Stage-wise collaborative feature enhancement for accurate maize insect pest detection in complex agricultural scenes
Abstract
IntroductionAccurate maize pest detection is hindered by morphological similarity among pest categories, complex backgrounds, and the loss of spatial detail in small targetsMethodsTo address these challenges, we developed a stage-wise collaborative feature enhancement framework based on RT-DETR-R18. A Similarity-aware Decoupling Module combines orthogonal feature transformation with channel excitation to enhance discriminative backbone features. Adaptive Sparse Self-Attention regulates high-level inter-region interactions, whereas an Efficient Up-Convolution Block recovers spatial detail during top-down feature fusion. The method was evaluated on a primary dataset comprising 2,311 images from four maize pest categories and on the separately constructed IP102-Maize dataset, which was derived from the public IP102 benchmark and comprised 4,533 images from 13 categories.ResultsOn the primary dataset, the proposed method achieved 90.4% mAP@50 and 67.9% mAP@50:95, improving mAP@50 by 2.7 percentage points over RT-DETR-R18. The model contained 20.9 M parameters and required 58.9 GFLOPs. On IP102-Maize, it achieved 64.1% mAP@50, exceeding RT-DETR-R18, RT-DETR-R34, and YOLOv10-S by 2.7, 4.0, and 1.7 percentage points, respectively.DiscussionThese results support the effectiveness of the tested stage-wise enhancement configuration for maize pest detection in two public-dataset settings. Further field and deployment evaluation is required to establish its practical applicability.
Related
- FSF-RTDETR: a frequency–spatial feature fusion network for eggplant fruit disease and pest detection in complex agricultural scenes
- Enhancing medical image segmentation with HCE-Net: A heterogeneous feature collaborative extraction network
- DAF-EfficientNetV2: distribution-aware feature enhancement for plant leaf disease recognition in complex environments
- MEF-TBN: Multi-exposure image fusion approach based on a three-branch network for feature extraction and color enhancement
- Explainable fine-grained feature enhancement framework for strawberry diagnostic decision making
BibTeX
@article{zuo2026stagewise,
title = {{Stage-wise collaborative feature enhancement for accurate maize insect pest detection in complex agricultural scenes}},
author = {Jiayi Zuo and Cong Zhang and Donghan Liu and Zheng Wang},
journal = {Frontiers in Plant Science},
year = {2026},
month = nov,
doi = {10.3389/fpls.2026.2011789},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.2011789},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Stage-wise collaborative feature enhancement for accurate maize insect pest detection in complex agricultural scenes
AU - Jiayi Zuo
AU - Cong Zhang
AU - Donghan Liu
AU - Zheng Wang
PY - 2026
DA - 2026/11/02
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.2011789
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.2011789
AB - IntroductionAccurate maize pest detection is hindered by morphological similarity among pest categories, complex backgrounds, and the loss of spatial detail in small targetsMethodsTo address these challenges, we developed a stage-wise collaborative feature enhancement framework based on RT-DETR-R18. A Similarity-aware Decoupling Module combines orthogonal feature transformation with channel excitation to enhance discriminative backbone features. Adaptive Sparse Self-Attention regulates high-level inter-region interactions, whereas an Efficient Up-Convolution Block recovers spatial detail during top-down feature fusion. The method was evaluated on a primary dataset comprising 2,311 images from four maize pest categories and on the separately constructed IP102-Maize dataset, which was derived from the public IP102 benchmark and comprised 4,533 images from 13 categories.ResultsOn the primary dataset, the proposed method achieved 90.4% mAP@50 and 67.9% mAP@50:95, improving mAP@50 by 2.7 percentage points over RT-DETR-R18. The model contained 20.9 M parameters and required 58.9 GFLOPs. On IP102-Maize, it achieved 64.1% mAP@50, exceeding RT-DETR-R18, RT-DETR-R34, and YOLOv10-S by 2.7, 4.0, and 1.7 percentage points, respectively.DiscussionThese results support the effectiveness of the tested stage-wise enhancement configuration for maize pest detection in two public-dataset settings. Further field and deployment evaluation is required to establish its practical applicability.
SN - 1664-462X
ER -