Pear blossom recognition is an important branch of agricultural intelligence, and models with higher recognition accuracy are more helpful for the automatic pollination of robotic arms in agriculture. In pear blossom recognition, it is important to reduce the leakage rate of flowers to improve pollination accuracy. This paper is based on the improvement of the YOLOv9 target detection algorithm and proposes the idea of organically combining the RCS-OSA attention mechanism with the YOLOv9 model. It also introduces the MPDIoU loss function to further optimize the bounding box regression loss in order to detect more occluded targets, which improves computational efficiency while ensuring that the target detection model maintains a high degree of accuracy. Comprehensive experiments carried out on a self-collected dataset validated the efficacy of the proposed method. The improved YOLOv9_c model attained a mean Average Precision (mAP) of 95.276% and a mean Average Recall (mAR) of 89.973%. These results signify notable enhancements, with mAP increasing by 0.696% and mAR by 0.133%, respectively, when compared to the baseline model. Additionally, the improved YOLOv9_e model attained a mAP of 95.454% and a mAR of 90.512%, with enhancements of 0.315% and 0.177%.