This study utilizes the IP102 dataset as the primary data source for experimentation. As the first large-scale benchmark dataset dedicated to crop pest detection in agricultural computer vision, IP102, a dataset specifically designed for agricultural pest detection, is extensively utilized across various research domains, including plant image recognition and pest management. The IP102 dataset, for instance, encompasses a wide range of crops such as strawberries, beans, and tomatoes, and includes various conditions like mold and leaf-spot diseases. Unlike smaller datasets, IP102’s key advantage is that all images were captured in real field environments, exhibiting complex background clutter, partial occlusions, and overlapping leaves. The dataset’s intrinsic characteristics closely replicate the real-world challenges encountered by detection algorithms in grassroots agricultural monitoring, rendering it ideal for assessing the robustness of proposed models, particularly their capacity to detect small and inconspicuous targets.

To enhance the experiment’s relevance, this study manually selected images from the IP102 dataset and chose three crops—bean, strawberry, and tomato—for evaluation. This study retained 5,483 annotated images representing 12 pest and disease types. Typical examples are shown in Figure 1.

The detailed distribution of samples is as follows:

Strawberry samples included 7 pest and disease categories: “Angular Leaf Spot” (541 images), “Anthracnose Fruit Rot” (194), “Blossom Blight” (356 images), “Gray Mold” (539 images), “Leaf Spot” (650 images), “Powdery Mildew Fruit” (258 images), and “Powdery Mildew Leaf” (558 images). Tomato samples included 3 categories: “Tomato Blight” (531 images), “Leaf Mold” (477 images), and”Spider Mites” (446 images). Bean samples included 2 categories: “Angular Leaf Spot” (489 images) and “Bean Rust” (414 images). The detailed images of the 12 selected pests and diseases are shown in Table 1.

To prevent bias from data partitioning and ensure reproducibility, the dataset was randomly divided into training, validation, and test sets in a 7:1:2 ratio. The training set contains 3,838 images used to learn model parameters, the validation set contains 548 images used for hyperparameter tuning and periodic performance checks during training, and the test set contains 1,097 images reserved for independent evaluation of the final model’s detection performance.

YOLOv5, a typical single-stage object detector under the YOLO family, simplifies object detection into an end-to-end regression task. This allows it to predict both the location and category of objects in one single forward pass. YOLOv5 was selected as the baseline architecture due to its proven efficiency on

embedded platforms and well-established implementation ecosystem. While RT-DETR offers advantages in certain scenarios, YOLOv5’s balance of accuracy, speed, and resource efficiency makes it more suitable for agricultural edge deployment. Such a structural design cuts down computational overhead markedly and achieves high inference speed, making it a promising option for real-time detection on resource-limited embedded platforms—well-aligned with the on-site monitoring needs of agricultural micropest detection. As illustrated in Figure 2, the YOLOv5 architecture comprises four core modules: input module, Backbone network, Neck network, and detection Head.

For multi-scale feature extraction, the Backbone employs CSPDarknet53 to extract multi-scale features 156 from the input images. The input image (640×640) is downsampled via convolutional layers to yield three 157 feature maps at different scales: P3, P4, and P5. P3 is a high-resolution feature map with rich spatial detail 158 and is best suited for detecting small objects, whereas P4 and P5 have lower spatial resolution but contain 159 more abstract semantic information and are therefore better for medium and large objects.

The Neck module integrates Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) to realize multi-scale feature fusion: FPN transmits high-level semantic information from top to bottom to lower layers, while PAN conveys low-level precise localization information from bottom to top. This bidirectional aggregation approach merges contextual and spatial information across different scales, thereby enhancing the detection capability for objects of various sizes. Nevertheless, in practical agricultural micropest detection scenarios, the original YOLOv5 still has certain limitations: the feature fusion in the Neck module lacks targeted enhancement for small targets, and the computational cost of the backbone network still needs optimization to adapt to long-term stable operation on embedded devices (e.g., Raspberry Pi 4B). For the final prediction, an anchor-based strategy is adopted: each feature map is preconfigured with fixed-size anchor boxes, and the network predicts bounding-box coordinate offsets, confidence scores, and class probabilities via convolution operations. Subsequently, non-maximum suppression (NMS) is used for post-processing to remove redundant detections and obtain the final results.

