ABSTRACT Grain cleaning operations constitute a critical preliminary step in the grain storage process, enhancing product quality and storage efficiency. The development of intelligent cleaning systems represents an effective approach to improving cleaning performance. The key to such systems lies in real‐time quality assessment, wherein the evaluation module must be cost‐effective to meet agricultural production requirements. To address this challenge, we propose the YOLO‐CIDNet grain impurity detection network based on the YOLO architecture, featuring compact parameters and high accuracy. Specifically, using maize as an example, we employ Darknet as the backbone network. Building upon YOLOv11, we integrate context‐aware coordinate attention‐enhanced deformable convolutions within the C3k2 module. The bottleneck block is encapsulated to optimize the neck, thereby enhancing the network's ability to learn irregularly shaped impurity features. Furthermore, we designed a lightweight detection head employing local convolutions and utilizing Inner‐MPDIoU as the bounding box regression loss. Throughout the training process, knowledge distillation was applied to further enhance model performance. Experimental results demonstrate that the model achieves 97.8% accuracy, 93.4% recall, 96.4% mAP50, and 85.6% mAP50:95 on the test set, with 2.2 M parameters and 6.7 GFLOPs. Among lightweight models under 10 GFLOPs, these metrics surpass the latest YOLOv11 by 0.3%, 1.9%, 1.2%, and 2.0%, respectively. Results confirm that YOLO‐CIDNet achieves high detection accuracy with minimal computational overhead, enabling precise impurity quantification in maize imagery. This provides practical guidance for designing intelligent cleaning systems in agricultural applications.
Zhang et al. (Mon,) studied this question.