Lychee(Litchi chinensis) is an important fruit crop in tropical and subtropical regions. However, severe infestations of the lychee stem borer(Conopomorpha sinensis) can reduce yield by more than 60% and compromise fruit quality. Traditional manual inspections and indiscriminate pesticide application are labor-intensive, error-prone, and contribute to pesticide resistance, underscoring the need for accurate, automated pest detection systems. Manual inspection inefficiency has been highlighted in (Sahu et al., 2023a; Dang and Wang, 2025). In recent years, with the rapid development of deep learning technology, computer vision-based object detection technologies have shown great potential in agricultural pest detection (Zhou et al., 2024; Chen et al., 2025).

YOLO series models, known for their fast detection capabilities and high accuracy, have been widely applied in object detection tasks. However, existing YOLO models still have shortcomings in small target detection, complex background interference, and positioning accuracy (Xue et al., 2025). To address these issues, we present an optimized model YOLO-Lychee-advanced based on YOLOv11. We have systematically optimized YOLOv11 for lychee pest detection. Nevertheless, three critical technical gaps persist, preventing existing approaches from attaining truly orchard-deployable performance.

Current YOLO variants lose sub-millimeter pest-hole details after eight-fold down-sampling, misclassify peel textures under variable orchard lighting, and suffer from a scarcity of lychee-specific annotated data, all of which hinder deployment in real orchards.

Our main contributions are:

The remainder of the paper reviews related work, presents the methodology, experiments, and conclusions.

This section reviews recent advances in YOLO-based pest detection, highlighting limitations addressed by our method.

In recent years, deep learning technologies have been widely applied in the field of agricultural pest detection. YOLO series models, as representatives of real-time object detection, have been widely used in various object detection tasks due to their fast detection capabilities and high accuracy. YOLO’s real-time capability has been validated in agricultural tasks (Ma et al., 2023; Fu et al., 2024).YOLOv3 (Redmon and Farhadi, 2018) introduced multi-scale predictions, YOLOv4 (Bochkovskiy et al., 2020) consolidated bag-of-freebies and bag-of-specials for speed–accuracy trade-offs, YOLOv5 (Nelson. and Solawetz., 2020) streamlined the training pipeline for production, and YOLOv8 (Gallagher, 2024) adopted anchor-free heads. These limitations are detailed in Section I.

Mainstream backbones for plant disease detection include ResNet (He et al., 2016), DenseNet (Tahir et al., 2022) and Inception (Ali et al., 2021; Bachhal et al., 2024), which we use as references for lightweight design.

Lightweight operators—depthwise separable convolution (Ma et al., 2023), attention routing (Fu et al., 2024), and slim-neck modules (Redmon and Farhadi, 2018)—as well as enlarged receptive field techniques (Bochkovskiy et al., 2020) have been widely adopted to improve small-target detection efficiency. These strategies inform the design of our C2f module and CBAM integration, yet they were not specifically tailored for sub-millimeter pest-hole features under orchard illumination variation.

There have also been many valuable studies in convolutional modules (Ma et al., 2023). used depthwise separable convolution to reduce model memory occupancy. DSC decomposes standard convolution into depthwise convolution and pointwise convolution, reducing the number of parameters and computational volume, making the model more suitable for resource-constrained environments (Fu et al., 2024). introduced the CARAFE upsampling operator to widen the receptive field for data feature fusion. CARAFE uses feature perception recombination to upsample features, predicting recombination kernels for each position based on underlying information and defining recombined features, thereby enhancing the model’s ability to capture image details. They also used the C2f Faster structure in the Backbone and Neck of YOLO v8, enhancing the model’s feature extraction capabilities. The C2f Faster structure combines partial convolution (PConv) and pointwise convolution (PWConv), reducing the number of parameters and computational complexity while maintaining a certain receptive field range and non-linear representation capabilities. This method is of great reference value in our computational tasks.

