Objective To address the demand for online detection of grain-impurity mixtures during wheat combine harvesting, this study developed a rapid detection model based on a deep learning instance segmentation algorithm to enable real-time measurement of harvest impurity rates.

Methods To overcome the limitations of traditional manual detection, which is inefficient and inadequate for real-time monitoring in harvesting fields, a lightweight wheat impurity rate detection model named YOLOv8n-SCW was proposed. The baseline network's C2 module was reconstructed by introducing the StarNetBlock module combined with the CAA attention mechanism to optimize feature extraction and fusion capabilities. WIoUv3 was adopted as the bounding box loss function, incorporating an adaptive gradient adjustment mechanism to mitigate the adverse effects of outlier samples on the optimization process. Mosaic data augmentation was enabled throughout the training process to enhance model generalization and robustness.

Results Evaluated on a self-constructed wheat kernel-impurity image dataset, the improved model achieved a 1.3% increase in impurity segmentation precision, a 3.1% increase in recall, and a 2.24% increase in the F1 score compared to the original YOLOv8n-seg baseline. Wheat kernel segmentation accuracy was also improved, with inference time of only 0.033 s per image. Compared to the baseline model, YOLOv8n-SCW increased the number of parameters by 1.3M and computational cost by 1.7 GFLOPS, while maintaining a relatively low computational overhead alongside the accuracy improvement.

Conclusion This lightweight improved model demonstrates potential for application in vision-based detection systems for wheat combine harvesters, providing a lightweight algorithmic reference and technical support for real-time online monitoring of impurity rates during wheat harvesting operations.