本研究採用Mask R-CNN影像辨識演算法模型以及Detectron2框架，應用於愛文芒果影像進行五類不良品分類。主旨在於針對愛文芒果五類不同病種進行多物件實例分割影像辨識，五類病種分別為著色不良、炭疽病、乳汁吸附、機械傷害以及黑斑病，總共五類不良病種。因廠商提供之資料集(training set + validation set共計59650張)資料雜訊嚴重，在資料清洗(Data Cleaning)以及資料前處理(Data Preprocessing)階段花費超過500小時以上，以及將不良資料汰除8995張，最終資料集共為50655張。資料集蒐集不易以及資料雜訊多的情形，為人工智慧應用於真實場域所會面臨的挑戰。為求辨識精準度而非速度，此研究主要應用技術為深度神經網路影像辨識演算法Mask R-CNN以及基於COCO Dataset Pretrained Model應用遷移式學習(Transfer Learning)。再運用GrabCut演算法，在資料前處理階段使實例分割(Instance Segmentation)達到99.9%的精準度。以及應用X101-FPN骨幹，以增加神經網路深度，相較R50-PFN，使辨識精準有效提升90%，最終達成67.2之AP。

This paper used the image recognition algorithm model and Detectron2 framework to detect five types of defective Irwin mangos. The principal object was to recognize five different diseases concerning Irwin mangos by multi-object instance segmentation. There are five diseases respectively, poor coloration, anthrax, latex attached, mechanical harming and ink spot disease. We spent over 500 hours on data cleaning and data pre-processing due to the dataset (training set + validation set totally 59650 images) offered by vendor is inferior in quality. We also eliminated 8995 bad images. Finally, the dataset remained 50655 images. Data collection and data hazard are the significant challenges will face when apply AI to the real-world. Our research mainly used Mask R-CNN which is the image recognition algorithm of deep neural networks and transfer learning based on COCO Dataset Pretrained Model makes the detection results more precise. Then, we used Grabcut algorithm which makes accuracy of instance segmentation up to 99.9% in data pre-processing stage. Further, we applied X101-FPN backbone for making neural network deeper which compared to R50-FPN was effectively improved 90% accuracy. Eventually, we achieve the 67.2 AP in ours experiment.