A Multi-Task Model for Quality Recognition of Seedling Transplantation

The mechanical characteristics of duckbill seeders, combined with complex environmental disturbances in the field, often lead to quality problems such as exposed seedlings and stem-buried seedlings when mechanically transplanted. These issues directly affect the subsequent yield and efficiency of automated harvesting. This paper is based on the YOLOv11 architecture and studies an automatic recognition model (Yolov11-ARM) for the transplanting quality of cabbage seedlings. The model applies a custom CPNMViTBv3 backbone that integrates Transformer-based global context modeling with efficient CNN feature reuse, replacing the standard C3f module to more effectively capture structural details in seedlings. The detection head is augmented with multi-scale feature aggregation and a lightweight SENetV2 attention mechanism, enhancing discriminative capability for small seedling instances. Additionally, a dynamically adaptive loss function for keypoint estimation is introduced, which adjusts according to object scale and error distribution, thereby ensuring training stability and robust localization performance. Experimental results show a high mean Average Precision (mAP@0.5) of 98.7% for seedling status (object) detection and 99.5% for keypoint localization with the studied model. This research provides an effective technical solution for the real-time, multi-parameter quality assessment of automated transplanting.