近年隨著機器學習、電腦視覺以及深度學習領域日新月異，提升自動光學檢測的效能與效率，各種產業紛紛嘗試使用自動光學檢測系統取代檢測員來進行產品上的瑕疵偵測。本研究主要目的為協助針織布廠商處理在針織布上進行自動光學檢測的過程中遭遇的困境，我們結合深度學習模型與經驗法則運算，設計一套能正確分析針織布影像上是否存在瑕疵之方法，解決廠商希望引進新檢測技術卻無法順利執行之問題，以取代傳統的肉眼檢測方式。 本研究設計從現場廠域中即時蒐集影像資料，並對影像進行標記，作為瑕疵偵測方法之訓練資料集。我們提出一套剖布線取樣方法，透過影像分割模型以及經驗法則運算，分析針織布中良品資料的特徵，並執行剖布線抽樣作業，進行影像挑選並製作成訓練資料集。接著我們提出另外一套瑕疵偵測方法，透過影像分割模型以及門檻值的設定，將每一張影像依照紋理進行判別，最後再分析連續影像之間的時序性關聯，執行投票運算，計算出輸入的連續影像是否為瑕疵品，作為系統輸出。後續能將系統偵測的瑕疵結果輸入到機台上的警報裝置，即時通知現場作業人員，維護針織布機台。;Since machine learning, computer vision and deep learning have been developed quickly these years, the efficiency and effectiveness of AOI have been improved a lot. Recently, inspection in many industries has been replaced inspectors with AOI system. The main purpose of our research is trying to deal with problems which knitted fabric manufacturers encountered when deploying AOI system for inspection. We have developed a system which is capable to detect defects on knitted fabric correctly with combination of deep learning models and rule -based calculation, so that the manufactures are able to do inspection with AOI system rather than inspectors. We plan to collect and label data on spot. After processing, those data will become training data of defect detection method. We proposed a “Cutline Sampling Method” to sample Cutline images by analyzing features with semantic segmentation model and rule-based calculation. We also proposed a “Rule-based Defect Detection Method”. Classification of each image will be done by semantic segmentation model and thresholds. Every single image classified result will be sent to time series voting. According to the relation between adjacent images, the system will distinguish the consecutive input images between “Good Product” and “Defective Product” as output. The output could be transmitted to alert system on knitted fabric weaving machine and inform workers so that they are able to maintain the weaving machine immediately.