Although YOLOv5 can achieve efficient real-time detection, in actual agricultural scenarios, due to the complex background and the small size of pest and disease targets, it still encounters some challenges. To overcome these limitations and maintain computational efficiency, we proposed the F-HAIN method and made targeted improvements to it. To address this limitation, this paper proposes a novel model, F-HAIN, whose network architecture is illustrated in Figure 3. The model replaces the core SPPF module in YOLOv5 with a newly designed F-SPPELAN module, which enhances multi-scale target feature expression through parallel multi-scale pooling and dynamic feature weighting while preserving real-time performance. Additionally, F-HAIN incorporates the efficient hybrid encoder’s HAIN module, which processes high-level backbone features. By implementing global interaction on feature maps via a multi-head attention mechanism, the HAIN module reduces computational overhead and improves processing speed. These architectural modifications collectively mitigate critical challenges in complex scenarios, including multi-target occlusion, small target missed detection, and insufficient detection speed.

To clarify the structural novelty of Focal-HAIN, we contrast its design rationale with incremental YOLO neck variants through two fundamental paradigm shifts tailored for agricultural monitoring.

(1) From Static Pooling to Focal-Aware Modulation: Unlike recent YOLO models (e.g., YOLOv9 to v11) that utilize static hierarchical pooling (such as SPPELAN) to expand receptive fields, our F-SPPELAN implements a focal-aware spatial modulation mechanism. While static pooling often leads to the loss of fine-grained textures in micro-pests, Focal Modulation adaptively weights spatial contexts. This allows the neck to selectively amplify pest-related features while attenuating high-frequency environmental noise like soil and leaf veins.

(2) From Sequential Fusion to Deep Interlaced Interaction: As summarized in Table 2, the HAIN module distinguishes itself from existing FPN/PAN variants through its structural positioning. While traditional adaptive-weighting structures primarily rely on simple learnable scalars for cross-scale summation, HAIN introduces a Deep Interlaced Fusion strategy. This strategy embeds Intra-scale Adaptive Interaction (AIFI) directly into the Hierarchical Feature Aggregation (HIFA) pathway. By doing so, it ensures that semantic refinement and multi-scale fusion occur simultaneously, effectively mitigating the semantic inconsistency often encountered when detecting micro-scale pests against chaotic agricultural backgrounds.

All training and evaluations used a single, consistent parameter set. Experiments ran on an Intel(R) Core(TM) i9-14900K (24 cores) with CUDA 12.6 and an NVIDIA RTX 4090 GPU (24GB VRAM), using the PyTorch 2.0.1 framework on Windows 11. Models were trained from scratch without using any pre-trained weights to ensure that performance gains were solely attributable to architectural innovations. Hyperparameters were: batch size 16, total epochs 300, and an initial learning rate of 1×10⁻⁴, with image 307 resolution standardized to 640×640. To enhance data diversity, we applied standard data augmentation strategies, including Mosaic, Mixup, and random horizontal flipping.

Considering the structural differences in the Neck between Focal-HAIN and the baseline models, a specialized learning rate protocol was adopted: a Warm-up strategy (first 3 epochs) was used to stabilize initial gradients, followed by a Cosine Annealing scheduler (minimum learning rate of 1 × 10⁻⁶ to ensure convergence. During the inference phase, we set the Non-Maximum Suppression (NMS) IoU threshold at 0.45 and the confidence threshold at 0.25. GIoU loss was applied uniformly. This standardized configuration demonstrates that the improvements in Focal-HAIN are a direct result of architectural innovations rather than hyperparameter bias.

To assess model performance across categories, we visualized the predictions and computed a normalized confusion matrix. The normalized confusion matrix reports, for each category, the proportion of correct and incorrect predictions, thereby removing the effects of class imbalance and more intuitively conveying classification performance. Figure 6 presents the model’s normalized confusion matrix on the data. Overall, the model performs well in most categories, as evidenced by the high diagonal values. However, a granular analysis of the off-diagonal elements reveals two primary failure patterns. First, inter-class misclassification is observed among visually similar diseases on the same host, such as Strawberry Leaf Spot (S-LS) and Strawberry Angular Leaf Spot (S-ALS), due to their overlapping color distributions and lesion morphologies. Second, background confusion is most noticeable in Beans Angular Leaf Spot (B-ALS) and Tomato Spider Mites (T-SM). For B-ALS, the sharp geometric edges of the lesions are occasionally confused with fragmented ground textures or soil shadows in the background. In the case of T-SM, the micro-scale, granular appearance of mites against the leaf veins poses a significant challenge for distinguishing target features from high-frequency background noise.

