- 學位論文
基於統一計算架構的單內核多執行緒運算模型之即時表面瑕疵檢測
Real-Time Surface Defect Inspection based on Single Kernel Multiple Threads Computing Model of Compute Unified Device Architecture
摘要
在產品的檢驗過程中，表面瑕疵檢測技術佔有非常重要的地位；人工檢測方式耗時且費力，不但容易因不同的檢測標準，造成主觀的判斷錯誤，且需要較高的人力成本。隨著機器視覺、影像處理、與高效能運算技術的快速發展，結合這些技術的應用領域愈來越廣泛，其中一個主要的產業應用，係自動化表面瑕疵檢測。 本論文使用統一計算架構(Compute Unified Device Architecture，CUDA)技術，以針對大面積的待測物表面，達到高速、高精度之瑕疵檢測需求。CUDA係整合中央處理器(Central Processing Unit，CPU)與圖形處理器(Graphics Processing Unit，GPU)的異構式運算平台及程式設計模型。本論文提出之方法，係以GPU容許的最大執行緒數量為基準，將待測物影像劃分為不同區塊，依序由GPU以大量執行緒，同時進行單一影像區塊內，瑕疵的邊緣與個數之判定；另一方面，由CPU進行跨越相鄰影像區塊的瑕疵之判定，從而獲得在待測物影像中正確的瑕疵數量。實驗結果顯示，本論文開發之演算法，可在小於1秒的時間內，在具有2.4576×〖10〗^8個像素的待測物影像中，獲得正確的瑕疵邊緣與數量。
並列摘要
Techniques of surface defect detection play an important role in the product quality control. Human visual inspection is time-consuming, error-prone, and labor intensive. Due to the rapid development in the fields of machine vision, image processing, and high performance computing, a variety of different applications can benefit from the combination of technologies in these fields. A primary industry application is the automatic surface defect detection. In the thesis, for objects with a large surface area, the CUDA (Compute Unified Device Architecture) technology is adopted to satisfy both the high-speed and high-precision requirements on the surface defect detection. The CUDA is a heterogeneous computing platform and programming model that integrates the CPU (Central Processing Unit) and GPU (Graphics Processing Unit) components. In the proposed method, based on a GPU’s maximum allowable threads, the image of an object is divided into different image blocks. The image blocks are successively processed by the GPU, via concurrent threads, to determine the edges and number of defects in each single image block. On the other hand, the CPU determines the number of defects that spans adjacent image blocks, in order to obtain the correct number of defects in the entire image. Experimental results show that the algorithms developed in the thesis can accurately obtain the edges and number of defects inside an image, containing 2.4576×〖10〗^8 pixels, in less than one second.