(0.81),可正確辨識青花菜不同成熟度的數量較多,惟精確率相對較低,於0.81-0.83之間;150張圖資訓練之模型有較低的召回率(0.69-0.71),但精準率相對較高,於0.89-0.91之間。經綜合評估該軟體在一般訓練和資料擴增兩者間之模型表現上並無顯著差異,以Model A(600/一般)之mAP=0.77和Model A^+(600/擴增)之mAP=0.76兩種模型有較佳的影像辨識結果,青花菜成熟度辨識模型將搭載至青花菜採摘機構,以實現自動選擇青花菜之採摘作業。
This research uses Solomon AccuPick 3D machine vision software to build a broccoli maturity identification model. The algorithm uses Mark R-CNN as the training base. We use different numbers of images (600, 300, 150) and different training methods (generally training, data augmentation) to build 6 models, and use model evaluation metrics to compare differences between models. The results show that 6 models have similar identification performance, and the models trained with 600 images have a high recall rate (0.81), which can correctly identify a large number of different maturity levels of broccoli, but the precision is relatively low, between 0.81 and 0.83; the models trained with 150 images have a low recall (0.67 and 0.71), but a relatively high precision , between 0.89 and 0.91. Under the comprehensive evaluation, the software has no significant difference in model performance between general training and data augmentation. Two models with mAP = 0.77 for Model A (600/general) and mAP = 0.76 for Model A^+ (600/ augmentation) have better image recognition results. The broccoli maturity identification will be loaded into the broccoli picking mechanism to realize the automatic selection of broccoli picking.