Aiming at online impurity detection of ampoule injection, a system based on machine vision is developed and analyzed. The hardware structures, configuration of imaging card and so on are discussed. According to the requirements of application, we designed a reasonable imaging method for the detection. Moreover, the way of high spinning-emergency stop was adopted to successfully distinguish the impurity from the blemish. With regard to the software, an oriental fast and effective way of segmentation is proposed as to the low level gray of the impurity and an effective way to detect impurity based on machine vision is achieved. At last, the experiment result was given, which proves the feasibility of the program.