Rehmannia Pest-Hole Dataset (RPHD): A High-Quality Image Dataset for Leaf Pest-Hole Detection

This dataset is the Rehmannia Pest-Hole Dataset (RPHD), which consists of two subsets: the raw version and the enhanced version.All images were collected from June to September 2024 during the critical growth period of Rehmannia glutinosa in the core planting area of Wen County, Henan Province, under natural field conditions. The acquisition devices included a DJI Mini 3 Pro drone and a XIAOMI 14 smartphone, with photography conducted during favorable natural lighting conditions from 9:00–11:00 AM and 2:00–4:00 PM. A total of 298 raw images were captured, covering diverse real-world scenarios such as close-ups of single pest holes, complex scenes with multiple pest holes, varying illumination directions, and different degrees of leaf occlusion.The raw version of the dataset retains the original image resolution, natural noise, and complex backgrounds without any preprocessing, faithfully representing the natural appearance of Rehmannia pest holes. It contains 5,059 annotated pest hole instances. The enhanced version was constructed by applying systematic image preprocessing to the raw images, including guided cropping to remove background redundancy, uniform resizing to 1024×1024 pixels, adaptive bilateral filtering for noise reduction while preserving edge sharpness, and brightness–contrast equalization to enhance the visual contrast between pest holes and healthy leaf tissue. The enhanced version comprises 2,678 high-quality pest hole annotations.Data annotation was performed using LabelImg, with bounding boxes saved in YOLO format. All targets are pest holes, with a unified class ID of 0. Each image was independently annotated by at least two annotators, and any discrepancies were resolved by a third annotator. Annotation consistency was ensured by requiring an Intersection over Union (IoU) greater than 0.8, guaranteeing the accuracy and reliability of the annotations.The dataset is organized into two separate subfolders by version. In the raw version folder, the images directory contains 298 raw images, and the labels directory contains the corresponding 5,059 annotation files. The dataset is split into training (208 images), validation (59 images), and test (31 images) sets according to a 7:2:1 ratio, with the split information provided in train.txt, val.txt, and test.txt files. In the enhanced version folder, the images directory is further divided into train, val, and test subdirectories, containing 203, 58, and 30 images respectively, while the labels directory mirrors this structure with the corresponding annotation files. Image files are named following the pattern “device abbreviation_date_sequence.jpg” (e.g., DJI_20240720_001.jpg), and the enhanced version appends “_cv” to the original filename (e.g., IMG_20240720_121547_cv.jpg). Annotation files share the same name as their corresponding images with a .txt extension. All images are in JPEG/JPG format; the raw version retains the original resolution, while the enhanced version is uniformly resized to 1024×1024 pixels. Annotation files follow the YOLO format, with each line containing the class ID and normalized bounding box coordinates (center x, center y, width, height).During the cropping process for the enhanced version, some regions with insignificant pest holes from the raw images may have been discarded, but all cropped images maintain a strict one-to-one correspondence with their annotations. Annotation errors are minimized through rigorous cross-checking and IoU threshold control, ensuring accurate bounding box placement and size. The preprocessing steps introduced no geometric distortions or annotation shifts. Both image and annotation files are in common formats widely used in object detection and can be directly utilized by mainstream deep learning frameworks and detection algorithms.Comparative experiments using YOLOv11n demonstrate that models trained on the raw version achieve a mean Average Precision (mAP@0.5) of 40.3%, while those trained on the enhanced version achieve an mAP@0.5 of 89.2%, confirming the critical role of high-quality data in improving model performance.