Breaking the camouflage: a transformer-based approach for soybean pest detection in complex field environments

Accurate soybean pest detection is essential for agricultural management, yet it remains challenging because of small target size, strong camouflage, and severe foliage occlusion. Traditional convolutional neural networks (CNNs) may exhibit limited robustness under these conditions. To address these challenges, we propose an enhanced RT-DETR-ResNet18 framework that incorporates three targeted components. First, the Global Semantic Enhancement Encoding Module (CFBC2f) integrates multi-scale contextual modeling to improve semantic representation under occlusion. Second, the Local Region Self-Attention AIFI (LRSAAIFI) module emphasizes fine-grained local textures to reduce feature dilution for small targets. Third, the Multi-source Feature Adaptive Modulation Module (MFM) adaptively reweights features to reduce background interference. Experiments across 12 soybean pest categories show that our method achieves 77.4% mAP@0.5 and a 79.9% F1-score, outperforming the evaluated YOLOv11n and YOLOv12s baselines under the reported protocol. Notably, AP@0.5 for the highly camouflaged Nezara viridula adult increases from 77.7% to 91.8%, corresponding to an absolute improvement of 14.1 percentage points over YOLOv12s. Furthermore, results on PASCAL VOC and VisDrone provide supplementary evidence that the proposed architectural modifications can transfer beyond the INSECT12C benchmark; however, these experiments should be interpreted as evidence of transferability under the evaluated settings rather than proof of universal cross-domain generalization.