In order to detect the impurities in bottle, a real-time visual inspection system was designed. A mathematical model for double-COD stereoscopic system was set up according to geometrical imaging principle and the equation applied to get the position of the impurity from two images was deduced. Morphological wavelet was used to simplify the image. Both adaptive thresholding and edge detector were applied to separate impurities from the image, and then Feature vectors were used for feature-match between two images. When image resolution is near 4 pixel/mm,the identification accuracy is about 98% and the time of detecting a bottle is less than 200 ms.