The accurate acquisition of safflower filament information is the prerequisite for robotic picking operations. To detect safflower filaments accurately in different illumination, branch and leaf occlusion, and weather conditions, an improved Faster R-CNN model for filaments was proposed. Due to the characteristics of safflower filaments being dense and small in the safflower images, the model selected ResNeSt-101 with residual network structure as the backbone feature extraction network to enhance the expressive power of extracted features. Then, using Region of Interest (ROI) Align improved ROI Pooling to reduce the feature errors caused by double quantization. In addition, employing the partitioning around medoids (PAM) clustering was chosen to optimize the scale and number of initial anchors of the network to improve the detection accuracy of small-sized safflower filaments. The test results showed that the mean Average Precision (mAP) of the improved Faster R-CNN reached 91.49%. Comparing with Faster R-CNN, YOLOv3, YOLOv4, YOLOv5, and YOLOv6, the improved Faster R-CNN increased the mAP by 9.52%, 2.49%, 5.95%, 3.56%, and 1.47%, respectively. The mAP of safflower filaments detection was higher than 91% on a sunny, cloudy, and overcast day, in sunlight, backlight, branch and leaf occlusion, and dense occlusion. The improved Faster R-CNN can accurately realize the detection of safflower filaments in natural environments. It can provide technical support for the recognition of small-sized crops.

Accurate safflower recognition is a critical research challenge in the field of automated safflower harvesting. The growing environment of safflowers, including factors such as variable weather conditions in unstructured environments, shooting distances, and diverse morphological characteristics, presents significant difficulties for detection. To address these challenges and enable precise safflower target recognition in complex environments, this study proposes an improved safflower detection model, WED-YOLO, based on YOLOv8n. Firstly, the original bounding box loss function is replaced with the dynamic non-monotonic focusing mechanism Wise Intersection over Union (WIoU), which enhances the model’s bounding box fitting ability and accelerates network convergence. Then, the upsampling module in the network’s neck is substituted with the more efficient and versatile dynamic upsampling module, DySample, to improve the precision of feature map upsampling. Meanwhile, the EMA attention mechanism is integrated into the C2f module of the backbone network to strengthen the model’s feature extraction capabilities. Finally, a small-target detection layer is incorporated into the detection head, enabling the model to focus on small safflower targets. The model is trained and validated using a custom-built safflower dataset. The experimental results demonstrate that the improved model achieves Precision (P), Recall (R), mean Average Precision (mAP), and F₁ score values of 93.15%, 86.71%, 95.03%, and 89.64%, respectively. These results represent improvements of 2.9%, 6.69%, 4.5%, and 6.22% over the baseline model. Compared with Faster R-CNN, YOLOv5, YOLOv7, and YOLOv10, the WED-YOLO achieved the highest mAP value. It outperforms the module mentioned by 13.06%, 4.85%, 4.86%, and 4.82%, respectively. The enhanced model exhibits superior precision and lower miss detection rates in safflower recognition tasks, providing a robust algorithmic foundation for the intelligent harvesting of safflowers.

Path planning for picking safflowers is a key component in ensuring the efficient operation of robotic safflower-picking systems. However, existing single-arm picking devices have become a bottleneck due to their limited operating range, and a breakthrough in multi-arm cooperative picking is urgently needed. To address the issue of inadequate adaptability in current path planning strategies for dual-arm systems, this paper proposes a novel path planning method for dual-arm picking (LTSACO). The technique centers on a dynamic-weight heuristic strategy and achieves optimization through the following steps: first, the K-means clustering algorithm divides the target area; second, the heuristic mechanism of the Ant Colony Optimization (ACO) algorithm is improved by dynamically adjusting the weight factor of the state transition probability, thereby enhancing the diversity of path selection; third, a 2-OPT local search strategy eliminates path crossings through neighborhood search; finally, a cubic Bézier curve heuristically smooths and optimizes the picking trajectory, ensuring the continuity of the trajectory’s curvature. Experimental results show that the length of the parallelogram trajectory, after smoothing with the Bézier curve, is reduced by 20.52% compared to the gantry trajectory. In terms of average picking time, the LTSACO algorithm reduces the time by 2.00%, 2.60%, and 5.60% compared to DCACO, IACO, and the traditional ACO algorithm, respectively. In conclusion, the LTSACO algorithm demonstrates high efficiency and strong robustness, providing an effective optimization solution for multi-arm cooperative picking and significantly contributing to the advancement of multi-arm robotic picking systems.

Clamping root-cutting end-effector for harvesting fresh safflower