- 學位論文
改良式行程編碼之膨脹與侵蝕運算應用於太陽能晶片影像定位之研究
Improved Dilation and Erosion Algorithms Based on Run-Length Encoding for Solar Cell Alignment
摘要
矽晶太陽能晶片在製程上有數個階段，在過程中會有旋轉或平移的情形發生，由於太陽能晶片本身單薄且易碎裂，所以不宜使用傳統機構方式來定位。因此，透過電腦視覺處理技術與程序，在影像中估測太陽能晶片的位移量與旋轉量，以利於後續加工或檢測等相關處理。 本論文研究太陽能晶片影像定位系統之架構主要可分成三大部分；分別為：教導階段（Teaching Phase）、檢測階段（Inspection Phase）及定位參數估測階段（Alignment Phase）。在教導階段中，首先將所要學習之樣本物件以ROI（Region of Interest）方式選定，接著使用Otsu’s演算法計算適合的二值化閥值，在影像二值化的過程中同時做行程編碼（Run-Length Encoding），然後使用本論文所提出的改良式行程編碼之膨脹與侵蝕運算，來加速修補物件的完整性，接著使用輪廓追蹤技術擷取樣本物件的輪廓，先轉換為至中心的歐基里德距離之一維訊號，再將此一維輪廓訊號轉換為傅立葉描述子。在檢測階段中，首先執行與教導階段相同之影像前處理，之後，接著使用快速物件偵測法，將太陽能晶片切割出為目標物件，然後將目標物件做輪廓特徵擷取與描述，亦可得到傅立葉描述子。在定位參數估測階段中，首先從樣本物件與目標物件的中心座標，估測出太陽能晶片的位移量，接著基於傅立葉描述子，應用相位移旋轉角度之估測技術，可求得樣本物件與目標物件的輪廓訊號序列之平移點數，然後再以最小平方法與線性內插法，估測出精確的太陽能晶片之旋轉量。 最後，使用不同太陽能晶片於二種製程上具有各種不同位移及旋轉的影像；來實驗演算法則的性能，並與坊間商用軟體進行測試與比較，以驗證本論文所提的影像定位系統之精度、可靠度與執行速度。
並列摘要
The solar cell is a solid state device that converts the energy of sunlight directly into electricity by the photovoltaic effect. In general, solar cell has different manufacturing processes that will yield translation and rotation effects. And these effects will degenerate or damage the solar cell in the posterior processes. Since the solar cell is thin and easily be broken by inappropriate mechanical operation, the precisely solar cell alignment using automatic opical inspection is a superior stragey. In this thesis, the solar cell image alignment is divided into three stages, namely: teaching phase, inspection phase and alignment phase. The teaching phase involves six steps: 1) selecting the region of interest(ROI) for learning solar cell in the reference image; 2) thresholding by Otsu’s method; 3) image binarization and run-length encoding (RLE) conversion; 4) proposing improved RLE dilation and erosion algorithms; 5) contour tracing of solar cell; 6) Fourier descriptor transformation for extracting contour. In the inspection phase, there are also six steps be involved and only the step four is different with teaching phase. In setp four, the target solar cell in the inspected image must be segment using RLE blobs under the maximum area cinstrain. In the alignment phase, the translation and rotation transformation parameters for aligning the inspected image with the reference image are estimated based on both solar cells’ poids and the phase-shift in the above images, respectively. Furthermore, the refined rotation angle of the sloar cell in the spatial domain is obtained with the phase-shifted information by using the least-squares estimation method and linear interpolation. Finally, the experiments with two different processes and a variety of translation and rotation of the solar cell are designed to prove the performance of our proposed algorithms. The indexes of alignemt accuracy and computation effieicene has been compared with two comerrical image alignment libraries, easy match of eVision and shape base operator of Halcon, to further verify our proposd algorithms owning a dramatic computation advantage and simulanteously keep the similar alignemt accuracy.