In terms of loss functions, Wang et al. (2024) used the MPDIoU (Minimum Point Distance IoU) loss function for YOLO v8n. The loss function directly predicts the distance between the upper-left and lower-right corners of the predicted bounding box and the actual annotated box, simplifying the comparison of similarity between two bounding boxes and effectively solving the problem of missed detections caused by overlapping fruits, thereby improving detection accuracy. Similarly (Fu et al., 2024), introduced the Focal SIoU loss function to address the issues of unbalanced positive and negative sample allocation and the limitations of CIoU. Focal SIoU combines Focal Loss and SIoU loss functions, reducing the weight of simple negative samples and enabling the model to focus more on hard-to-classify samples, thereby improving the model’s performance when dealing with imbalanced datasets.

In the direction of feature fusion, Lin et al. (2017) proposed the FPN structure (Feature Pyramid Network), which constructs a feature pyramid to fuse features from different levels, enabling the model to capture both global and local feature information simultaneously. In plant disease detection, this multi-scale feature fusion helps accurately identify different sizes and shapes of disease regions (Li et al., 2022b; Ghayoumi, 2024), especially suitable for processing high-resolution agricultural images. This helps improve the accuracy and robustness of disease detection. PANet (Liu et al., 2018)further optimizes the feature fusion path based on FPN, improving feature propagation efficiency and model performance through bidirectional feature fusion. BoTNet’s (Srinivas et al., 2021) MHSA module can handle feature maps of different scales, enabling the model to better capture global information and local details of targets, improving the model’s recognition capabilities in complex backgrounds and occlusion situations (Ma et al., 2023).

Attention mechanisms enhance the model’s focus on key features by automatically learning important regions in images. For example, in plant disease detection (Ramamurthy et al., 2020; Guo et al., 2024; Han et al., 2024), attention mechanisms can help the model more accurately locate disease regions, thereby improving detection accuracy. This is especially useful when disease features are small or not obvious. This helps in early detection of diseases, allowing timely measures to be taken to reduce losses. Related work includes Guo et al (Guo et al., 2022), who introduced attention mechanisms such as SE, ECA, and CBAM into target detection models like Faster R-CNN, YOLOx, and SSD, significantly improving the detection accuracy of grape leaf diseases and model operational efficiency. Li et al. (2022a) introduced attention mechanisms such as scSE and CA into the backbone network, enabling the improved network to more quickly and accurately identify and locate defect regions, with stronger generalization capabilities for defect categories and significantly improved image defect detection accuracy. SENet (Hu et al., 2018) enhances the model’s expression of key features by adding channel attention modules between convolutional layers, dynamically adjusting the importance weights of each channel (Fu et al., 2024). introduced the BiFormer attention mechanism to focus adaptively on small area features, improving the model’s detection accuracy for small targets (Redmon and Farhadi, 2018). introduced the CBAM attention mechanism, combining channel attention and spatial attention to enhance the model’s feature extraction capabilities, reducing background interference and improving model robustness.

MobileNet, through the use of depthwise separable convolution, significantly reduces the model’s parameter count and computational volume, making it more suitable for mobile devices. This is very important for practical applications in plant disease detection, as many detection tasks need to be performed in the field in real-time (Ridnik et al., 2021; Sangjan et al., 2021). Other lightweight designs include (Ma et al., 2023), who used depthwise separable convolution to significantly reduce the model’s parameter count and computational volume, making the model more suitable for mobile and embedded devices (Redmon and Farhadi, 2018). improved the detection of strawberries and peduncles through the lightweight design of the Slim-neck module, reducing the model’s computational complexity and improving operational efficiency, further promoting the deployment of models in practical applications.

In terms of model optimization through knowledge distillation and pruning, Hinton et al. (2015) transferred the knowledge of large complex models to smaller models (Hoang and Jo, 2021; Qian et al., 2021). Knowledge distillation can significantly reduce the model’s parameter count and computational volume without significantly reducing performance. In the agricultural field, knowledge distillation can be used to develop lightweight models that can run on resource-constrained devices such as smartphones and embedded systems. Han et al. (2015) used pruning techniques to remove unimportant weights or neurons from the model, further compressing the model size and improving operational efficiency (Fu et al., 2024). optimized the model by using the C2f Faster structure and CARAFE upsampling operator, reducing the model’s parameter count and computational volume while maintaining high detection accuracy. In agricultural image analysis (Ekanayake et al., 2022; Park and Kim, 2022), model pruning can reduce the demand for computational resources, enabling the model to process image data faster and improve detection real-time performance.

