Rubber tree (Hevea brasiliensis) powdery mildew (PM) is a devastating foliar disease causing annual yield losses exceeding 30% in severely affected plantations. Automated multi-grade detection in complex field conditions is hampered by three technical gaps: (i) CNN-based detectors lack global context for disambiguating fine-grained early lesions; (ii) Transformer detectors incur quadratic complexity <InlineEquation ID="IEq1"><EquationSource Format="TEX">$O(n^2)$</EquationSource><EquationSource Format="MATHML"><math><mrow><mi>O</mi><mo stretchy="false">(</mo><msup><mi>n</mi><mn>2</mn></msup><mo stretchy="false">)</mo></mrow></math></EquationSource></InlineEquation> impractical for edge deployment; and (iii) conventional multi-scale fusion inadequately bridges the semantic gap across disease scales. We propose LightPM-DETR, a lightweight detection transformer addressing these gaps through a co-designed architecture–compression pipeline. The encoder incorporates four modules: (1) Selective Feature Scanning (SFS), a newly designed tri-stream Mamba adaptation for O(n)-complexity long-range modelling; (2) Poly-Kernel Inception Net (PKI Net), adapted for multi-scale receptive field aggregation; (3) Wavelet Efficient Attention (WEA), a new Haar-basis frequency-domain key-value compression mechanism for lesion-sensitive attention; and (4) Global Query Aggregator (GQA), a new hierarchical cross-attention module for query enrichment. Structured channel pruning followed by IoU-weighted knowledge distillation further compresses the model by 85.3% in parameters relative to RT-DETR-L (from 32.7 M to 4.8 M). On the self-built PM-Dataset-Plus (6-grade, 8,412 images, 47,836 instances) and PD40 (40-class cross-species) benchmarks, LightPM-DETR achieves mAP<InlineEquation ID="IEq2"><EquationSource Format="TEX">$_{50}$</EquationSource><EquationSource Format="MATHML"><math><mmultiscripts><mrow /><mn>50</mn><mrow /></mmultiscripts></math></EquationSource></InlineEquation> of 91.6% and 84.3% respectively (mean over three runs), surpassing all baselines. The compressed variant operates at 4.8 M parameters and 6.2 G FLOPs, achieving 62.4 FPS on a desktop GPU and 23.8 FPS on an NVIDIA Jetson Orin Nano. Code and data will be released at <a href="https://github.com/wfcyliyuheng-dev/PM-Dataset-PLus">https://github.com/wfcyliyuheng-dev/PM-Dataset-PLus</a> upon acceptance.

LightPM-DETR: a lightweight transformer for grading detection of rubber tree powdery mildew

期刊论文
出版模式：

开放获取
发表日期：

2026年6月30日

查看全文

Licheng Zhang,
Yuheng Li

摘要

Rubber tree (Hevea brasiliensis) powdery mildew (PM) is a devastating foliar disease causing annual yield losses exceeding 30% in severely affected plantations. Automated multi-grade detection in complex field conditions is hampered by three technical gaps: (i) CNN-based detectors lack global context for disambiguating fine-grained early lesions; (ii) Transformer detectors incur quadratic complexity $O(n^2)$ $O (n^{2})$ impractical for edge deployment; and (iii) conventional multi-scale fusion inadequately bridges the semantic gap across disease scales. We propose LightPM-DETR, a lightweight detection transformer addressing these gaps through a co-designed architecture–compression pipeline. The encoder incorporates four modules: (1) Selective Feature Scanning (SFS), a newly designed tri-stream Mamba adaptation for O(n)-complexity long-range modelling; (2) Poly-Kernel Inception Net (PKI Net), adapted for multi-scale receptive field aggregation; (3) Wavelet Efficient Attention (WEA), a new Haar-basis frequency-domain key-value compression mechanism for lesion-sensitive attention; and (4) Global Query Aggregator (GQA), a new hierarchical cross-attention module for query enrichment. Structured channel pruning followed by IoU-weighted knowledge distillation further compresses the model by 85.3% in parameters relative to RT-DETR-L (from 32.7 M to 4.8 M). On the self-built PM-Dataset-Plus (6-grade, 8,412 images, 47,836 instances) and PD40 (40-class cross-species) benchmarks, LightPM-DETR achieves mAP$_{50}$ $_{50}$ of 91.6% and 84.3% respectively (mean over three runs), surpassing all baselines. The compressed variant operates at 4.8 M parameters and 6.2 G FLOPs, achieving 62.4 FPS on a desktop GPU and 23.8 FPS on an NVIDIA Jetson Orin Nano. Code and data will be released at https://github.com/wfcyliyuheng-dev/PM-Dataset-PLus upon acceptance.