In object detection network design, researchers have proposed multiple improvement strategies. With the development of deep learning technology, object detection networks continue to evolve. The Pruned-YOLO v5s+Shuffle model proposed by Xu et al. (2022) employs channel pruning method, achieving 93.2% detection accuracy in complex backgrounds. The Yolov5-ECA-ASFF network proposed by Zhang et al. (2024) enhances detection performance by integrating ECA and ASFF modules. Lin et al. (2024) optimized the YOLO model through combining mixed data augmentation and osprey search strategy, realizing tomato biotic stress detection. The WCG-VMamba model developed by Wang et al. (2024) introduces wavy vision Mamba network, effectively capturing semantic correlations between image features and text features, further improving detection performance in complex backgrounds. The cross-domain dynamic attention mechanism designed by Mo and Wei (2024) effectively solves uneven illumination problems. Mhala et al. (2025) addressed class imbalance issues through model compression and knowledge distillation techniques, achieving efficient model deployment. These studies indicate that object detection network design is evolving towards better adaptation to complex environments and higher accuracy.

Despite promising results in agriculture, existing YOLO-based methods still face fundamental limitations in greenhouse applications: (1) Standard strided convolutions in YOLO backbones sacrifice spatial resolution for computational efficiency, but disease symptoms often manifest as subtle texture changes requiring preservation of fine-grained details; (2) Traditional feature pyramid networks inadequately handle the extreme scale variation of disease symptoms, from macro-level leaf discoloration spanning hundreds of pixels to micro-level lesions occupying fewer than 20 pixels; (3) Fixed feature fusion weights in existing architectures cannot adapt to the dynamic visual complexity of greenhouse environments where lighting conditions, plant density, and background complexity vary significantly. Our ADEConv module specifically addresses the feature preservation challenge while maintaining computational efficiency.

In feature extraction optimization, the introduction of various innovative mechanisms has led to significant breakthroughs. Liu et al. (2021) proposed region and loss reweighting methods, providing new insights for feature extraction optimization. The EFDet model developed by Liu et al. (2024) improves detection effects in complex backgrounds by fusing features from different levels. Yan et al. (2024) proposed an adaptive deep transfer learning framework for mixed subdomains, significantly improving cross-species disease diagnosis performance. Notably, Kang et al. (2024) proposed a cascade framework combining detector and tracker, significantly reducing computational complexity while maintaining high accuracy, providing a feasible solution for practical application scenarios. Sun et al. (2025) proposed a new tomato disease recognition method based on the DeiT model, significantly improving detection accuracy in complex environments through improved feature extraction and multi-scale feature fusion mechanisms.

While attention-based approaches show promise, most existing methods apply static attention weights. Chang et al. (2024) improved wheat disease recognition through DenseNet modifications, but their approach lacks the dynamic adaptability required for greenhouse environmental variations. The AAFS mechanism differs fundamentally from existing attention approaches: Unlike SE-Net which focuses solely on channel attention through global average pooling, AAFS integrates both channel and spatial attention through parallel pathways. Compared to CBAM which applies channel and spatial attention sequentially, our approach enables simultaneous processing and dynamic weight fusion. Unlike ECA-Net’s 1D convolution for channel attention, AAFS employs adaptive group convolution with channel shuffling for enhanced feature interaction.

In data augmentation strategies, researchers have proposed a series of innovative methods to address the unique challenges in greenhouse environments. The multi-scale feature enhancement strategy proposed by Tian et al. (2022) significantly improved the model’s recognition ability for disease regions. Karantoumanis et al. (2024) developed a strategic data augmentation method achieving a 37% accuracy improvement in legume crop disease detection. Zhang et al. (2024) proposed feature transfer and small target oversampling methods based on CycleGAN, effectively improving sample imbalance issues and successfully achieving precise recognition of early eggplant wilt disease. Johri et al. (2024) combined deep transfer learning with data augmentation, achieving significant results in small sample scenarios, providing new ideas for solving data insufficiency problems.

In multi-modal fusion, researchers have gradually begun to focus on the synergistic use of multi-source information. Yang et al. (2024) innovatively proposed a language-vision fusion framework, demonstrating excellent performance in tomato disease segmentation tasks. Hu et al. (2024) achieved deep fusion of spectral information and RGB images, significantly improving disease detection accuracy. Zhao et al. (2025) successfully implemented complementary fusion of healthy and diseased leaf information using Double Generative Adversarial Networks (DoubleGAN), providing new ideas for disease detection in small sample scenarios.