Data augmentation techniques (such as rotation, flipping, cropping, and color adjustment) (Ryo, 2022; Shoaib et al., 2023; Isinkaye et al., 2024) can increase the diversity of training data and improve the model’s generalization ability. In agricultural image data, data augmentation can simulate different lighting conditions, shooting angles, and disease development stages, thereby improving the model’s robustness in practical applications.

Chen et al. (2020) used pre-trained models (such as IMa, HgeNet pre-trained models) trained on large-scale datasets and fine-tuned them through transfer learning, significantly improving model performance in plant disease detection tasks. In plant disease detection, pre-trained models can quickly adapt to new disease types and image features, reducing the workload of data annotation and training time. For example, models such as VGG16, Inception V3, and ResNet50 have been fine-tuned through transfer learning (Bhatti et al., 2023; Sahu et al., 2023a).

YOLOv11 is an important variant of the YOLO series, optimized for real-time object detection tasks. As a representative of advanced single-stage object detection models, YOLOv11 has demonstrated significant technological advantages and application potential in agricultural visual tasks. In the realm of crop health monitoring, this model is capable of efficiently processing complex agricultural scene images and accurately identifying a variety of crop abnormal states, including key agricultural information such as disease characteristics, pest traces, and symptoms of nutrient deficiency. The unique lightweight network structure and multiscale feature fusion mechanism of YOLOv11 enable it to maintain high detection accuracy while adapting to the practical challenges of variable target scales and complex backgrounds commonly encountered in agricultural scenarios. Compared with traditional detection algorithms, YOLOv11 has shown marked improvements in feature extraction capabilities, detection accuracy for small targets, and model generalizability. The experimental results confirmed that the integration of the CBAM attention mechanism and CIoU loss function led to a definitive performance breakthrough, with the model attaining a mAP50–95 of 61.6% (95% CI: 60.1–63.1%). These enhancements provide an ideal framework for developing high-precision intelligent agricultural monitoring systems (Shen et al., 2025; Wang K. et al., 2025).

Detailed architectural parameters are presented in Table 1 after the model improvements.

Model performance evaluation is a core aspect of object detection tasks. To objectively quantify the performance of the proposed model, our model adopts precision (P), recall (R), mean average precision (mAP), and mAP50–95 as the primary evaluation metrics (Everingham et al., 2010; Hosang et al., 2016), defined as follows:

Precision (P): Reflects the proportion of samples predicted as positive that are truly positive, calculated as:

where TP (True Positive) is the number of true positives, and FP (False Positive) is the number of false positives.

Recall (R): Reflects the proportion of truly positive samples that are correctly predicted by the model, calculated as:

where FN (False Negative) is the number of false negatives.

Mean Average Precision (mAP): First, the average precision (AP) for a single category is calculated as the area under the precision-recall curve (PR curve):

Then, the mAP is obtained by averaging the APs of all categories:

where N is the total number of detection categories.

1) mAP50-95: The mAP is calculated for each IoU threshold from 0.5 to 0.95 (with a step size of 0.05), and the average of these mAP values is taken to comprehensively evaluate the model’s robustness under different localization accuracy requirements.

To facilitate reproducibility, the four cases in the confusion matrix are defined as follows:

Because the task is single-class, TNs are not involved in mAP but are considered when quantifying background false alarms (FP).

Priority Explanation: In object detection systems, mAP50–95 is the most comprehensive due to its coverage of multiple IoU thresholds and is prioritized as the core metric. mAP, precision, and recall are used as auxiliary analysis bases. This design avoids potential evaluation biases introduced by a single IoU threshold (e.g., mAP50).

Due to insufficient sample sizes (<50 images per subgroup) across varieties and lighting orientations, no formal subgroup analyses were performed. We therefore added a “Subgroup Analysis and Generalizability” paragraph at the end of Section V.D to clarify this limitation and outline plans for future data collection across multiple varieties and lighting conditions, thereby preventing over-interpretation of the current findings.

2) Data Annotation and Storage

The expanded dataset was uniformly stored in JSON format and manually annotated using the Labelme tool. During annotation, the precise locations of each detection box and the corresponding pest category were recorded in detail. To ensure annotation consistency, the same annotation standards and procedures were applied to the original data, direct light augmented data, and backlighting augmented data ( Figure 8 ). After annotation, the data was converted into txt format label files and stored in a designated label folder with a standardized naming convention, ensuring the accuracy and standardization of data annotation and laying a solid foundation for subsequent model training ( Figure 9 ).

