CEAM-DETR: An NMS-free lightweight transformer for weed detection in soybean fields under complex conditions
Cheng Zhang1, Jianyu Xiao2, Yiqun Chang1
1School of Computer Science and Technology, Huaibei Normal University, Huaibei, 235000, China.
Scientific Reports
|May 23, 2026
Summary
This study introduces CEAM-DETR, a lightweight transformer model for efficient and accurate weed detection in agriculture. It overcomes limitations of traditional methods, improving performance in challenging conditions without non-maximum suppression (NMS).
Area of Science:
- Agricultural Engineering
- Computer Vision
- Machine Learning
Background:
- Weed detection in agriculture is crucial but challenging due to illumination variations, crop-weed similarity, and background interference.
- Existing object detection methods often rely on non-maximum suppression (NMS), which suffers from threshold sensitivity, poor adaptability, and inference latency.
- These limitations hinder accurate and real-time weed detection in complex agricultural settings.
Purpose of the Study:
- To propose a novel, lightweight, adaptive transformer-based model (CEAM-DETR) for efficient and stable weed detection.
- To develop an end-to-end detection paradigm that eliminates the need for non-maximum suppression (NMS).
- To enhance detection accuracy and real-time performance in complex agricultural environments.
Main Methods:
- Developed a cross-stage efficient attention backbone integrating partial connections and single-head self-attention for reduced computational overhead and enhanced small-target representation.
- Introduced an adaptive sparse feature interaction (ASFI) module to dynamically fuse attention branches for improved discriminative information concentration.
- Designed a multi-scale dilated re-parameterization block (MSDRB) to expand receptive fields and incorporate multi-scale context without increasing computational cost.
Main Results:
- CEAM-DETR demonstrated superior detection accuracy and robustness compared to state-of-the-art methods on public datasets.
- The model achieved a 1.5% increase in accuracy, reduced parameters by 36.5%, decreased GFLOPs by 26.0%, and improved Frames Per Second (FPS) by 32.9% compared to RT-DETR.
- Validation confirmed the model's effectiveness in complex agricultural environments.
Conclusions:
- CEAM-DETR offers an efficient and stable solution for weed detection in challenging agricultural scenarios.
- The proposed NMS-free, end-to-end detection paradigm significantly enhances performance metrics.
- The model's lightweight and adaptive nature makes it suitable for practical agricultural applications.
