Accurately detecting citrus pollen vitality is crucial for cultivating seedless citrus varieties and satisfying the demand for high-quality citrus products. Developing intelligent tools for this detection is therefore of great significance. In this study, we built a pollen vitality detection dataset through manual collection and labeling, addressing challenges such as the clustering, small size, and complex background of citrus pollen particles. To tackle the issues of imbalanced pollen samples and complex backgrounds in the dataset, we improved the YOLOv11 deep neural network, introducing the YOLOv11-FS model. During detection, we replaced the EIOU loss in YOLOv11 with the Focal EIOU loss function to boost performance on imbalanced samples and combined it with Soft NMS to enhance detection box accuracy. We also leveraged the backbone and neck networks of YOLOv11-FS to boost feature extraction and key pixel region recognition, improving small target detection accuracy. Experimental results showed that our improved YOLOv11-FS model performed outstandingly in pollen detection tasks, with a pollen vitality rate error of only 0.70%. For fertile pollen detection, the recall rate, accuracy rate, and F1 score reached 98.76%, 99.67%, and 99.22%, respectively, while for sterility pollen detection, these metrics were 94.87%, 98.89%, and 96.84%, respectively. These results met the basic requirements for pollen vitality detection and provided reliable support for seedless citrus breeding. This method offers technical support for pollen vitality detection and variety improvement in intelligent citrus orchard management and can also serve as a reference for pollen vitality detection in other plants.