4) Evaluation of Data Augmentation Effects

To assess the effects of data augmentation, model training was conducted on both the original and augmented datasets, and the results were compared. Using an NVIDIA GeForce RTX 3060 server, the model was trained with settings of batch=32, imgsz=416, epochs=200, based on YOLOv11n with a pre-trained model YOLOv11n.pt. The comparison before and after data augmentation is shown in Table 6 :

The study confirmed that the data augmentation strategy based on lighting conditions significantly improved the comprehensive performance of the YOLO model in lychee pest detection, especially in high-precision localization and difficult sample recognition( Figure 10 ). This transitioned the target detection from being “data quantity driven” to “data quality driven.” Therefore, the augmented dataset was selected for subsequent model training and optimization.

Although the YOLO-Lychee-basic model showed some improvement in mAP50–95 performance, increasing from 0.574 to 0.592, the increase was relatively limited. This means that under higher IoU thresholds, there is still considerable room for optimizing the model’s detection accuracy. As noted in I-A, the gaps are addressed below.

The traditional IoU loss function is overly sensitive to the aspect ratio of predicted bounding boxes when calculating the overlap between predicted and ground-truth boxes. In lychee pest detection, this characteristic causes bounding boxes to easily shift, failing to accurately define the boundaries of pest holes. Inaccurate bounding boxes lead to incorrect judgments of the position and size of pest holes, severely affecting the model’s localization accuracy and, consequently, the performance of the mAP50–95 metric.

To effectively address these issues and significantly enhance the model’s performance in mAP50-95, the subsequent improvements introduced the CBAM attention mechanism. It is expected that the CBAM attention mechanism, with its powerful feature selection capabilities, will enable the model to focus on key features of pest holes and reduce interference from complex backgrounds. Meanwhile, the CIoU loss function, with its more rational calculation method, will optimize the model’s localization of predicted boxes, improving the accuracy of bounding boxes and thereby comprehensively enhancing the model’s detection accuracy and performance across different IoU thresholds.

In lychee pest detection, the post-processing stage directly affects the model’s detection performance. Our module optimized the non-maximum suppression (NMS) parameters for the YOLO-Lychee-advanced model, significantly improving its performance in small target detection scenarios. Even though pest holes are not densely distributed, these optimizations are still significant, as follows:

• Reducing the IoU Threshold:

To accommodate the irregular shapes of lychee fruits and the highly variable locations and sizes of pest holes, we lowered the IoU threshold from 0.70 to 0.45—even though the holes themselves are sparsely distributed. Lowering the IoU threshold makes the model’s requirements for matching detection boxes more flexible. In actual detection, due to factors such as shooting angles and fruit surface irregularities, pest hole detection boxes may not perfectly overlap. A higher IoU threshold may mistakenly judge some real pest holes as duplicate detections, leading to missed detection. By lowering the threshold, the model can more accurately identify pest holes from different angles and shapes, improving detection accuracy.

• Fine-tuning the Confidence Threshold:

The confidence threshold was reduced from 0.25 to 0.18. Lychee stem borers cause pest holes of varying sizes, and some initial or minor infestations form pest holes with less obvious features and weaker signals. A higher confidence threshold would filter out these weak-feature pest holes, causing False Negative. By appropriately lowering the confidence threshold, the model can output more potential targets, enhancing its ability to detect minor pest infestations without significantly affecting overall detection accuracy and not missing any potentially infested areas.

• Limiting the Maximum Number of Detections per Image:

The maximum number of detections per image was decreased from 300 to 10. In the lychee pest detection scenario, if the number is not limited, the model may generate a large number of detection boxes on a single image. Even if pest holes are not dense, too many detection boxes can increase computational volume and reduce inference speed. Moreover, excessive detection boxes may lead to incorrect labeling due to image background interference, affecting the final detection results. By limiting the number, computational resources can be concentrated on truly potentially infested areas, reducing redundant calculations, improving inference speed, and enhancing detection accuracy.

