Objective To solve the problems of low efficiency and high cost in cherry screening.Methods A heatmap regression method HRNet-YT was proposed to automatically identify cherry size and the presence of fruit stems, thereby realizing efficient screening. HRNet-YT utilized multiple parallel subnetworks to achieve multi-scale information fusion while maintaining high-resolution representations, ensuring the spatial accuracy of heatmaps for stem and calyx keypoints. By leveraging heatmap techniques to capture rich contextual information and optimizing the loss function, the model's robustness and precision were enhanced.Results HRNet-YT-W48 (384×288) achieves a detection precision of 87.3% and an keypoint average precision (AP, OKS=0.5) of 0.22 on the dataset.Conclusion The proposed method demonstrates high precision and adaptability in the cherry keypoint detection task.