Abstract
INTRODUCTION: Wild Ussurian Pear germplasm resource has rich genetic diversity, which is the basis for genetic improvement of pear varieties. Accurately and efficiently identifying wild Ussurian Pear accession is a prerequisite for germplasm conservation and utilization. METHODS: We proposed YOLOv10n-MCS, an improved model featuring: (1) Mixed Local Channel Attention (MLCA) module for enhanced feature extraction, (2) Simplified Spatial Pyramid Pooling-Fast (SimSPPF) for multi-scale feature capture, and (3) C2f_SCConv backbone to reduce computational redundancy. The model was trained on a self-made dataset of 16,079 wild Ussurian Pear leaves images. RESULTS: Experiment results demonstrate that the precision, recall, mAP50, parameters, FLOPs, and model size of YOLOv10n-MCS reached 97.7(95% CI: 97.18 to 98.16)%, 93.5(95% CI: 92.57 to 94.36)%, 98.8(95% CI: 98.57 to 99.03)%, 2.52M, 8.2G, and 5.4MB, respectively. The precision, recall, and mAP50 are significant improved of 2.9%, 2.3%, and 1.5% respectively over the YOLOv10n model (p<0.05). Comparative experiments confirmed its advantages in precision, model complexity, model size, and other aspects. DISCUSSION: This lightweight model enables real-time wild Ussurian Pear identification in natural environments, providing technical support for germplasm conservation and crop variety identification.