SD-MSL-YOLO: A lightweight model for vegetable pot seedling detection and missing seedling localization

The vegetable replanting robot is essential for addressing the issue of missing seedlings and enhancing crop yield. However, current approaches entail substantial computational costs and memory requirements. Therefore, this study introduces a lightweight detection model, SD-MSL-YOLO (Seedling Detection and Missing Seedling Localization-YOLO), to detect vegetable pot seedlings and localize missing ones. A dataset of vegetable seedling samples under diverse illumination conditions was developed for this research. The backbone network incorporates PsConv and FasterNet-SimAM modules to improve small object detection and minimize computational redundancy. To enhance representation capabilities by capturing multi-scale context information, the GLSA-BiFPN structure is utilized. Furthermore, TADDH is introduced to mitigate environmental noise and address the lack of independence between classification and localization tasks. Ultimately, the model predicts the position of missing seedlings and completes localization by integrating RGB-D cameras. Experimental findings indicate that, compared to the baseline YOLOv8n, SD-MSL-YOLO achieves a 1.34% improvement in mAP@0.5, alongside a reduction of 56.78% in parameters and 41.97% in GFLOPs. Compared to mainstream object detection models, SD-MSL-YOLO demonstrates exceptional performance in both detection accuracy and model size, attaining a mean Average Precision (mAP@0.5) of 97.53% with only 1.31 M parameters and 4.7 GFLOPs. In terms of missing seedling localization, the average Euclidean distance error relative to the robot is 26.19 mm. These results indicate that the proposed method effectively achieves seedling detection and missing seedling localization, providing technical support for the practical deployment of vegetable replanting robots.