• Single-class Detection Configuration:

The agnostic_nms was enabled and single_cls was set to True. Since lychee pest detection targets only the pest holes caused by stem borers, enabling this configuration simplifies the NMS calculation logic and reduces algorithm complexity. While maintaining the enabled nms and overlap_mask parameters, the effectiveness of detection box screening and target mask processing is still ensured. This allows the model to more efficiently detect and screen pest holes, improving the completeness of detection results and avoiding detection omissions or errors due to high computational complexity ( Table 11 ).

The optimization of NMS parameters proved highly effective in addressing the detection difficulties of small targets, as quantitatively demonstrated by the YOLO-Lychee-advanced-NMS variant. Its precision sharply increased to 95.5% (95% CI: 94.2–96.8%) and its mAP50–95 reached 63.2% (95% CI: 61.7–64.7%) ( Table 12 ). The higher lower bound of its CI for mAP50-95 (61.7%) compared to the upper bound of the advanced model’s CI (63.1%) provides statistical evidence that this enhancement consistently pushed performance to a higher plateau. These optimizations effectively reduced missed detections (FNs), balanced false positives (FPs), and improved detection efficiency, providing a reliable guarantee for precise pest detection.

In terms of training configuration and dataset construction, this experiment used an NVIDIA GeForce RTX 3060 graphics processor as the core computing unit, equipped with the CUDA 12.4 computing platform, and completed model development in the Python 3.8 programming environment. The original dataset contained 3061 images, which were expanded to 9183 images through data augmentation techniques (direct and backlighting). Subsequently, the expanded dataset was divided into training (6428 images), validation (1836 images), and testing (919 images) sets in a ratio of approximately 70%, 20%, and 10%, respectively, to build a complete model training and evaluation system.

During model training, the hyperparameters were deeply optimized: the learning rate was set to 0.001, the momentum parameter to 0.937, the weight decay coefficient to 0.0005, the batch size to 32, and the input image size to 416×416 pixels. Table 13 shows the performance comparison. The model was trained for 200 epochs. This parameter combination balanced training efficiency and model generalization capabilities, laying a solid foundation for the reliability and effectiveness of the training results, as follows:

1) Comparison of Original and Improved Models ( Table 14 ).

FPS was measured at input resolution 416×416 with batch=1 on RTX 3060. Error bars represent 95% bootstrap confidence intervals.

In data processing, the strategy of annotating only the core area of pest holes was selected, combined with data augmentation techniques based on simulated lighting conditions. The annotation strategy improved the mAP50–95 metric (by 18.0%), enhancing the model’s focus on pest features. Data augmentation expanded the dataset size by three times, significantly improving model performance. Precision (P), recall (R), mAP50, and mAP50–95 increased by 4.5%, 7.3%, 6.1%, and 10.9%, respectively, enhancing the model’s adaptability to complex lighting conditions.

In terms of model improvements, the YOLO-Lychee-basic model introduced the C2f module to optimize the structure, increasing P by 2.5%, R by 0.24%, and mAP50 by 0.55%, strengthening feature processing capabilities. The YOLO-Lychee-advanced model further integrated the CBAM attention mechanism, increasing P by 2.56%, mAP50 by 1.10%, and mAP50–95 by 3.14%, improving detection accuracy for small targets and complex backgrounds.

In the post-processing stage, the NMS parameters of the YOLO-Lychee-advanced model were optimized, increasing P by 3.3% and mAP50–95 by 2.4%, reducing False Negative, balancing misdetections, and improving detection efficiency ( Figure 17 ).

After a series of optimizations, the model’s performance was significantly enhanced, providing an effective solution for precise lychee pest detection and offering scientific basis and technical references for research in the field of agricultural pest detection, promoting the application of related technologies in practical production.

Furthermore, to quantify the practical significance of the model improvements, we computed Cohen’s d for the difference in mAP50–95 between the YOLO-Lychee-advanced model and the baseline models (YOLOv9t and YOLOv10n). The effect sizes were d = 1.21 (vs. YOLOv9t) and d = 0.89 (vs. YOLOv10n). This result (where d > 0.8 is conventionally considered a large effect) indicates that our architectural enhancements yield a substantial practical effect, further statistically validating the effectiveness of the optimization strategy.

