The research utilized a dataset of tobacco leaf maturity images, which was established from the collection of leaves in the tobacco cultivation region of Luoning County, Luoyang City, within Henan Province. For the acquisition of field data, the study employed the rear camera of a Huawei Honor 20 smartphone, featuring a 32-megapixel high-resolution sensor. To minimize the impact of lighting conditions on the leaf maturity recognition, the data was collected exclusively during daylight and under clear skies. To further augment the complexity of the dataset and enhance the robustness of our model, we employed data augmentation techniques such as rotation, scaling, flipping, and the addition of noise. The tobacco leaves were classified into three distinct maturity stages: immature, mature, and over-mature. Immature leaves, characterized by their green color, are not harvest-ready. Mature leaves are identified as the optimal stage for harvesting without compromising the final product’s quality. Over-mature leaves, indicative of an excessive degree of maturity, are prone to significant losses during the harvesting and subsequent processing stages. In this study, the dataset was randomly partitioned following an 8:1:1 ratio into training, validation, and test sets, respectively. The training set consists of 1,752 images, the validation set contains 370 images, and the test set comprises 373 images. Representative images from the developed tobacco leaf maturity dataset are depicted in Figure 1 .

The present investigation applies Precision (P), Recall (R), mAP50, and mAP50-95 as the evaluative metrics for the tobacco leaf maturity detection model. Precision delineates the proportion of tobacco leaves that are accurately classified by the model into a specific maturity stage, signifying the model’s trustworthiness in predicting particular maturity levels. Recall measures the model’s effectiveness in identifying all instances of a given maturity stage, representing the ratio of correctly detected leaves to the total actual instances. mAP50 emerges as a holistic benchmark in the evaluation of tobacco leaf maturity, encapsulating the model’s aggregate proficiency in distinguishing among various stages. It is calculated by averaging the AP values across stages, thereby assessing the model’s comprehensive accuracy in classifying tobacco leaf maturity. mAP50-95 expands the IoU threshold scope, pivotal for nuanced visual feature differentiation across maturity stages. This metric furnishes an encompassing view of the model’s efficacy across a spectrum of matching stringencies. The respective computational formulas are articulated in Equations 2-5.

Where, TP is the tally of veracious positive instances, FP the tally of fallacious positive instances, and FN the tally of fallacious negative instances. N encapsulates the aggregate number of categories. A P i is the mean precision for the i-th category at an IoU threshold of precisely 0.5. A P i , j pertains to the mean precision for the i-th category at an IoU threshold incrementing from 0.5 by increments of 0.05 for each successive j, ranging up to 0.95. The term ‘91’ embodies the methodical computation of AP across this continuum, spaced into 91 uniform intervals for a meticulous assessment of AP.

The experimental procedures described herein were undertaken within a Windows 11 environment, leveraging the PyTorch deep learning framework at version 2.0.1, with Python 3.9 serving as the programming language of choice and PyCharm acting as the IDE for coding endeavors. The computational experiments were powered by an Intel Core i5-13500h CPU, complemented by 16 GB of system memory. The GPU designated for this research is the NVIDIA GeForce RTX 4050, endowed with 6 GB of graphics memory and 2560 CUDA cores for parallel processing capabilities. To ensure the reliability of our model, we adopted a consistent set of hyperparameters for all training runs. The hyperparameters for model training were sourced from https://github.com/THU-MIG/yolov10/blob/main/ultralytics/cfg/default.yaml. The specific values are summarized in the Table 1 .

In order to better evaluate the impact of C2f-DCNv3 on different parts of the model, this study utilizes C2f-DCNv3 to replace the C2f module in the backbone network and necking network, respectively, in order to enhance the performance of the model. The outcomes from integrating C2f-DCNv3 at these distinct phases are delineated in Table 2 .

As indicated in Table 2 , the overall model accuracy improved from 0.939 to 0.970, marking a 3.3% increase, when the C2f module in the backbone network was replaced in isolation. The mAP50 metric also saw a slight rise from 0.984 to 0.991, amounting to a 0.7% increase. Notably, within the “Immature” category, there was a significant leap in accuracy, with mAP50 and mAP50-95 experiencing boosts of 1.5% and 3.3%, respectively. Following the replacement of the neck network, the overall precision was further enhanced to 0.973, a 3.7% increase. The mAP50 metric mirrored the initial rise, while the mAP50-95 improved from 0.962 to 0.972, reflecting a 1.0% increase. Conversely, replacing the C2f modules in both the backbone and neck networks concurrently resulted in an overall precision of 0.971, yet the mAP50-95 dipped slightly to 0.958.

