Abstract: Table grape thinning has been commonly used to increase berry size for high-quality fruits in modern agricultural management. However, the table grape thinning in manual has also been a time-consuming and labor-intensive task in current large-scale production. Accurate and rapid identification of grape flowers, stems, and clamping points of fruit stem can contribute to the table grape thinning during automatic operation. In this study, a new method was proposed to locate the clamping points of fruit stem for table grape thinning using an improved Mask region-based convolutional neural networks (R-CNN). Firstly, an improved Mask R-CNN was established to fuse a bottom-up path augmentation using the ResNeXt network. The ResNeXt network was implemented to improve the feature extraction of flowers and stems, where the bottom-up path augmentation was to retain the low-level features. Then, a set logic was proposed, according to the prediction from the improved Mask R-CNN. The Intersection over Union (IoU) function was built to remove the duplicate targets in the detection. Nevertheless, some fruit stems were incorrectly recognized as leaf stems in the prediction, due to the lower detection accuracy of fruit stems than that of grape flowers. In doing so, the position relationship between table grapes flower and fruit stem was further filtered out from the prediction data, further to improve the detection precision. Meanwhile, the pairs of table grape flowers and fruit stems were built in the prediction dataset. The isolated pixel points in the fruit stem masks were removed by the morphological open operation, and the main fruit stem masks were calculated by the set logic relation. Eventually, the center point near the placenta of the main fruit stem was determined as the clamping point of the fruit stem, particularly for the main fruit stem bounding boxes with different aspect ratios. Six models were compared to verify the improved Mask R-CNN for detection and segmentation of the table grape flower and fruit stem using different backbone networks. The result showed that the improved Mask R-CNN worked performed best. The evaluation metrics Average Precision (AP0.5, AP0.75, AP0.5-0.95, APS) on the coco dataset increased by 2.3, 5.4, 6.0 and 23 percentage points, respectively, and the MIoU increased by 7.4 percentage points, compared with the ResNet50-based Mask R-CNN model. The ResNeXt101 in the backbone of the improved Mask R-CNN effectively improved the performance of the model, while the fusion of the path augmentation network effectively enhanced the detection performance of small targets. The improved Mask R-CNN performed more accuracy in some segmentation, compared with the others. Meanwhile, the speed of detection and segmentation was relatively comparable to the other models, indicating fully meeting the requirements of table grape thinning. Then, the location accuracy and speed were tested in the different conditions including the shooting weather, the number of flowers in one picture, and the growth stage of table grape flower. The images in the test set were randomly selected for the experiment, where the successful location rate of the fruit stem clamping point was 83.3%, the average location time was 0.325 s. The results showed that the location accuracy was 90% in sunny and sunlight weather, and the total time was 0.3 s. In all shooting conditions, the sunny and sunlight, single flower in one picture, and the inflorescence period can greatly contribute to the accuracy of the location. Then, the location error of the fruit stem clamping point was tested, where the maximum location error in x, y and the total location error were 10, 12, and 16 pixels, respectively, indicating fully meet the requirements of location accuracy and speed for table grape thinning.