地黄是我国重要的道地中药材，规模化种植的地黄常受多种害虫侵袭，导致叶片孔洞频现，严重影响地黄的产量与品质。然而，当前基于深度学习的虫害孔洞检测方法面临田间环境复杂、目标尺度小、形态不规则等挑战。本研究构建了面向田间复杂环境的地黄叶片虫害孔洞图像数据集（Rehmannia Pest-induced Hole Dataset, RPHD）。该数据集包含初始采集版本（298张图像，5 059个标注）与精校增强版本（291张图像，2 678个高质量标注）。通过系统化的图像预处理流程，包括基于内容感知的裁剪与分辨率统一（1024×1024）、自适应双边滤波去噪与光照均衡化，并结合精细化人工标注与多轮交叉校验质量控制机制，显著提升了数据集的目标判别性。实验采用YOLOv10n、YOLOv11n和YOLO12n三种主流目标检测模型对数据集进行基准评估，对比两个版本在不同模型上的性能表现。结果表明，在三个模型上，使用精校增强数据集训练均取得显著性能提升：YOLOv10n的mAP@0.5从40.3%提升至87.2%，YOLOv11n从48.9%提升至92.1%，YOLO12n从51.2%提升至93.0%。该数据集为地黄小目标病虫害检测算法研究提供了标准化、高质量的基准数据支持，可供其他农作物相关研究借鉴，有助于推动农业视觉检测技术的实用化与智能化发展。

数据摘要：

Rehmannia glutinosa is an important traditional Chinese medicinal herb. Its large-scale cultivation is often affected by various pests, leading to frequent leaf pest-induced holes that severely impact yield and quality. However, current deep learning-based Pest-induced Hole detection methods face challenges such as complex field environments, small target scales, and irregular morphological characteristics. To address these issues, this paper presents the first Rehmannia Leaf Pest-induced Hole Dataset (RPHD) specifically designed for complex field environments. The dataset comprises an initial version (298 images with 5,059 annotations) and a refined version (291 images with 2,678 high-quality annotations). Through a systematic image preprocessing pipeline, including content-aware cropping and resolution normalization (1024×1024), adaptive bilateral filtering for noise reduction, and illumination equalization, combined with meticulous manual annotation and multiple rounds of cross-validation quality control mechanisms, the annotation consistency and target discriminability of the dataset were significantly improved. Three mainstream object detection models—YOLOv10n, YOLOv11n, and YOLOv12n—were employed to evaluate the dataset, comparing the performance of the two versions across different models. Experimental results show that training on the refined dataset yields substantial performance improvements across all models:YOLOv10n: mAP@0.5 increased from 40.3% to 87.2%;YOLOv11n: mAP@0.5 increased from 48.9% to 92.1%;YOLO12n: mAP@0.5 increased from 51.2% to 93.0%. This dataset provides standardized, high-quality benchmark data for research on small-target pest and disease detection algorithms for Rehmannia and other crops, contributing to the practical and intelligent development of agricultural vision detection technologies.

Data summary: