<p>Agricultural pest detection is critical for ensuring food security and minimizing crop losses. However, accurate detection in field environments remains challenging due to extreme scale variation, dense overlapping instances, high inter-class similarity, and complex backgrounds. To address these challenges, we propose AEP-PEST, a lightweight and efficient object detection model based on the YOLOv11 framework. Our improvements are threefold, focusing on feature enhancement, relational reasoning, and adaptive optimization. First, a Lightweight Frequency and Edge Enhancement Fusion module combined with a Multi-branch Dilated Spatial Pyramid Pooling Fusion module is integrated into the backbone to preserve high-frequency details and capture multi-scale contextual information. Second, a Cross-scale Hypergraph Pyramid Fusion network is introduced to capture high-order dependencies among densely packed pests, effectively reducing false detections caused by overlapping instances. Third, a Scale-Quality Aware EIoU loss dynamically adjusts sample weights based on target size and prediction quality, focusing optimization on challenging small pests. Extensive experiments on the large-scale Pest24 dataset demonstrate that AEP-PEST achieves 75.2% mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation>, outperforming state-of-the-art methods including Litepest (74.1%) and Hyper-YOLO (74.0%). This model maintains a compact architecture with only 24.3M parameters and 82.5 GFLOPs, achieving an inference speed of 147.3 FPS. Compared to the YOLOv11-L baseline, AEP-PEST improves mAP<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> by 2.8% while reducing parameters by 4.0% and computational cost by 5.6%. Additional cross-dataset validation on the IP102 detect task further demonstrates its generalization ability, where AEP-PEST achieves the best mAP<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> of 60.1% and the highest recall of 61.8% among the compared methods.</p>

AEP-PEST: adaptive edge-enhanced pyramid network for real-time detection of small and occluded agricultural pests

期刊论文
发表日期：

2026年5月21日

查看全文

Zhe Tang,
Yusen Wang,
Dengpeng Zou,
Fang Qi

摘要

Agricultural pest detection is critical for ensuring food security and minimizing crop losses. However, accurate detection in field environments remains challenging due to extreme scale variation, dense overlapping instances, high inter-class similarity, and complex backgrounds. To address these challenges, we propose AEP-PEST, a lightweight and efficient object detection model based on the YOLOv11 framework. Our improvements are threefold, focusing on feature enhancement, relational reasoning, and adaptive optimization. First, a Lightweight Frequency and Edge Enhancement Fusion module combined with a Multi-branch Dilated Spatial Pyramid Pooling Fusion module is integrated into the backbone to preserve high-frequency details and capture multi-scale contextual information. Second, a Cross-scale Hypergraph Pyramid Fusion network is introduced to capture high-order dependencies among densely packed pests, effectively reducing false detections caused by overlapping instances. Third, a Scale-Quality Aware EIoU loss dynamically adjusts sample weights based on target size and prediction quality, focusing optimization on challenging small pests. Extensive experiments on the large-scale Pest24 dataset demonstrate that AEP-PEST achieves 75.2% mAP $_{50}$ $_{50}$ , outperforming state-of-the-art methods including Litepest (74.1%) and Hyper-YOLO (74.0%). This model maintains a compact architecture with only 24.3M parameters and 82.5 GFLOPs, achieving an inference speed of 147.3 FPS. Compared to the YOLOv11-L baseline, AEP-PEST improves mAP $_{50}$ $_{50}$ by 2.8% while reducing parameters by 4.0% and computational cost by 5.6%. Additional cross-dataset validation on the IP102 detect task further demonstrates its generalization ability, where AEP-PEST achieves the best mAP $_{50}$ $_{50}$ of 60.1% and the highest recall of 61.8% among the compared methods.