近年需要視覺檢測的物品不斷推陳出新,為了因應各類型之樣本,必需開發不同類型之檢測方法,且每次為了因應各項產品開發出之方法都必須花費冗長的時間,且只能針對單一產品進行檢測,不符合經濟效益。因此,本研究針對具有重複之週期性紋路特徵的品,開發出一套自動化視覺檢測方法。 適應性相減法(Adaptive subtraction)是一套不需具備參考影像的瑕疵檢測演算法,主要是利用相關係數法快速尋找影像特徵週期,然後於傳統的相減法增加一個遮罩(Filter),用以克服檢測影像些微旋轉與位移的問題。其第一部分藉由擷取測試影像之一維灰階訊號,以發現其具有週期性變化的特徵。因此,使用相關係數法來驗證特徵求取影像週期第二部份計算兩相鄰週期之相同位置像素點之灰階差異,最後設定臨界值(Threshold)判別出瑕疵區塊。本研究主要是針對具有週期性紋路之兩類型之LCD面板、三類型之彩色濾光片、鑄件與紡織品之表面瑕疵影像進行檢測,由實驗結果可發現本方法針對此七類樣本,共268張影像,總檢測正確率能達到91.79%,且處理一張640× 450 像素之影像僅需要0.125秒。因此可以將此方法實際廣泛應用於具有週期性紋路特徵之瑕疵檢測。
The products which required automatic visualinspection are increasing dramatically through the years. It is necessary to develop different inspection methods for various type of products. In addition, every inspection method development is very time consuming; they can only inspect a single product at a time,which is very inefficient. Therefore, this research intents to develop an automatic visual inspection method for inspecting the products which contain the repeated periodic texture characteristic. The adaptive subtraction method is a defects-checker algorithm that doesn’t need golden sample image. It mainly utilizes correlation coefficient method to rapidly seek the period of image feature, and add a filter on the traditional subtraction to overcome the problems of rotating and displacing of images. First, the signal of 1-D gray-level line image is retrieved by using a product image. Secondly, a correlation coefficient method is used to check the periodicity of the periodic feature. Thirdly, the gray-level differences between the pixels at the same place of the two adjacent periodicities are calculated. Finally, by setting a threshold we can differentiate these defects. This research is mainly designed to detect the surface image defects of two types of LCD panels, three types of color filters, casting samples, and fabrics which all possess the periodic textures characteristic. And according to the experiment results of these seven samples which have total of 268 images, the accuracy rate of this method can reaches 91.79%. It can achieve a fast computation of 0.125 seconds for a 640× 450 image. Therefore, this method can be applied for detecting defects of products with periodic textures characteristic.