Abstract: During the industrial production of pastries, foreign substances such as plastic and rubber can accidentally enter the processing chain, posing serious risks to consumer health and safety. Therefore, detecting foreign substances in pastries is a critical quality control step. X-ray computed tomography (CT) is a fast, non-contact, and non-destructive testing method that is widely used in quality inspection processes on of industrial pastry production lines. However, owing to the high-throughput detection requirements of such production lines, the analysis of a single product typically needs to be completed within 1 s. This limited time frame makes it impossible to capture a sufficient number of projection images, restricting the use of conventional CT methods. In this study, we propose a foreign-object detection method based on the U-Net network, which is trained using CT data from the same type of samples and foreign objects. The experimental results show that this method requires only a few projection images to accurately identify multiple foreign objects. It can quickly and efficiently detect foreign objects from CT data on industrial production lines, greatly improving detection efficiency in the pastry industry.