何檢測：血糖酵素試片瑕疵檢測

Abstract

血糖酵素試片是糖尿病管理的重要部分，然而在網版印刷方式的印製過程中可能會出現錯誤，例如血糖酵素試片的導電碳墨溢出或破損，可能導致檢測結果的不準確性，進而影響患者的健康管理。當前製程中會於流水線的過程中安排人工檢測血糖酵素試片，其成本不容小覷，本研究旨在開發能應用於工業場景的血糖酵素試片碳墨瑕疵檢測方法。

由於工業製程中的異常樣品難以獲得，故傳統監督式檢測方法並不適用。本論文提出的HeInspect是基於Mousakhan, Arian and Brox, Thomas and Tayyub, Jawad於2023所發表的DDAD (Anomaly Detection with Conditioned Denoising Diffusion Models)採用無監督的異常檢測架構[11]。考量其龐大的計算成本，本論文利用自編碼器等技術對影像進行特徵提取和重建，通過比較重建影像與真實影像誤差實現對異常樣品的檢測，優化其影像重建的檢測框架，以利實際場景之應用。

Blood glucose enzyme test strips are an integral component of diabetes management. However, errors may occur during the printing process of screen printing, such as the overflow or damage of the conductive carbon ink on blood glucose enzyme test strips. These errors could lead to inaccuracies in measurement results, thus affecting patient health management. Currently, the process involves manual inspection of blood glucose enzyme test strips on the production line, incurring significant costs. This study aims to develop a method for detecting carbon ink defects on blood glucose enzyme test strips that can be applied in industrial settings.

Due to the difficulty in obtaining abnormal samples in industrial processes, traditional supervised detection methods are not applicable. Our proposed method, HeInspect, is based on the unsupervised anomaly detection framework introduced in DDAD (Anomaly Detection with Conditioned Denoising Diffusion Models) by Mousakhan, Arian and Brox, Thomas and Tayyub, Jawad [8]. Considering its high computational cost, this paper utilizes techniques such as autoencoders for feature extraction and reconstruction of images. Detection of abnormal samples is achieved by comparing the error between reconstructed and original images, thereby optimizing the detection framework for image reconstruction and facilitating its application in practical scenarios.