Soybean seed germination rate is a key indicator of seed quality, influencing yield. However, traditional germination inspection fails to meet the requirements for full-cycle, non-destructive detection. To address this, this study proposes an enhanced neural network based on YOLOv8n for soybean seed germination detection, seamlessly integrated into the assessment process and coined as APE-YOLOv8n. In this study, a comprehensive dataset encompassing the entire soybean seed germination cycle was established. The ADown module was incorporated to substitute the original convolutional layer, facilitating cross-scale feature fusion and enhancing the capability of the model to capture defect-related information. In addition, PConv was incorporated into the Bottleneck structure of C2f to reduce feature redundancy and enhance feature extraction efficiency. Moreover, the EIOU loss function was adopted to further improve localization accuracy and strengthen of the model capability in capturing critical regional features. Experimental results demonstrated that the APE-YOLOv8n model achieved a mAP@0.5 of 94.9 % and a precision of 90.8 %, with parameter count and computational cost measured at 3.4 M and 5.1 GFLOPs, respectively. Compared to the original YOLOv8n, the proposed model exhibited reductions in parameters and FLOPs by 42.4 % and 37.0 %, respectively, while achieving improvements in precision and mAP@0.5 of 2.4 % and 1.3 %. Moreover, it outperforms not only the baseline but also other comparative models.
Shumei et al. (Sun,) studied this question.