The incorporation of the C2f-DCNv3 module has notably enhanced the YOLOv10 model’s performance, particularly within the neck network structure. The C2f-DCNv3’s design amalgamates the profound feature extraction capabilities of Convolutional Neural Networks (CNNs) with the adaptability of Deformable Convolutional Networks (DCNs), thus enabling the model to adeptly adjust to the variability in target shapes and spatial configurations. Acting as a conduit between the backbone and detection head, the neck network’s efficacy is pivotal to the detection precision. Replacing the C2f module with C2f-DCNv3 in the neck network has bolstered the model’s target recognition by enriching feature representation. However, the decline in mAP50-95 when both networks are updated with C2f-DCNv3 could be attributed to potential issues. It may stem from overfitting due to heightened model complexity, especially with limited data. Alternatively, suboptimal feature integration strategies between the backbone and neck networks could lead to information loss or redundancy.

In this research, the strategy of replacing the C2f module in the neck network with C2f-DCNv3 has been selected from the outcomes of employing C2f-DCNv3 at various stages, as it demonstrated the most substantial benefit in enhancing model performance. Consequently, the C2f-DCNv3 module is chosen to replace the C2f module in the neck network to augment the model’s capabilities.

In this research, we have made significant improvements to the YOLOv10n object detection model by incorporating the ELA (Efficient Layer-wise Attention) module to enhance the precision and efficiency of tobacco leaf maturation identification. Initially, we substituted the PSA (Pointwise Spatial Attention) mechanism in YOLOv10n with the ELA, creating the YOLOv10n+ELA1 model. Subsequently, we introduced an additional ELA module at the juncture of the backbone and neck networks within the YOLOv10n+ELA model to potentially elevate the model’s performance further. The precision of models with attention mechanisms modified at various stages is detailed in Table 3 .

From Table 3 , it is clear that the enhanced model has shown significant performance improvements across all maturity categories. Specifically, for the “Over-Mature” category, the YOLOv10n+ELA1 model’s accuracy has increased from 0.905 to 0.95, and the mAP50 has improved from 0.975 to 0.983. In the “Immature” category, both accuracy and mAP50 have reached 0.991 and 0.995, respectively, demonstrating an exceptionally high recognition rate. Moreover, the YOLOv10n+ELA model has achieved an overall precision and mAP50 of 0.972 and 0.992 for the “All” categories, which is a 3.3% and 0.8% increase compared to the original YOLOv10n model.

The incorporation of the ELA module has notably bolstered the model’s capability to capture features indicative of tobacco leaf maturity. The ELA’s design, leveraging inter-layer attention mechanisms, effectively enhances the interconnectivity of feature maps, thus improving the model’s differentiation between tobacco leaves of varying maturities. Additionally, by incorporating ELA at the interface of the backbone and neck networks, we have further strengthened the conveyance and integration of features, enabling the model to sustain high recognition accuracy even when dealing with images of tobacco leaves against complex backgrounds and under diverse lighting conditions.

However, we have also noted a decrease in mAP50-95 for the “Over-Mature” category in the YOLOv10n+ELA1 model compared to the original model. This may indicate that the model’s ability to recognize extreme cases of tobacco leaf maturity has been somewhat compromised during the enhancement process. This could be attributed to the introduction of the attention mechanism, which may have altered the distribution of features, potentially diminishing the model’s generalization capabilities in certain scenarios.

In this research, a comprehensive set of enhancements has been strategically applied to substantially elevate the performance of the YOLOv10 model. These improvements encompass the innovative replacement of the C2f module with the C2f-DCNv3 within the neck structure, alongside the sophisticated transition from the PSA (Pointwise Spatial Attention) mechanism to the ELA (Efficient Local Attention) mechanism within the backbone network. The seamless integration of an additional ELA attention mechanism at the interface of the backbone and neck networks has culminated in the development of a model that excels in the sophisticated recognition of tobacco leaf maturity. The model accuracy of the multi-stage improved fusion is shown in Table 4 .

As demonstrated in Table 4 , the enhanced model from this study surpasses the original YOLOv10n model in multiple indicators. In general, the precision (P) of our model across all categories has seen a rise from 0.939 to 0.973, which is a 3.4% increase; the recall (R) has also seen an improvement, increasing from 0.968 to 0.984, a 1.6% increase. The Mean Average Precision at 50% intersection over union (mAP50) has increased from 0.984 to 0.994, a 1.0% improvement; and the mAP50-95 has also shown an increase, moving from 0.962 to 0.973, a 1.1% increase.

In the granularity of specific categories, our model exhibits considerable improvement within the “Over-Mature” classification, with accuracy escalating from 0.905 to 0.973, reflecting a 6.8% enhancement; the recall rate has also witnessed an uptick from 0.949 to 0.969, a 2.0% gain; mAP50 has seen a boost from 0.975 to 0.991, a 1.6% advancement; and mAP50-95 has climbed from 0.946 to 0.970, a 2.4% escalation. Within the “Mature” classification, accuracy has surged from 0.923 to 0.970, amounting to a 4.7% enhancement; the recall rate has spiked from 0.966 to 0.992, a 2.6% augmentation; mAP50 has risen from 0.984 to 0.995, a 1.1% increment; and mAP50-95 has inched up from 0.972 to 0.981, a 0.9% increase. For the “Immature” classification, accuracy has slightly edged from 0.991 to 0.975; the recall rate has marginally improved from 0.989 to 0.991, a 0.2% increment; mAP50 has sustained its level at 0.995; and mAP50-95 has maintained its steadiness at 0.968.