Regarding small target localization and detection recognition in complex backgrounds, the unique characteristics of greenhouse environments bring distinct challenges to disease detection. Barbedo (2019) research showed that disease recognition faces challenges of small target size, blurred target features, and occlusion problems. Toda and Okura (2019) revealed the decision mechanism of CNNs in plant disease diagnosis under complex environments. Kumar et al. (2023) proposed a bidirectional feature attention pyramid network, effectively enhancing the model’s detection capability for targets of different scales. Zhou et al. (2023) innovatively introduced weakly supervised learning into disease feature segmentation, providing new approaches for small target detection. Ye et al. (2024) proposed an adaptive small target detection framework, significantly improving detection performance in low-light environments by integrating EnlightenGAN networks. Hari and Singh (2025) proposed an adaptive knowledge transfer method based on federated deep learning, significantly improving model convergence and accuracy through intelligent weight transfer technology optimizing knowledge integration between parent and child entities.

However, existing research still faces severe challenges in complex, unstructured greenhouse environments. First, image acquisition in greenhouse environments suffers from serious quality degradation issues, including image blur, noise interference, and uneven illumination, leading to significant false detections and missed detections in practical applications. Second, vegetable disease symptoms often manifest as small local areas of pathological changes, and these subtle features are easily lost during feature extraction, making them difficult to capture accurately (Qing et al., 2023). Furthermore, feature expression and multi-scale feature fusion mechanisms under complex background interference remain unresolved (Castillo-Girones et al., 2025).

To address these issues, this study proposes a high-precision localization and detection algorithm (YOLO-vegetable) for vegetable disease targets in greenhouse environments, based on the computationally efficient YOLOv10n single-stage object detection network. The algorithm contains three core innovative modules: First, we design the Adaptive Detail Enhancement Convolution (ADEConv) module, which significantly improves fine-grained feature retention capability while maintaining computational efficiency through dynamic adjustment of convolution kernel parameters; Second, we construct the Multi-granularity Feature Fusion Detection Layer (MFLayer), which achieves precise localization of small targets through hierarchical integration of feature information at different scales; Finally, we propose the Inter-layer Dynamic Fusion Pyramid Network (IDFNet), combining with Attention-guided Adaptive Feature Selection (AAFS) mechanism, significantly enhancing the model’s key information extraction capability by establishing dynamic association weights between feature layers.

Based on the characteristics of vegetable disease targets requiring detection, this study proposes a detection algorithm model YOLO-vegetable targeting greenhouse environments. Taking YOLOv10n, which has the smallest parameter count in the detection-performance-excellent YOLOv10 series, as the baseline model, we redesigned the backbone network and neck network of the original model. The structure of YOLO-vegetable is shown in Figure 1 .

The diagram illustrates the complete network structure with backbone (left), neck network with IDFNet (center), and detection heads (right). Red boxes highlight our proposed modules: ADEConv modules replace traditional strided convolutions in the backbone, MFLayer provides multi-granularity feature fusion for small target detection, and AAFS mechanisms enable adaptive feature selection throughout the neck network. Input images (640×640) flow through the backbone for feature extraction, then through IDFNet for multi-scale feature fusion, finally reaching dual detection heads for classification and regression outputs.

To ensure the dataset encompasses diverse greenhouse environments and meets the model’s requirements for handling complex backgrounds, occlusion, blurred disease features, and small target detection, this study employs our self-built Vegetable Disease Dataset (VDD), comprising 15,000 images involving 3 major facility vegetables (tomato, cucumber, pepper) and their 15 common diseases along with healthy samples ( Figure 7 ; Table 2 ). Data collection was conducted in controlled greenhouse facilities with temperature at 22-28°C (day) and 18-22°C (night), and relative humidity at 60-75%. Images were captured using professional high-resolution cameras at 30-50cm distance across four growth stages (seedling, vegetative, flowering, fruiting) under diverse weather conditions to ensure dataset robustness. The dataset is annotated following YOLO format specifications. Dataset annotation was performed by certified plant pathologists following standardized protocols. Each disease instance was annotated with precise bounding boxes. The dataset is divided into training set, validation set, and test set in a 7:2:1 ratio. This dataset contains vegetable disease targets under various weather conditions.

