Research on an Improved YOLOv11-Based Detection Method for Harvestable Safflower Filaments in Unstructured Environments.

To address the demands of intelligent safflower harvesting scenarios, which require a safflower recognition model with both low computational cost and high detection performance, this paper proposes a lightweight improved model based on YOLOv11n, termed YOLOv11n-Starnet-ADown. To reduce the model's size, the Starnet network is adopted to replace the backbone network of YOLOv11n. To enhance small-object detection capability while further reducing memory footprint, ADown is used to replace the standard convolutional downsampling module in the neck network of YOLOv11n. The YOLOv11n-Starnet-ADown model was experimentally validated on a self-constructed safflower dataset. The results show that the model achieves an overall precision of 91.6%, a recall of 87.8%, and an mAP@0.5 of 92%; the recognition accuracy for harvestable safflower filaments reaches 96.5%; and the model memory footprint is 3.66 MB, representing a 29.9% reduction compared to the baseline YOLOv11n model. Finally, the detection performance of YOLOv11n-Starnet-ADown was compared with that of four conventional models under different scenarios, confirming the effectiveness of the proposed model. The proposed model exhibits stable detection performance under diverse complex conditions, including overcast skies, occlusion, and backlighting, which adequately satisfies the fundamental requirements for safflower filament detection in real-world environments. Overall, this work offers a lightweight technical solution for the intelligent harvesting of safflower filaments in unstructured settings.