Abstract
The precise determination of tobacco leaf maturity is pivotal for safeguarding the taste and quality of tobacco products, augmenting the financial gains of tobacco growers, and propelling the industry's sustainable progression. This research addresses the inherent subjectivity and variability in conventional maturity evaluation techniques reliant on human expertise by introducing an innovative YOLOv10-based method for tobacco leaf maturity detection. This technique facilitates a rapid and non-invasive assessment of leaf maturity, significantly elevating the accuracy and efficiency of tobacco leaf quality evaluation. In our study, we have advanced the YOLOv10 framework by integrating DCNv3 with C2f to construct an enhanced neck network, designated as C2f-DCNv3. This integration is designed to augment the model's capability for feature integration, particularly concerning the morphological and edge characteristics of tobacco leaves. Furthermore, the incorporation of the Efficient Local Attention (ELA) mechanism at multiple stages of the model has substantially enhanced the efficiency and fidelity of feature extraction. The empirical results underscore the model's pronounced enhancement in performance across all maturity classifications. Notably, the overall precision (P) has been elevated from 0.939 to 0.973, the recall rate (R) has improved from 0.968 to 0.984, the mean average precision at 50% intersection over union (mAP50) has advanced from 0.984 to 0.994, and the mean average precision across the 50% to 95% intersection over union range (mAP50-95) has risen from 0.962 to 0.973. This research presents the tobacco industry with a novel rapid detection instrument for tobacco leaf maturity, endowed with substantial practical utility and broad prospects for application. Future research endeavors will be directed towards further optimization of the model's architecture to bolster its generalizability and to explore its implementation within the realm of actual tobacco cultivation and processing.