<p>Aiming at the current problem of Rosa roxburghii processing, which relies on manual work to remove the persistent calyx, with low efficiency and insufficient precision, this study proposes an automated positioning method of the persistent calyx of Rosa roxburghii based on an improved target detection model, to improve the detection precision and realize the automated processing of Rosa roxburghii products. By constructing a dataset containing 1120 images of Rosa roxburghii and combining data enhancement strategies such as flipping and rotating, the persistent calyx location is divided into six types of regions. Based on the YOLO-v5s model, the system compares the performance of four attention mechanisms with four loss functions. The experimental results show that the introduction of SimAM attention mechanism features of attention while Efficient-IoU can effectively optimize the bounding box regression accuracy. The optimized model achieves a mean average accuracy of 83.1% for the calyx detection mean value when the intersection and concurrency ratio threshold is 0.5, and the batch size is 5, which is a 4.8% improvement over the base model, and a 2.5% improvement in the overall mean average accuracy. In addition, the model training time is reduced to 0.528&#xa0;h, which can balance efficiency and accuracy. The study shows that the improved model can accurately locate the persistent calyx with high confidence and no leakage detection, which provides reliable technical support for the design of automated processing equipment for Rosa roxburghii.</p>