QseedNet: A lightweight deep-learning framework for detection of quarantine weed seeds in imported soybean inspection

Quarantine weed seeds in imported soybeans pose significant ecological and economic risks, necessitating highly accurate differentiation between quarantine and non-quarantine weed species. However, conventional manual identification methods are often slow and susceptible to human error, particularly due to the high morphological similarity among different weed seed species. This study proposes an integrated system that combines a novel multi-stage dynamic sorting apparatus with a lightweight deep learning framework, QseedNet, for the automated detection of quarantine weed seeds in imported soybeans. The proposed device physically separates weed seeds from samples, alleviating occlusion and improving imaging quality. QseedNet introduces three key architectural innovations: a Micro-scale Perception Head (MPH) for enhanced detection of ultra-small targets, structural re-parameterization (REP) based on RepVGG for reduced inference complexity, and a Multi-Scale Channel Attention (MSCA) module for improved feature discriminability. Extensive experiments on the Qseed-12 dataset, which includes 12 representative quarantine weed species, demonstrate that QseedNet achieves state-of-the-art performance with a precision of 92.0 %, recall of 90.4 %, and mAP@50 of 95.1 %, while maintaining a compact model size of 3.41 MB. Real-world deployment tests on the NVIDIA Jetson Orin NX platform confirm that QseedNet delivers real-time inference at 70.42 FPS without loss of accuracy. These results highlight the model's suitability for edge-based agricultural inspection workflows.