Random Rolling Attention Augmentation for Efficient Agricultural Disease and Pest Detection

Abstract

Abstract Accurate detection of agricultural diseases and pests is essential for crop protection and food security worldwide. Real-world field applications face challenges including small objects, complex backgrounds, high class similarity, and limited computational resources. This work proposes a Random Rolling Transformer (RRT), which introduces random circular shifts along channel and sequence dimensions into multi-head self-attention to enrich feature interactions without increasing parameters or computation. Integrated into YOLOv12, the proposed RRT-YOLO is evaluated on the IP102 pest dataset and a tomato leaf disease dataset. Results show that RRT-YOLO improves mAP@50 by 2.5% and mAP@50–95 by 3.9% on IP102, and by 8.8% and 4.2% on the tomato disease dataset, while maintaining identical model size and complexity. This attention perturbation strategy offers an effective and efficient solution for lightweight agricultural vision detection and can be extended to other visual computing tasks. The code and detailed descriptions can be accessed via the following repository: \href{https://github.com/glorioustory/Random-Rolling-Attention-Augmentation.git}{https://github.com/glorioustory/Random-Rolling-Attention-Augmentation.git}.

Plant phenotyping relevance

植物葉の病害状態を画像から検出する計算手法を開発・評価しており、植物病害の表現型取得が技術的中心である。害虫検出も含むが、トマト葉病害データセットで手法性能を検証している。

Code and data availability

The authors explicitly state their code and detailed descriptions are publicly available in a GitHub repository containing the RRT-YOLO implementation used for the paper's disease/pest detection experiments. The IP102 and tomato leaf disease datasets are third-party public resources, not paper-specific assets.