2) Performance Comparison between YOLO-Lychee-advanced and Various YOLO Versions ( Table 12 )

In the key research area of lychee stem borer recognition, YOLO series models have demonstrated significant value. Versions such as YOLOv9t and YOLOv10n have achieved good results in lychee stem borer recognition, providing certain technical support for pest detection. However, to further improve detection accuracy and efficiency, the YOLO-Lychee-advanced model was carefully developed based on YOLOv11. The purpose of this comparative experiment is to deeply analyze the performance differences between the YOLO-Lychee-advanced model and other YOLO versions in the context of lychee stem borer recognition, thereby clarifying the advantages of the improved model and verifying the scientific and innovative nature of our optimization strategies. As comprehensively summarized in Table 12 , our YOLO-Lychee-advanced model achieved a superior mAP50–95 of 61.6% (95% CI: 60.1–63.1%), outperforming both YOLOv9t (58.2%, 95% CI: 56.8–59.6%) and YOLOv10n (59.9%, 95% CI: 58.5–61.3%). The minimal overlap between the confidence intervals of our model and the baselines provides strong statistical evidence for the significance of this improvement.

In terms of recall (R), the values of the various models are relatively close. Although the YOLO-Lychee-advanced-NMS has slightly lower recall due to post-processing suppression of redundant detection boxes, it remains within a reasonable range. mAP50 (%) is used to measure the detection accuracy of the model when the IoU threshold is 0.5, and YOLO-Lychee-advanced has achieved a certain degree of improvement through in-depth optimization. mAP50-95 (%) comprehensively reflects the model’s average precision at different IoU thresholds (0.5 - 0.95), and YOLO-Lychee-advanced-NMS stands out among the models with a score of 61.6%, demonstrating its superior comprehensive capabilities in different strict IoU thresholds for target boundary localization of lychee stem borers ( Figure 18 ).

The comparative results across multiple metrics provide preliminary evidence supporting the effectiveness and potential innovation of the incremental improvements introduced in the YOLOv11 framework, as the YOLO-Lychee-advanced model generally outperforms other YOLO variants in lychee stem borer recognition. These findings may offer insights for future research directions and model optimization aimed at improving the precise detection and control of lychee stem borers ( Figure 19 ).

1) System Architecture and Core Functionalities

We have developed a cross-platform intelligent detection system for lychee stem borers (LSBVS), which deeply integrates deep learning-based object detection technology with the PyQt5 graphical interface framework ( Figure 20 ). The core of the system lies in the visualization of the detection process and results, with specific functionalities including:

• Multisource Input Visualization Processing:

Supports input from static images, video streams, and real-time cameras, and clearly displays the original images within the interface.

• Real-time Detection Result Visualization:

Utilizes a dual-view comparative interface (original image vs. detection result image) to highlight and annotate the detected lychee stem borer targets (bounding boxes) in real time ( Figure 21 ).

• Dynamic Model Loading and Resource Visualization Feedback:

Users can load custom models (in *.pt format) through the interface. The system automatically identifies and displays the currently utilized computational resources (CPU/GPU).

• Batch Data Analysis Visualization:

Supports batch detection of image folders, automatically generates Excel reports containing detection results, and visualizes statistical charts (e.g., histograms of pest distribution).

• Visualization Optimization of Interaction Processes:

Enhances operational intuitiveness and user experience through visual designs such as image transition animations (fade-in and fade-out) and immediate feedback on button states.

2) Implementation of Key Technologies

• Multimodal Input Visualization Pipeline:

A unified interface is designed to process various input sources, ensuring that the original images and detection results are visualized smoothly and synchronously within the interface.

• Static Images:

Display the original image alongside the annotated result image.

• Video Streams/Cameras:

Real-time display of processed video frames with detection result annotations.

• Efficient Visualization Rendering:

OpenCV is utilized for image processing (annotation), and the detection results are efficiently displayed in the Qt interface with adaptive scaling through QImage/QPixmap, while maintaining the aspect ratio.

• Data Visualization and Management:

After detection, statistical charts (e.g., pest distribution) are generated and visualized. The system supports exporting and saving these charts along with structured detection reports (in Excel format) in various formats (PNG/JPEG/Excel), facilitating result viewing and analysis ( Figure 22 ).

