Method for the estimation of the cutting points in tomato seedling grafting based on improved YOLO11n

A grafting robot needs to obtain the position information for the plant seedlings to perform automatic grafting operations. Accurately measuring the cutting points required during grafting plays a pivotal role in completing high-quality grafting tasks. Traditional visual detection models exhibit suboptimal performance on edge devices due to their large model size and suffer from limited detection efficiency. To achieve rapid and precise cutting point localization, this study proposes an all-new module termed the Stimulative Upsample Block (SUB). Additionally, the Spatial and Channel Reconstruction Convolution (SCConv) and a Local Importance-based Attention (LIA) mechanism are incorporated into the YOLO11n architecture, culminating in an enhanced model named YOLO11n-LSS. Our model achieved mean average precision (mAP) values of 93.2% for the instance segmentation task and 98.9% for the key point detection task. Compared to YOLOv8n and YOLO11n, our model reduces the number of parameters and computational cost by 4.6% and 3.8%, respectively, making it a high-performance and lightweight solution. The successful application of the new algorithm will significantly improve the production efficiency of automated tomato grafting and contribute to the advancement of the tomato cultivation industry.