Figure 7 showcases representative samples from our dataset, illustrating the diversity of disease manifestations across different vegetable types and growth stages. The images demonstrate varying symptom presentations, from early-stage subtle discolorations to advanced necrotic lesions, captured under diverse lighting conditions and viewing angles. This diversity ensures models trained on our dataset develop robust generalization capabilities applicable to real-world greenhouse scenarios.

The experimental platform uses Ubuntu22.04 as the operating system, equipped with Intel(R) Xeon(R) Gold 5418Y processor with a main frequency of 2.00 GHz. The system memory is 32GB, with an Nvidia GeForce RTX 4090 graphics card having 24GB memory capacity. The PyTorch framework version is 2.2.2+cu121, and Python version is 3.10.0. Input image resolution is uniformly set to 640×640 to ensure the clarity of targets at different scales in feature maps, adapting to the model’s requirements for small target detection. During training, the model’s initial learning rate is set to 0.01, batch size to 16, momentum to 0.937, weight decay coefficient to 0.0005, and training epochs to 100. To further enhance the model’s robustness, all experiments are conducted without any form of pre-trained weights, and all experiments use consistent hyperparameters for training and validation to ensure comparability of experimental results.

Data augmentation techniques, including mosaic and mixup, were applied to enhance dataset diversity. Our augmentation pipeline also included random rotation (± 15°), horizontal and vertical flipping, and adjustments to brightness (± 25%), contrast (± 20%), and saturation (± 15%) to simulate the variable lighting conditions in greenhouse environments. To address class imbalance issues, we employed oversampling for minority disease classes, ensuring balanced representation during training while maintaining authentic image characteristics.

For hyperparameter optimization, we conducted a systematic grid search to identify optimal values. The learning rate was initialized at 0.01 and adjusted using a cosine annealing scheduler with warm restarts. Weight decay was set to 0.0005, and momentum maintained at 0.937 throughout training. These parameters were selected after evaluating 16 different configurations, with the final values providing the best balance between convergence speed and model generalization.

To comprehensively evaluate YOLO-vegetable model’s balanced performance in terms of speed and accuracy, this study selects Precision, Recall, Average Precision (AP), Mean Average Precision (mAP), Parameters, FLOPs, and Inference Time as evaluation metrics. In object detection tasks, mAP@0.5 and mAP@0.5:0.95 serve as primary evaluation metrics, capable of comprehensively evaluating model performance. Specifically, mAP@0.5 represents average precision at Intersection over Union (IoU) threshold of 0.5; mAP@0.5:0.95 reflects model stability under different IoU thresholds. Meanwhile, through evaluating parameter count and computational complexity, we provide important references for practical deployment.

To significantly improve the accuracy and efficiency of vegetable disease target detection, we propose YOLO-vegetable. The experimental process includes three key phases, as shown in Figure 8 .

To validate the effectiveness of the proposed model, experiments were conducted on our self-built vegetable disease dataset. The training and validation curves of the proposed model’s box loss, dfl loss, classification loss, and other performance metrics including precision, recall, mAP@0.5, and mAP@0.5:0.95 are shown in Figure 9 , with iteration count on the horizontal axis.

As shown in Figure 9 , during the 100 training iterations, the loss exhibit stable convergence patterns, gradually stabilizing as training progresses. Similarly, the validation loss demonstrate consistent convergence behavior, reaching steady states by the final epochs. Observing the model’s mAP@0.5 and mAP@0.5:0.95 convergence curves, performance metric curves begin to stabilize after 50 iterations. Finally, the model achieves excellent performance on the test set: mAP@0.5 reaches around 95%, and mAP@0.5:0.95 reaches approximately 60%. Meanwhile, the model demonstrates good precision and recall performance, indicating strong generalization ability and stability in vegetable detection tasks. The YOLO-vegetable model achieves a parameter count of 3.8M and a computational complexity of 14.7 GFLOPs, making it highly efficient for real-time deployment in resource-constrained environments.

To comprehensively evaluate YOLO-vegetable model’s performance in detecting various vegetable diseases, testing was conducted based on our self-built dataset. Table 3 shows the detection results of YOLO-vegetable model for these different types of diseases.