3) Innovations and Contributions

• Multimodal Visualization Detection Framework for Agricultural Scenarios:

We realizes the deep integration of deep learning-based detection and cross-platform graphical user interfaces (GUIs), constructing a closed-loop visualization detection process that covers images, videos, and real-time cameras. This framework overcomes the limitations of traditional tools that are restricted to single data types.

• Lightweight and Smooth Visualization Interaction Experience:

By combining progressive animations with multithreading technology, we provides a smooth and low-fatigue visualization operation interface while ensuring real-time processing capabilities. The system also supports offline usage.

• End-to-End Visualization Decision Support:

Beyond offering intuitive visualizations of pest target annotations, we further assists users in intuitively identifying pest distribution patterns through batch result statistical charts. This visual basis supports decision-making processes.

Lychee stem borer causes >60% yield loss and chemical residues; an accurate yet lightweight detection tool is therefore urgently needed.

This paper focuses on the problem of lychee pest detection and conducts gradual optimization research based on the YOLOv11 model, achieving a series of important results. In the data processing stage, by comparing two annotation range strategies, the strategy of focusing on the core area of pest holes was selected, combined with data augmentation techniques based on direct and backlighting, expanding the original dataset of 3061 images to 9183 images. Based on the YOLOv11 model, this data augmentation method, without the need for complex modifications to the model architecture, has achieved significant improvements in model performance through a low-cost data expansion approach, fully verifying that data augmentation can be an efficient and low-cost solution for improving YOLO model performance in resource-constrained scenarios.

In terms of model construction, the YOLO-Lychee-basic model was first proposed based on YOLOv11. By adjusting the main structure of the backbone network, such as module replacement and optimization of stacking layers, the model’s feature extraction and fusion capabilities were enhanced, resulting in improvements in precision, recall, and mAP50 metrics. Compared with two recent YOLO baselines (YOLOv9t and YOLOv10n) on the same lychee test set, YOLO-Lychee-advanced raises mAP50–95 from 58.2% → 61.6% (+3.4%) and 59.9% → 61.6% (+1.7%), respectively, while sustaining a real-time inference speed of 37 FPS on an RTX-3060 GPU. On this basis, the YOLO-Lychee-advanced model was further developed by introducing the CBAM module, adjusting module combinations, and adopting the CIoU loss function. These in-depth optimization strategies significantly enhanced the model’s ability to capture key features of tiny pest targets, resulting in excellent performance in key metrics such as mAP50-95.In conclusion, the YOLO-Lychee-advanced model significantly raises the bar for lychee stem borer detection, achieving a state-of-the-art mAP50–95 of 61.6% (95% CI: 60.1–63.1%) — a statistically significant improvement of 3.4 and 1.7 percentage points over YOLOv9t (58.2%, 95% CI: 56.8–59.6%) and YOLOv10n (59.9%, 95% CI: 58.5–61.3%), respectively. After post-processing optimization, the precision was further boosted to 95.5% (95% CI: 94.2–96.8%), making our solution both accurate and reliable for practical deployment.

Finally, post-processing optimization was performed on the YOLO-Lychee-advanced model by carefully adjusting NMS-related parameters, such as reducing the IoU threshold and fine-tuning the confidence threshold, resulting in the YOLO-Lychee-advanced-NMS model. This model achieved significant improvements in precision and mAP50–95 metrics. Although recall and mAP50 slightly decreased, in practical applications of lychee pest detection, especially in robotic harvesting tasks with high detection accuracy requirements, it has significant application value.

Compared with other versions of the YOLO series, our model improved based on YOLOv11 has shown clear advantages in key performance metrics such as precision and mAP50-95, verifying the effectiveness and innovativeness of the gradual optimization strategy. In the future, more optimization solutions will be continuously explored, such as further research on dynamic adaptive data augmentation strategies to enhance the model’s robustness in complex and changing orchard environments; in-depth exploration of model lightweighting techniques to promote efficient deployment of the model on edge devices, providing stronger and more convenient technical support for the intelligent pest control of the lychee industry (Wang et al., 2025; Zhang et al., 2025).

Future work will address dynamic illumination and model compression, while current limits remain modest data, controlled lighting, single-class scope, and regional validation.

Additional Objective Metrics( Table 10 ).