Figure 7 presents distributions of bounding-box labels. The top-left bar chart reports the instance count per label category, indicating the categorical distribution. The top-right plot maps bounding boxes in coordinate space: the central light-cyan cluster denotes many small, concentrated boxes, whereas the peripheral purple frames denote larger boxes, indicating hierarchical size variation. The bottom-left scatter of (x, y) coordinates shows a dense cluster near (0.4, 0.6), indicating most objects lie in the central region. The bottom-right (width, height) heatmap reveals a strong positive correlation: small widths align with small heights (dark cluster at low values) and larger widths align with larger heights, indicating bounding boxes in this dataset generally preserve relatively stable aspect ratios.

Figure 8 assesses model performance via three curves: precision-recall (a), precision-confidence (b), and recall-confidence (c). In Figure 8a classes like S-BB, T-LM and T-SM maintain around 0.9 precision across wide recall ranges, reflecting robust feature learning, while B-ALS underperforms (0.654) likely due to insufficient samples or high intra-class variability. The all classes curve yields an mAP50 of 0.899, verifying strong cross-category generalization. Linking subplots, Figure 8b shows precision surges with confidence and plateaus near 1.0, hitting perfect precision at a 0.983 threshold—proving high-confidence predictions are error-free. Figure 8c reveals a precision-recall trade-off: recall peaks at 0.96 at low confidence but drops sharply above 0.8. Collectively, these curves guide practical deployment: low thresholds maximize capture, while 0.983 filters noise for reliable detections.

Figure 9 depicts the 300-epoch training dynamics of the F-HAIN model, with raw (blue) and smoothed (orange) curves tracking both loss metrics and detection performance. During the first 50 epochs, all loss curves—including train/box loss, train/cls loss, train/dfl loss and validation losses—plummet sharply, then stabilize at 0.5–1.0; the tight alignment between training and validation losses clearly rules out overfitting, a key advantage for model generalization. In parallel, performance metrics show steady learning progress: metrics/precision(B) and metrics/recall(B) rise rapidly in the initial 50 epochs and plateau at 0.8-0.9, while metrics/mAP50(B) hits 0.85 by epoch 100 and peaks near 0.9. Notably, the more rigorous metrics/mAP50-95(B) stabilizes at 0.6-0.7, implying the model still has room to improve on hard targets. Collectively, these trends confirm F-HAIN’s fast convergence, stable training process and robust detection capability across target difficulty levels.

To systematically evaluate the contribution of each proposed module, we conducted ablation experiments based on the YOLOv5 baseline model. The ablation study has followed a systematic, incremental design to isolate the contribution of each proposed module, where we have defined three experimental groups: the Baseline (original YOLOv5 model without any modifications), single-module variants (Model A: YOLOv5 integrated only with the F-SPPELAN module; Model B: YOLOv5 integrated only with the HAIN module), and the Complete model (F-HAIN, which has incorporated both F-SPPELAN and HAIN modules). This incremental design has enabled us to quantify the individual contribution of each proposed module, assess potential synergistic effects between the two modules, and maintain experimental consistency by adopting YOLOv5 as the unified baseline across all experimental groups. The specific data are shown in Table 3.

The ablation results systematically demonstrate the contribution of each proposed module, linking specific structural behaviors to performance gains. The original YOLOv5 baseline achieved a mAP50 of 84.3%. Model A (YOLOv5 + F-SPPELAN) incorporates the F-SPPELAN module, which raises mAP50 to 87.1% (Δ+2.8). This improvement stems from F-SPPELAN’s focal-aware spatial modulation behavior, which adaptively suppresses high-frequency environmental noise (e.g., soil and leaf textures) while amplifying pest-related spatial contexts. This targeted modulation ensures higher recall for targets in cluttered backgrounds with only a marginal GFLOPs increase (+0.2G). Model B (YOLOv5 + HAIN) introduces the HAIN module, raising mAP50 to 87.8% (Δ+3.5). The gain is attributed to HAIN’s deep interlaced interaction behavior, where Intra-scale Adaptive Interaction (AIFI) eliminates the semantic gap between feature levels during the aggregation process. While the self-attention-based Query-Key-Value (QKV) interactions increase computational complexity (+0.7G), they provide the necessary global context to resolve semantic inconsistencies in micro-pest detection. The complete Focal-HAIN model achieves a mAP50 of 90.1% (Δ+5.8). This synergy demonstrates that the model effectively balances localized focal refinement with hierarchical global interaction, maintaining high inference speed (161.29 FPS) while ensuring the precision required for embedded agricultural deployment.