As shown in Table 3 , YOLO-vegetable model achieves Precision, Recall, and AP values above 90% for 15 vegetable diseases and healthy samples, demonstrating high precision and recall rates. The model’s mAP reaches 95.6%, fully proving its excellent performance in handling different types of vegetable diseases. Additionally, the model’s outstanding performance in healthy sample recognition helps reduce misdiagnosis and unnecessary treatments.

Figure 10 presents the confusion matrix of our proposed YOLO-vegetable model, showing the proportion of detection results for each category. The horizontal axis represents predicted class numbers, while the vertical axis represents annotated class numbers. In the matrix, squares where predicted classes match annotated classes represent correct algorithm predictions, while other squares represent class confusion cases. From the prediction results, the model demonstrates high detection accuracy with minimal confusion overall, showing only a small proportion of class confusion cases.

To comprehensively evaluate the detection performance of the proposed YOLO-vegetable model compared to the baseline model, we analyzed the Precision-Recall (PR) curves, which illustrate the trade-off between precision and recall at different confidence thresholds. Figure 11 presents the PR curves for both models.

The PR curve analysis reveals that the proposed YOLO-vegetable model ( Figure 11a ) achieves superior performance with an Average Precision (AP) of 0.956, representing a significant improvement over the baseline model’s AP of 0.892 ( Figure 11b ). The YOLO-vegetable model maintains higher precision values across a wider range of recall values, indicating its ability to identify disease instances correctly while minimizing false positives. The enhanced performance demonstrated by the PR curves further validates the effectiveness of our architectural improvements—specifically the ADEConv module for preserving fine-grained features, the MFLayer for accurate small target localization, and the IDFNet for enhanced feature fusion. These components collectively contribute to the model’s ability to maintain high precision even at high recall thresholds, making it well-suited for real-world greenhouse disease detection scenarios with varying lighting conditions and complex backgrounds.

To systematically evaluate the performance contribution of each core module in the YOLO-vegetable algorithm, this study uses YOLOv10n as the baseline model, progressively introducing the ADEConv, MFLayer, and IDFNet modules. Through comprehensive analysis of model accuracy, computational complexity, and inference time, we validate the optimization effect of each module. Detailed experimental results are shown in Table 4 .

The experimental results show that introducing the ADEConv module improves mAP@0.5 from 89.2% to 94.3%, significantly enhancing the network’s fine-grained feature extraction capability. Although parameter count increases from 2.2M to 3.6M and computational cost increases to 9.5 GFLOPs, it only brings a 0.5ms inference time delay (15.6ms to 16.1ms). While the MFLayer module leads to computational cost increasing to 15.1 GFLOPs with a 2.6ms inference time increase, it performs excellently in maintaining small target detail features, achieving 93.2% mAP@0.5 with only 2.8M parameters. The introduction of IDFNet demonstrates superior feature fusion effects, achieving 94.3% mAP@0.5 with just 2.7M parameters, while maintaining comparable computational cost (8.3 GFLOPs) and inference time (15.9ms).

Further research reveals that the combination of ADEConv and MFLayer achieves 94.5% mAP@0.5, surpassing single-module applications. Although computational cost increases to 16.7 GFLOPs, through reasonable parameter configuration (3.8M), the inference time increase (20.1ms) remains acceptable. This result demonstrates the synergistic effect between detail feature extraction and feature preservation. Building upon this foundation, introducing IDFNet to form the complete YOLO-vegetable model not only further improves mAP@0.5 to 95.6% but also achieves optimization in computational resource utilization: maintaining parameter count at 3.8M, reducing computational cost to 14.7 GFLOPs, and controlling inference time to 18.6ms. This balance between performance improvement and computational overhead fully validates the necessity of innovative modules and their excellent synergistic effects.

To more intuitively demonstrate the performance improvement effects of different modules on the model, Figure 12 illustrates the trends of model performance as different modules are introduced. From the overall trends in Figure 12 , the progressive introduction of the three innovative modules shows steady performance improvement, with balanced enhancement across all metrics, demonstrating no significant degradation in any indicator while others improve. This balanced performance improvement validates that our proposed improvement strategies are not only necessary but can work collaboratively and mutually reinforce each other, achieving overall optimization of model performance.

