應用SAM與轉移學習的U-Net模型於刀具磨耗之研究

本文提出一套創新的刀具磨耗檢測系統，結合 U-Net 語義分割與 Segment Anything Model（SAM），可自動辨識刀具磨耗位置。傳統深度學習方法仰賴大量人工標註，耗時且費力。為此，本文利用 SAM 模型自動產生遮罩，並與預訓練模型輸出比對，選取重疊度（IoU）最高的遮罩作為偽標註，實現無需人工介入的資料擴充流程。檢測系統採用影像疊圖展開技術，將端銑刀轉為全景影像，透過 U-Net 精準分割刀刃區域，並應用轉移學習以提升模型對新樣式刀具的辨識能力。實驗結果顯示，該方法在新刀具上的mIoU提升超過 35%，同時保留原有刀具的辨識效果，證明其具備應用於智慧製造之潛力。

This study proposes an innovative tool wear detection system. It integrates U-Net semantic segmentation with the Segment Anything Model (SAM) to automatically identify wear on cutting tools. Traditional deep learning approaches rely heavily on manual data annotation, which is time-consuming and labor-intensive. To address this issue, the proposed method utilizes SAM to automatically generate segmentation masks and compares them with the output of a pre-trained model. The mask with the highest Intersection over Union (IoU) is selected as a pseudo-label, enabling dataset expansion without manual intervention. The system adopts an image stitching technique to transform end mills into panoramic images, which are then processed by U-Net for accurate cutting edge segmentation. Transfer learning is applied to enhance the model 's performance on new tool types. Experimental results show that the proposed approach improves mIoU by over 35% on unseen tools. It also maintains high accuracy on known tool types, demonstrating its potential for practical applications in smart manufacturing.