Abstract: Abstract: Rice has been widely planted in tropical Asia, where South China is the main rice-producing area. Rice production has accounted for about 50% of the total grain in China. The phenotype of rice can greatly contribute to clarifying the physical, physiological, and biochemical characteristics, thereby effectively promoting the rice varieties and yield. However, the morphology of a single tiller can usually be characterized by the contact destructive measurement in traditional rice cultivation, indicating a cost-consuming, labor-intensive and subjective task. In this study, a rice key point detection was proposed to rapidly extract the parameters of rice phenotype using a Stacked Hourglass Network (SHN) of computer vision. Firstly, a total of 1 080 RGB images of rice were collected. Some operations were then selected to preprocess the raw images of single tiller rice with the monochromatic background, including clipping, gamma correction, and Laplacian sharpening convolution. The data was also labeled to form the supervised dataset, according to the preset key point number of each rice component. Secondly, an SHN was adapted to detect the rice key points with the fixed and unfixed leaf number. A block of hourglass was then formed to combine the sampling, the maximum pooling, and the residual together. After that, intermediate supervision was adapted to generate the prediction from all connected hourglass blocks. As such, a multi-scale fusion was contributed to a better prediction from different scales. Two ways were also proposed to add the blank data for the unfixed leaf number. Finally, an optimal parametric combination was achieved in the experiments. The results showed that the accuracy rates reached 85.84% and 82.09% in the training and test set, respectively, when the epoch was equaled to 119 in the key point regression task with unfixed leaf numbers. The highest prediction accuracy reached 96.48% in the case of the fixed number of leaves. Correspondingly, the predicted key points of single tiller rice were connected to form a plant skeleton, according to the semantic information, indicating a better performance for the actual growth of plants. Furthermore, five parameters were selected to verify the model, including the stem length, leaf length, ear length, leaf stem angle, and stem node position. It was found that the Root Mean Square Errors (RMSE) were 5.82 cm, 3.09 cm, 1.71 cm, 3.22°, and 2.035 6 cm, respectively, indicating a better matching for the actual growth of rice. More importantly, the error of phenotypic parameters and some damage to plants were significantly reduced, compared with the previous manual measurement. The finding can also provide a new idea to efficiently extract the plant skeleton for the phenotypic parameters in rice production.