To better understand the model’s decision-making process, we visualized the feature activation maps using Grad-CAM techniques ( Figure 13 ). These visualizations demonstrate that our YOLO-vegetable model correctly focuses on disease-affected regions while effectively filtering out background noise. For smaller lesions, the model exhibits precise localization, confirming the effectiveness of our detail-preserving modules. Comparative analysis of activation maps between the baseline model and YOLO-vegetable reveals distinct differences in feature focus. While the baseline model tends to activate broadly across leaf surfaces with disease-like coloration patterns, our model demonstrates more precise localization specifically on the actual disease lesions. This is particularly evident in the second row of Figure 13 , where the baseline model shows diffuse activation across multiple spots, while YOLO-vegetable concentrates activation intensity precisely on the primary disease lesions. This improved focus significantly reduces false positives in complex backgrounds with similar color patterns to diseases but different textural features, a common challenge in greenhouse environments with varying light conditions creating shadowing effects that resemble disease symptoms.

Figure 14 presents the comparative experimental results between the proposed YOLO-vegetable model and the baseline model during the training process. The left subfigure (a) shows the mAP@0.5 curves of YOLO-vegetable and the baseline model. From the figure, it is evident that the proposed model achieves higher mAP during training and converges faster, ultimately reaching 95.6% mAP, significantly outperforming the baseline model’s 89.2%. The right subfigure (b) displays the Loss curves of YOLO-vegetable and the baseline model. The baseline model exhibits higher Loss values, slower convergence speed, and ultimately higher final Loss values compared to the proposed model. This indicates that the proposed YOLO-vegetable model not only surpasses the baseline model in accuracy but also demonstrates better convergence behavior and lower loss during training.

To comprehensively evaluate the performance of YOLO-vegetable, we conducted extensive comparisons with mainstream object detection models. The experimental results are summarized in Table 5 . On the same vegetable disease dataset, the proposed model exhibits superior comprehensive performance. In terms of detection accuracy, YOLO-vegetable achieves 95.6% mAP@0.5, significantly exceeding the baseline model YOLOv10n (89.2%) and outperforming other mainstream detection algorithms such as Faster-RCNN (89.6%), SSD (94.2%), and YOLOv5s (93.9%). Notably, the proposed model achieves comparable performance to YOLOv10s (95.5%) while demonstrating superior resource efficiency.

From the perspective of model complexity, YOLO-vegetable exhibits significant advantages. Compared to Faster-RCNN’s 63.2M parameters, the proposed model requires only 3.8M parameters, reducing storage demands by approximately 94%. In terms of computational efficiency, YOLO-vegetable achieves 14.7 GFLOPs, substantially lower than Faster-RCNN (370.0 GFLOPs) and SSD (63.2 GFLOPs), and also outperforms YOLOv5s (23.8 GFLOPs) and YOLOv8s (28.5 GFLOPs). This marked reduction in computational cost makes the model more suitable for deployment in resource-constrained practical applications.

Regarding real-time performance, YOLO-vegetable achieves an average inference time of 18.6ms per frame, significantly outperforming two-stage detectors such as Faster-RCNN (114.8ms/frame) and single-stage detectors like SSD (22.2ms/frame). Although there is a slight increase compared to the baseline model YOLOv10n (15.6ms/frame), this latency increment is acceptable given the substantial improvement in detection accuracy (from 89.2% to 95.6%). Particularly, compared to YOLOv10s (24.8ms/frame) and YOLOv11s (21.8ms/frame), the proposed model achieves lower inference latency while maintaining comparable detection accuracy, which is crucial for real-time disease monitoring in greenhouse environments.

Through comparisons with various YOLO series variants, it is evident that YOLO-vegetable achieves an optimal balance between performance and lightweight design. Compared to lightweight models such as YOLOv10n (89.2%) and YOLOv11n (91.7%), the proposed model achieves significant accuracy improvements with only moderate increases in parameter count. When compared to YOLOv10s (94.5%) and YOLOv11s (94.0%), it maintains comparable accuracy while substantially reducing model complexity and computational overhead. This balanced performance fully validates the effectiveness of the proposed improvement strategies and provides an efficient and practical solution for vegetable disease detection in real-world applications.