The ELA demonstrates excellent performance in terms of computational efficiency and the enhancement of model capabilities. By adeptly capturing local features and providing advanced feature representation, ELA markedly boosts the model’s precision and generalization ability. Its primary strengths are the efficient capture of local features, optimization of channel dimensions, and a simplified structure, circumventing the redundancy and increased computational complexity inherent in global feature extraction. These attributes render ELA especially fitting for compact models and real-time applications, thus augmenting overall computational efficiency.

The C2f-DCNv3 module, a fusion of DCNv3 and the C2f module, strengthens the model’s adaptability to varied shape changes and spatial configurations. It leverages the adaptability of DCNv3 and the profound feature extraction capabilities of convolutional neural networks to further refine the model’s detection precision and robustness. The integration of the C2f-DCNv3 module into the neck network facilitates superior integration of multi-scale features, enhancing the accuracy of target recognition. Additionally, employing the ELA attention mechanism in conjunction with the C2f-DCNv3 module not only enhances detection precision but also bolsters the model’s robustness and generalization capabilities.

To better demonstrate the capabilities of our model, this study compared it against four existing YOLO series models (specifically, YOLOv5n, YOLOv6n, YOLOv8n, and YOLOv10n). The comparative accuracy of these models is detailed in Table 5 . The results of tobacco maturity detection for different models are shown in Figure 7 .

As depicted in Table 5 , our model exhibited superior performance in the task of recognizing the maturity of tobacco leaves. Specifically, it achieved a P of 0.973, a R of 0.984, and mAP50 and mAP50-95 of 0.994 and 0.973, respectively, indicating exceptionally high detection accuracy. Within the subcategories representing different stages of maturity, our model continued to excel, maintaining mAP50-95 values above 0.968 for Over-Mature, Mature, and Immature categories, thereby highlighting the model’s high accuracy and robustness in identifying tobacco leaves at various stages of maturity.

In the horizontal analysis of the tobacco leaf maturity recognition models, our model demonstrated significant superiority across all four key performance metrics for all categories. For instance, when compared to the YOLOv10n model, our model showed improvements of 3.4% in precision, 1.6% in recall, 0.10% in mAP50, and 1.1% in mAP50-95. The performance gains were even more pronounced when compared to the YOLOv5n model, with increases of 5.4%, 1.8%, 1.1%, and 4.0% in these metrics, respectively. Similarly, when compared to the YOLOv6n model, our model’s improvements were 4.5% in precision, 7.5% in recall, 3.2% in mAP50, and 4.2% in mAP50-95. Although the YOLOv5n and YOLOv6n models showed good performance in certain metrics—YOLOv5n, for example, achieved an mAP50 of 0.995 for the immature category—our model overall exhibited a more outstanding comprehensive performance across all categories. While the YOLOv8n model was comparable to ours in some subcategories, such as a mAP50-95 of 0.945 for the Over-Mature category, our model showed higher consistency and stability across all maturity categories.

To visually represent the performance of the models, a confusion matrix was employed to directly illustrate the detection capabilities. As shown in Figure 8 , our model had the fewest misclassifications across the three maturity stages, followed by YOLOv10n, which had a lower total number of misclassifications. The YOLOv6n model performed the poorest, with the highest total number of misclassifications across all categories.

To provide a more comprehensive evaluation of the model, this study employs the PR curve to assess the overall performance of the model in terms of recall and precision. The PR curves for different models are illustrated in Figure 9 .

As shown in Figure 9 , the ours model achieves a mAP50 of 0.994 at all categories, significantly surpassing other models. Specifically, YOLOv10n reaches a mAP50 of 0.984, while YOLOv5n and YOLOv8n achieve a mAP50 of 0.983 and 0.962, respectively. This indicates that the ours model has a distinct advantage in precision and recall, particularly maintaining a high level of precision in the high-recall region. Furthermore, the ours model also demonstrates outstanding performance in specific categories, achieving a mAP of 0.942 in the Over-Mature category, compared to YOLOv10n’s 0.975, suggesting that the ours model is slightly less effective in this category. However, in the immature and Over-Mature categories, the ours model achieves an mAP of 0.995 at a threshold of 0.5, showcasing its robust performance in these areas. Overall, the ours model exhibits excellent performance across multiple evaluation metrics, particularly with its overall performance of 0.994 mAP at 0.5, which is markedly higher than that of other models, underscoring its exceptional capabilities and potential in object detection tasks.

In conclusion, our model offers an efficient and precise solution in the domain of tobacco leaf maturity recognition. Its exceptional performance in key performance metrics, coupled with its clear advantages over existing models, underscores its significant potential for practical applications in agriculture. Future efforts will focus on further optimizing the model to minimize computational resource consumption and exploring its applicability in a broader range of agricultural monitoring tasks.