植基於GPU及CPU之快速大面積瑕疵檢測

Fast Large Scale Defect Detection Based on GPU and CPU Cooperated Architecture

10.29768/JNTUT.201106.0001

陳金聖(Chin-Sheng Chen)；楊勝任(Sheng-Ren Yang)

大面積瑕疵檢測 ； GPU ； CUDA ； 影像前處理 ； 物件偵測 ； 以物件為基礎之階層式群聚法 ； fast large scale defect detection ； GPU ； CUDA ； image preprocessing ； object detection ； hierarchical object clustering

臺北科技大學學報

44卷1期（2011 / 06 / 01）

1 - 19

繁體中文

本文開發一套圖形處理單元(GPU)與中央處理單元(CPU)整合運用之高效能大面積瑕疵檢測系統，GPU採用Nvidia之CUDA，CPU採用Intel之Pentium處理器。著眼於CUDA之高速平行處理能力，首先利用CUDA技術來進行演算法中大量運算的前處理部份，然後，將不適合平行處理部份，在本文中為以物件為基礎之階層式群聚法於CPU端建構，透過CPU與GPU的相互合作，發展出適合大面積檢測之視覺演算法。最後，利用AOI檢測機台所取得之抗反射玻璃待測影像進行測試，可明顯看出此技術應用在大面積檢測之優勢。本文整合演算法設計、CPU與GPU適用性分析、以及軟體設計，完成植基於GPU及CPU之快速大面積瑕疵檢測演算法之開發。

This paper develops integrated fast large scale defect detection algorithms based on Graphic Processor Unit (GPU) and Central Processor Unit (CPU) cooperated architecture. The platforms are Nvidia CUDA (Compute Unified Device Architecture) and Intel Pentium processor corresponding to GPU and CPU, respectively. The large scale defect detection involves three stages: image preprocessing, object detection and object clustering. Since the high speed parallel processing capabilities of CUDA, the image preprocessing algorithms, which spent much computation time, are implemented in CUDA. The object detection and object clustering are not suitable be implemented with parallel processing because they are constructed with many reasoning and sequential steps. Hence, the object detection and a proposed hierarchical object clustering are implemented in CPU to detect the defects. Finally, the experiments with an automatic optical inspected anti-refection glass are provided to verify the performance of our proposed architecture. The experimental results reveal that the accuracy is almost sustained and the computation time is dramatically decreased by the GPU and CPU cooperated architecture.