基于关键特征图的可变形 DETR 在复杂场景中对伪装蝗虫目标的检测
key-fg DETR based camouflaged locust objects in complex fields.
作者信息
Chen Dongmei, Cao Peipei, Diao Zhihua, Dong Yingying, Zhang Jingcheng
机构信息
College of Artificial Intelligence, Hangzhou Dianzi University, Hangzhou, China.
School of Electrical Information Engineering, Zhengzhou University of Light Industry, Zhengzhou, China.
出版信息
Front Plant Sci. 2025 Jul 23;16:1565739. doi: 10.3389/fpls.2025.1565739. eCollection 2025.
INTRODUCTION
In real agricultural environments, many pests camouflage themselves against complex backgrounds, significantly increasing detection difficulty. This study addresses the challenge of camouflaged pest detection.
METHODS
We propose a Transformer-based detection framework that integrates three key modules: 1.Fine-Grained Score Predictor (FGSP) - guides object queries to potential foreground regions; 2.MaskMLP generates instance-aware pixel-level masks; 3.Denoising Module and DropKey strategy - enhance training stability and attention robustness.
RESULTS
Evaluated on the COD10k and Locust datasets, our model achieves AP scores of 36.31 and 75.07, respectively, outperforming Deformable DETR by 2.3% and 3.1%. On the Locust dataset, Recall and F1-score improve by 6.15% and 6.52%, respectively. Ablation studies confirm the contribution of each module.
DISCUSSION
These results demonstrate that our method significantly improves detection of camouflaged pests in complex field environments. It offers a robust solution for agricultural pest monitoring and crop protection applications.
引言
在真实的农业环境中，许多害虫会在复杂背景下伪装自己，这显著增加了检测难度。本研究旨在应对伪装害虫检测的挑战。
方法
我们提出了一种基于Transformer的检测框架，该框架集成了三个关键模块：1. 细粒度分数预测器（FGSP）——引导目标查询到潜在的前景区域；2. MaskMLP生成实例感知的像素级掩码；3. 去噪模块和DropKey策略——增强训练稳定性和注意力鲁棒性。
结果
在COD10k和蝗虫数据集上进行评估，我们的模型分别取得了36.31和75.07的AP分数，比可变形DETR分别高出2.3%和3.1%。在蝗虫数据集上，召回率和F1分数分别提高了6.15%和6.52%。消融研究证实了每个模块的贡献。
讨论
这些结果表明，我们的方法显著提高了在复杂田间环境中伪装害虫的检测能力。它为农业害虫监测和作物保护应用提供了一个强大的解决方案。
Zotero 插件