Abstract: Aiming at the current lack of online detection of impurity rate of domestic combine harvesters, a segmentation method for rice stems based on improved mask R-CNN was proposed. According to the shape and location characteristics of the stalk impurities, the network layer in the original Mask R-CNN was optimized. The image augmentation was introduced to expand the image samples and solve the lack of image training data. The trained model was used to segment the images in the verification set and compare them with the original Mask R-CNN and other algorithms. The results showed that the comprehensive evaluation index（F1） of the improved Mask R-CNN algorithm reaches 91.12%, which is better than other models. The segmentation time can reach 3.57 s, which proves that the method can meet the requirements of real-time detection and provides a technical reference for the subsequent implementation of an online detection system of impurity rate.