CED-RTDETR: a contour-aware evidence-guided decoupled network for rice leaf disease detection.

Abstract

To address the challenges posed by large lesion-scale variations, weak boundary cues, and high inter-class similarity in rice leaf disease detection, this study proposes an improved RT-DETRv2-R50-based rice disease detection model, termed CED-RTDETR. First, a Lightweight Contour-Guided Aggregation Backbone (LCGA-Backbone) is constructed. In the PResNet residual blocks, an InceptionDWConv2d-based direction-aware depthwise separable spatial mixing strategy is introduced to capture local, horizontal, and vertical lesion texture patterns with low computational overhead. Meanwhile, a Contour-guided Efficient Global Aggregation Block (CEGA Block) is embedded after the outputs of the C3, C4, and C5 stages. Through contour-difference enhancement, channel shuffle, group-wise efficient global aggregation, and bottleneck channel mixing, the proposed block strengthens the representation of the boundaries of small lesions, weak textures, and contextual semantics. Second, a Multi-scale Evidence Interaction Fusion Neck (MEIF-Neck) is used to perform multi-scale evidence interaction and salient-region competition after cross-scale feature concatenation. A lightweight feature reconstruction process is further implemented using Spatial-Evidence RepNCSPELAN (SE-RepNCSPELAN), which is developed from a YOLO-style feature fusion structure. Finally, a Direction-Amplitude Decoupled Deformable Attention (DAD-DA) mechanism is introduced to decompose sampling offsets into direction rotation residuals and radius gains while incorporating an aspect-ratio-aware geometric compression-restoration strategy, thereby improving the geometric stability of sampling locations during the decoding stage. Experimental results on the constructed rice leaf disease dataset show that CED-RTDETR achieves AP, AP50, and AP75 values of 29.5%, 75.2%, and 17.6%, respectively, outperforming RT-DETRv2-R50 by 4.1, 4.9, and 3.8 percentage points. To further evaluate the model on an additional public benchmark dataset, experiments were also conducted on the public Rice Disease Dataset. On this dataset, CED-RTDETR achieves AP, AP50, and AP75 values of 34.1%, 77.1%, and 23.2%, respectively, improving upon RT-DETRv2-R50 by 4.3, 5.6, and 3.4 percentage points. These results indicate that the proposed method achieves consistent overall performance improvements on both the constructed dataset and the public benchmark dataset.

Plant phenotyping relevance

イネ葉の病斑・病害を画像から検出する深層学習モデルを開発し、構築データセットと公開ベンチマークで性能検証しているため、植物病害状態のフェノタイピング手法が中心です。

Code and data availability

The supplied blocks contain no data availability statement, no public dataset URL, and no author code/model deposit. The constructed rice leaf disease dataset and CED-RTDETR code are not described as publicly available. The only URL present (https://github.com/ultralytics/ultralytics) is a cited generic library, not a纸

No evidence-backed public reproduction asset is currently recorded.