In order to solve the problem of the synchronously identification on multiple scattered tobacco leaves in the intelligent classification and identification of tobacco leaves, a method for simultaneous recognition of multiple tobacco leaves position was proposed based on improved Mask R-CNN. The K-means clustering algorithm was introduced into the region proposal network of Mask R-CNN, to cluster the marked target detection frames. It realized the optimization of 5 preset anchor point sizes of sizes and 3 aspect ratio of the anchor point, and made it more in line with the distribution characteristics of tobacco leaf image data, so as to improve the accuracy of the generated suggestion box and shorten the recognition time. Based on the collected tobacco leaf image dataset, the effectiveness of the proposed method was verified. The experimental results show that when the IoU is 0.5, the single-sample time of this improved Mask R-CNN is 313 ms with the improved mAP value of 3.56% on the test set, which is faster than the 326 ms of the original Mask R-CNN. Comparing the target detection algorithms of Faster R-CNN and SSD, it also shows advantages in precision and recall rate.