Abstract
To address the problems of low detection accuracy, frequent missed detection and false detection in maize seedling counting tasks, this paper proposes a high-precision counting model for maize seedlings in dense scenes based on improved YOLOv8n, named DAQYOLO. Firstly, a dynamic and efficient multi-scale attention module is integrated into the neck network, which effectively alleviates the feature dilution of small targets in dense scenes and enhances the model’s ability to capture global and local information. Secondly, an adaptive quality-aware focal loss is adopted to perform collaborative optimization from the two aspects of classification accuracy and localization quality coupling, so as to improve the detection accuracy for overlapping and occluded targets in dense scenes. Finally, a quality-enhanced heterogeneous adaptive non-maximum suppression method is used to adaptively adjust the IoU threshold according to the density of target regions, which eliminates redundant detection boxes and improves counting accuracy simultaneously. Validated on a self-constructed dataset, experimental results indicate that—while maintaining a near-constant parameter—the proposed model outperforms the baseline YOLOv8n, with precision, recall, mAP50 and mAP50–95 increases by 2.0%, 3.3%, 2.1% and 8.0%, respectively. This method provides a technical foundation for efficient and accurate counting of maize seedlings.