To evaluate the statistical significance and stability of the F-HAIN model, we conducted six independent runs of the F-HAIN model using different random seeds under the same hardware and hyperparameter settings. The final average mAP50 of the model was 89.73%, with a standard deviation of ±0.47%. Although random factors during the training process may introduce slight fluctuations, the performance of the model remained within a high-precision range. This variance analysis confirmed that the proposed architectural improvements have extremely high robustness and training stability.

To further verify the validity of the proposed model, various performance indicators of the model were compared with those of other models, including RT-DETR, YOLOv5, YOLOv8, YOLOv10, and YOLOv11. The specific data are shown in Table 4.

Compared with RT-DETR, YOLOv5, YOLOv8, YOLOv10 and YOLOv11, the mAP50 of the improved model increased by 5.3%, 6.6%, 4.7%, 5.2% and 2.8% respectively. Meanwhile, compared with these comparison models, the mAP50:95 of the improved model increased by 7.7%, 6.8%, 8.4%, 6.9% and 8.2% respectively. These results indicate that our model outperforms the compared detectors for crop pest and disease detection in natural environments.

Controlled experiments were conducted on RT-DETR, YOLOv5, YOLOv8, YOLOv10, YOLOv11, and F-HAIN under identical datasets, training epochs, and hyperparameters to guarantee a fair comparison. Training logs were parsed to plot the comprehensive performance metrics, as illustrated in Figure 10.

Specifically, the two upper panels display the mAP50 and mAP50:95 curves throughout the training process, where yellow, green, red, and blue curves correspond to YOLOv5, YOLOv8, RT-DETR, and F-HAIN, respectively. To further evaluate the models under specific agricultural constraints, granular testing was performed on targeted disease subsets, as shown in the bottom panels of Figure 10. The lower-left bar chart presents the mAPsmall results, which were evaluated using 806 strawberry samples across five categories (S-AFR, S-BB, S-GM, S-PMF, and S-PML) where pest instances are characterized by micro-scale dimensions. The lower-right bar chart illustrates the False Discovery Rate (FDR), calculated on 518 soybean samples (B-ALS and B-R) featuring complex backgrounds such as soil and leaf interference.

Quantitative results demonstrate that F-HAIN outperforms all comparative models by notable margins across all four metrics, achieving the highest mAPsmall (82.6%) and the lowest FDR (5.6%). These performance gains are directly attributed to the HAIN and F-SPPELAN modules. Unlike conventional attention-based or Transformer-based models that often suffer from over-smoothing of fine textures, the Focal Modulation mechanism within our architecture preserves sharper boundary information through hierarchical gated aggregation. This spatial awareness allows F-HAIN to effectively enhance micro-scale feature representation for strawberry pests while suppressing non-target environmental noise in complex soybean field backgrounds, providing a more robust solution than recent YOLO variants.

To evaluate the statistical significance and stability of the Focal-HAIN model, we conducted three independent training runs using different random seeds. The model achieved a mean mAP@0.5 of 86.10% with a standard deviation of ±0.87%. Although stochastic factors in training, such as data augmentation and weight initialization, introduce minor fluctuations, the performance consistently remains within a high-accuracy range. This variance analysis confirms that the proposed architectural improvements provide robust and reproducible gains over the baseline models.

Specifically, yellow, green, red and blue curves correspond to YOLOv5, YOLOv8, RT-DETR and F-HAIN, respectively. Quantitative results demonstrate that F-HAIN outperforms all comparative models by notable margins in both metrics, with the performance gain attributed to its F-SPPELAN and HAIN modules that strengthen multi-scale feature representation and localization precision.

Figure 11 presents detection results for four pest and disease types affecting beans and tomatoes. The figure shows that F-HAIN achieves higher confidence scores than the other three models. Detection accuracy is also improved: for example, F-HAIN identified areas in B-ALS and B-R cases that the other models missed.

Figure 12 presents detection results for four strawberry diseases and pests. The F-HAIN detections exhibit higher confidence than those produced by the other three models. F-HAIN also yields fewer false positives for the S-LS and S-PML categories compared with the other three models.

In order to compare the performance differences of the three models in the object detection task more clearly, the EigenGradCAM heat map method is adopted to conduct a visual analysis of the three diseases and pests. The specific results are shown in Figure 13.

The heat-map graphs show that YOLOv5 and YOLOv8 produce widely dispersed response patterns, particularly near object boundaries. RT-DETR’s feature extraction within target areas is limited, which can lead to false detections; for example, in the third type there is a distinct thermal pattern beneath the petals. By contrast, the improved model yields broader high-response regions centered on key targets and shows greater robustness in edge and other non-target areas. These results indicate that the improved model captures targets more completely and thus reduces missed detections.