游檢測 : 水閥瑕疵檢測

Abstract

本論文提出游檢測: 一個結合人工智慧和傳統影像處理的水閥檢測機台，透過機台的檢測可以知道水閥產線產出的水閥是否為瑕疵品，取代用人力辨別每個水閥並找出瑕疵品的人力成本。

總共有十一種型態的水閥且正反兩面皆須檢測，對於不同水閥的每一面客製化一個專屬於它的演算法，需被檢測的瑕疵類別有黑斑、紅斑、裂痕、缺角、無反光、缺角裂痕。

水閥會先經過震動盤初步篩選正反面，然後依次放入旋轉玻璃盤，藉由上下兩站相機得到水閥正反面照片。照片首先經由傳統演算法進行降噪和去背處理，然後送入訓練好的人工智慧模型判斷是否有瑕疵，最後決定這個水閥送往瑕疵產品或好的產品或非上述兩者的出料口。

In this thesis, we propose YuInspect: a water valve inspection machine, which combines artificial intelligence (AI) and traditional image processing. Through machine inspection, we can determine whether the water valves produced by the production line are defective or not, thereby replacing the manual labor required to identify each valve and find defective products.

There are eleven types of water valves to be inspected on both sides. For each side of different water valves, an algorithm is customized for it. The defects to be inspected include black spots, red spots, cracks, chipping, no-lapping, and chipping-cracks.

The water valves are initially screened for their orientation using a vibrating plate, and then sequentially placed on a rotating glass plate. The top and bottom cameras capture photographs of both sides of the water valve. We use traditional algorithms for noise reduction and background removal before feeding photographs into a trained AI model to detect any defects. Finally, the decision is made whether to send the water valve to the defective product, good product, or undetermined output port.