在實務上,自動視覺檢測(Automatic Optical Inspection, AOI)設備在使用上的瓶頸為必需要搭配檢測對象來調整檢測環境、檢測方法與硬體設備相關參數。所以,本研究提出共用平臺來減少使用者重複開發與學習整合的時間;同時,讓AOI演算法有共通介面以加速開發。另外,在工廠實際環境或是製造過程中擷取影像時,很難獲得與實驗室同樣良好的影像品質,導致影像前處理部分,影像強化方法扮演重要角色。然而,在實際應用上,影像強化方法的選擇通常是透過試誤法或是經驗法則來實現。因此,由開發系統的角度來看,影像強化自動化將對於後續影像檢測有很大助益。綜合以上所述,本研究將建立一個自動化選取影像強化方法的共用平臺。 在系統開發上,本研究使用(Dynamic Link Library, DLL)進行不同影像函式庫的整合以建立共用平臺,並搭配多執行緒處理進行加速,及提出以二元搜尋樹搭配變異數為單一指標以縮減演算時間,改善後速度比傳統快2.84倍,儲存空間節省6.9倍。在自動化影像強化方法選取上,本研究提出特徵子集的架構,首先區分新輸入影像是否在系統中已存在相似特徵子集,本研究提出使用奇異值分解(Singular Value Decomposition, SVD)萃取影像資料,並保留特定奇異值還原出特徵矩陣進行分類,最後建立影像強化方法自動選擇程序。在選擇程序中判斷最佳影像強化方法的指標,使用結構相似指數(Structural SIMilarity, SSIM)。透過文獻中的75張影像驗證發現,本研究自動選擇的影像強化方法40%與文獻一致。另外60%以本研究的方法較文獻的Entropy值平均增加27.93%。研究結果顯示本系統能有效提高影像品質,而不至於在增強影像品質的同時產生過多雜訊。在實務上也因為非監督式學習不需要訓練的特性提供使用者可以快速更迭新樣本的彈性。
In practice, the bottleneck of automatic optical inspection (AOI) equipment is to adjust the environment, methods and hardware. Therefore, this study suggests a common platform that reduces the duplication of development and the learning process. In addition, it is usually difficult to grab good images in a manufacturing process; and therefore, the image enhancement methods used in the pre-processing phase play an important role. However, the image enhancement method is usually selected by trial-and-error or experience. So the present study is to establish a common platform for selecting automatic image enhancement method. For system development, this study used Dynamic Link Library (DLL) to integrate different libraries. And created a common platform with multi-thread processing for acceleration. Then, presented a binary search tree through variance, it reduces the calculation time. After improvement, the speed was 2.84 times faster than traditional ones, and resulted in storage space savings of 6.9 times. On automatic selection of image enhancement, the research proposed an architecture of feature subsets. First, to check if a newly input image is the same as any subset that already exists in the system. If similar features exist, an image enhancement method can be determined through the pre-stored subset information. Next, the research proposed an efficient feature matrix based on singular value decomposition (SVD) to represent an image. Finally, the structural similarity index (SSIM) was used to select the optimal enhancement. To verify the procedures, 75 images from literature were used in the experiment. Forty percent of the automatically selected enhancement methods were consistent with the literature. Another 60% using different enhancement method had an increased entropy value of 27.93%. The study results implied that the system could effectively improve the image quality and not causing over enhancement to produce noise, and the non-supervised learning nature allows users can quickly apply to new samples.