DSA-DET: a tea disease detection algorithm based on dynamic spatial pyramid and polarized linear attention.

Abstract

Introduction Traditional tea disease detection methods suffer from low efficiency and strong subjectivity, while existing deep learning approaches often demonstrate inadequate detection accuracy and poor real-time performance in complex and variable environments. Methods Here, we present an intelligent tea disease detection method based on an improved Real-Time Detection Transformer, named DSA-DET. We propose a backbone network based on dynamic attention spatial pyramid modeling to achieve more accurate collaborative modeling of local features and global context. We design an encoder combining polarized linear attention with parallel spatial enhancement networks and multi-scale adaptive enhancement to improve feature extraction capabilities. We further develop an upsampling module employing efficient spatial-channel upsampling and shift mixing mechanisms to enhance the quality of reconstructed features. Results Experimental results show that the improved model achieves a precision of 94.73%, a recall of 89.65%, and an mAP 50 of 93.68%. Compared to the baseline model RT-DETR-R18, its precision is improved by 3.56%, recall by 2.96%, and mAP 50 by 3.02%; meanwhile, the model maintains a lightweight parameter scale of 15.4M and a real-time detection speed of 71.5 FPS. Discussion The improvement scheme in this study successfully enhances detection accuracy while maintaining a good balance between model complexity and inference speed, providing a practical and reliable technical solution for the intelligent diagnosis of tea diseases.

Plant phenotyping relevance

茶葉の病害状態を画像等から検出する深層学習手法を開発し、精度・再現速度を比較検証しており、植物病害表現型の取得手法が中心である。

Code and data availability

The paper describes a self-constructed AGTea tea-disease image dataset (5,972 images) and a DSA-DET model, but the supplied blocks contain no data availability statement, repository deposit, or authors' public URL for the dataset, images, annotations, code, or trained model. No paper-specific public asset is actionable

No evidence-backed public reproduction asset is currently recorded.