WMC-RTDETR: a lightweight tea disease detection model.

Abstract

Tea pest and disease detection is crucial in tea plantation management, however, challenges such as multi-target occlusion and complex background impact detection accuracy and efficiency. To address these issues, this paper proposes an improved lightweight model, WMC-RTDETR, based on the RT-DETR model. The model significantly enhances the ability to capture multi-scale features by introducing wavelet transform convolution, improving the feature extraction accuracy in complex backgrounds, and increasing detection efficiency while reducing the number of model parameters. Combined with multiscale multihead self-attention, global feature fusion across scales is realized, which effectively overcomes the shortcomings of traditional attention mechanisms in small target detection. Additionally, a context-guided spatial feature reconstruction feature pyramid network is designed to refine the target feature reconstruction through contextual information, thereby improving the robustness and accuracy of target detection in complex scenes. Experimental results show that the proposed model achieves 97.7% and 83.1% respectively in mAP50 and mAP50:95 indicators, which outperform the original model. In addition, the number of parameters and floating-point operations are reduced by 35.48% and 40.42% respectively, enabling highly efficient and accurate detection of pests and diseases in complex scenarios. Furthermore, this paper successfully deploys the lightweight model on the Raspberry Pi platform, which proves that it has good real-time performance in resource-constrained embedded environments, providing a practical solution for low-cost disease monitoring in agricultural scenarios.

Plant phenotyping relevance

茶葉の病害虫を画像から検出するモデルを開発・評価し、病害状態の推定性能と組込み実装まで検証しているため、植物フェノタイピング手法が中心である。

Code and data availability

The article describes a custom tea pest/disease image dataset (160 images collected at Maoshan Tea Garden, augmented to 1280) and the WMC-RTDETR model, but no block contains any data availability statement, public repository deposit, or author-provided URL for the dataset, images, code, or trained model. No qualifying,

No evidence-backed public reproduction asset is currently recorded.