Abstract: To identify foreign matters in tobacco strips more accurately, a foreign matter identifying method based on object saliency self-supervised learning was proposed by combining deep learning image processing with clustering algorithm. First, a deep learning network is used to identify salient objects in tobacco strip images, and then clustering analysis is performed on the features of the identified salient objects, and then tobacco strips are removed from the images. Finally, the state accumulation detection method based on time series is used to authenticate the identified foreign matters. The results showed that the average IoU (Intersection Over Union) and MAE (Mean Absolute Error) of the established two-stage U-Net model were 0.90 and 0.054, respectively, which were superior to those of the BASNet and U-Net models. The average identifying accuracy was 96.6%. The time needed for processing a single image was 21 ms, which met the requirements of real-time detection. This method improves the classification and detection efficiency of foreign matters in tobacco strips.