In order to improve the detection accuracy of pre-packaged food, a defect detection system based on machine vision was designed. The detection system mainly includes image acquisition module, image processing and analysis module, output execution module and so on. The image processing method was described in detail. The image denoising model based on partial differential equation was used. The defect region was segmented by double threshold segmentation method. Finally, BP neural network was used to classify defects according to circumference, area and roundness. The feasibility and effectiveness of the method ware verified by experiments. The experimental results show that the overall omission rate is 0.17% and the detection accuracy is relatively high. The detection time of each package is about 70 milliseconds, so the detection efficiency is relatively high. The system can well meet the real-time, rapid, accurate and stable testing requirements of food packaging.