MedHerbs: An Image Dataset of Bangladeshi Medicinal Herbs

MedHerbs is a publicly available image dataset of 35 Bangladeshi medicinal herb categories, comprising 29,225 RGB images (835 images per class) in PNG format. Physical herb specimens were sourced from Hamdard Laboratories (WAQF), Bangladesh, ensuring authenticity and correct labelling across all categories. Images were captured using two smartphone cameras (Samsung Galaxy M31 and Apple iPhone 13) under varying lighting conditions (daylight, low-light, shadowed, mixed-light) and viewpoints. Raw images were processed through a seven-stage pipeline YOLO-based object detection, SAM3 segmentation, background removal, cropping and spatial alignment, and CLAHE lighting normalisation to produce clean, aligned RGB images. Each of the 35 classes is stored in a numerically named folder (0–34) following the path convention Images/{folder_id}. An accompanying metadata file (metadata.csv / metadata.json) maps each folder_id to its scientific name, Bangla name, medicinal plant part (bark, leaves, fruit, seed, root, or flower), a short description, and image count. A classes.txt file provides a simple folder_id → label mapping for direct use with ImageFolder-style data loaders. This dataset is intended to support research in medicinal plant recognition, fine-grained image classification, ethnobotany, and mobile health application development within the context of South Asian traditional medicine.