To evaluate the robustness of the proposed model under noisy conditions, we conducted experiments with Gaussian and salt-and-pepper noise. The results demonstrate that YOLO-vegetable maintains high detection accuracy, with mAP@0.5 above 90% in both noise scenarios, highlighting its robustness in real-world applications.

Figure 15 presents representative detection results with bounding boxes across various greenhouse scenarios, including different lighting conditions, planting densities, and disease severities. The visualizations demonstrate YOLO-vegetable’s superior detection performance particularly in challenging cases such as partially occluded leaves, early-stage disease symptoms, and complex backgrounds with shadows. Compared to baseline models, our approach shows notably fewer false positives on healthy plant parts with similar color patterns to diseased regions, indicating enhanced feature discrimination capabilities.

To validate the generalization capability of the proposed YOLO-vegetable model, a public dataset downloaded from the Baidu PaddlePaddle platform in China was selected for generalization testing. This dataset contains 534 images and corresponding annotation files, exhibiting strong scene diversity and challenges. The download link is: https://aistudio.baidu.com/datasetdetail/292158. Training was conducted under consistent hardware conditions. The experimental results are shown in Table 6 .

The experimental results demonstrate that YOLO-vegetable exhibits excellent performance advantages on cross-scene datasets. Compared to the baseline model, precision increased by 8.3% (from 86.5% to 94.8%), indicating that the improved model maintains high detection accuracy in unknown scenarios. Recall increased by 9.4% (from 84.2% to 93.6%), proving that the model has a stronger ability to detect diseases. The mean average precision (mAP@0.5) increased by 8.8% (from 85.4% to 94.2%), demonstrating a significant improvement in the model’s overall performance. This comprehensive performance enhancement fully validates the effectiveness of the proposed improvement strategies.

In-depth analysis reveals that the superior generalization performance of YOLO-vegetable is primarily attributed to its enhanced feature representation capability. Through the adaptive detail enhancement mechanism of the ADEConv module, the model can better extract and retain fine-grained features of diseases, enabling accurate recognition across different scenarios. The multi-granularity feature fusion mechanism of the MFLayer allows the model to adaptively handle disease targets of different scales, effectively addressing the issue of target scale variation in cross-scene data. Additionally, the dynamic feature fusion strategy of the IDFNet significantly enhances the model’s adaptability to complex backgrounds, ensuring stable detection performance under varying lighting, angles, and occlusion conditions.

Qualitative analysis shows that YOLO-vegetable exhibits clear advantages in handling complex scenarios (e.g., lighting variations, partial occlusion, complex backgrounds), with both false detection and missed detection rates lower than those of the baseline model. This fully demonstrates that the proposed improvement strategies not only enhance the model’s detection accuracy but also improve its generalization capability and environmental adaptability. The experimental results indicate that the YOLO-vegetable model has excellent generalization performance, maintaining stable detection performance when faced with new, unseen data, laying a technical foundation for the large-scale application of the model in practical agricultural production.

This study successfully addresses critical challenges in greenhouse vegetable disease detection by developing YOLO-vegetable, an enhanced deep learning architecture that significantly improves detection accuracy while maintaining computational efficiency. Our approach represents a substantial advancement in applying AI technology to support agricultural new quality productive forces.

The key contributions of this work include: (1) innovative architectural designs that preserve fine-grained features while enabling multi-scale detection; (2) comprehensive experimental validation demonstrating superior performance across diverse greenhouse conditions; and (3) practical deployment considerations with optimized parameter efficiency. The proposed method achieves state-of-the-art performance on our comprehensive dataset while maintaining real-time capabilities essential for practical applications.

Experimental results validate the effectiveness of our approach, with significant improvements in detection accuracy and computational efficiency compared to existing methods. The model’s robust performance across different disease types, growth stages, and environmental conditions demonstrates its potential for widespread adoption in intelligent greenhouse systems. This work provides a foundation for advancing precision agriculture through AI-driven disease monitoring and contributes to the development of sustainable agricultural practices.

Future research directions include: (1) comprehensive cross-regional validation to establish model generalizability across diverse agricultural settings; (2) development of lightweight architectures for edge computing deployment; (3) integration with IoT systems for automated greenhouse monitoring; and (4) extension to additional crop varieties and disease types. Long-term goals focus on creating comprehensive agricultural AI platforms that support large-scale implementation of intelligent disease management systems in modern farming operations.