A Medicinal Plant Leaf Image Dataset from Bangladesh for Deep Learning–Based Recognition
Description
The dataset titled “A Medicinal Plant Leaf Image Dataset from Bangladesh for Deep Learning–Based Recognition” comprises a total of 6,419 high-resolution JPG images of six medicinal plant species collected from the rural jungle area of Guadanga village, Phulpur, Mymensingh, Bangladesh. Leaf samples were photographed indoors under controlled lighting conditions with varied natural backgrounds using Google Pixel 5 and Google Pixel 6a smartphones to ensure consistent color fidelity and sharpness suitable for computer-vision and deep-learning research. The dataset is organized into six species-specific folders, each representing a commonly found medicinal plant in Bangladesh: • Centella Asiatica • Coccinia Grandis • Eclipta Prostrata • Mikania Micrantha • Murraya Koenigii • Stephania Japonica The collection includes 1,678 original images, 1,678 annotated images, and 3,063 augmented images generated through horizontal and vertical flipping, grayscale conversion, and geometric transformations. All images were auto-oriented and resized uniformly to 640 × 640 pixels to maintain consistency across the dataset. These augmentations increase data diversity and support the development of robust deep-learning models for medicinal-leaf classification, detection, and segmentation.
Files
Steps to reproduce
Object Detection and Annotation For object-detection tasks, leaf regions were annotated with bounding boxes using the Roboflow tool. Each annotation was manually verified to ensure labeling consistency and accuracy across all six plant species. The annotated subset facilitates applications in object detection, image segmentation, and automated plant recognition. Dataset Partitioning To support reproducible training and evaluation, the dataset was stratified into three subsets as follows: Training Set: 77% of total images (2,363 samples) Validation Set: 13% of total images (400 samples) Test Set: 10% of total images (300 samples) This split ensures balanced representation of all species across subsets and allows consistent evaluation of deep-learning models. Preprocessing and Augmentation Summary Process Details Auto-Orientation Applied to all images Resize Uniformly stretched to 640 × 640 pixels Augmentations per Example 3 Applied Augmentations Horizontal & vertical flips; grayscale (15%) Dataset Attributes Attribute Details Total Images 6,419 (JPG) Original Images 1,678 Annotated Images 1,678 (TXT in YOLO format) Augmented Images 3,063 Plant Species (Classes) 6 Capture Devices Google Pixel 5 and Google Pixel 6a Resolution (After Processing)640 × 640 pixels Source Location Jungle of Guadanga, Phulpur, Mymensingh, Bangladesh Annotation Tool Roboflow Applications Deep learning, medicinal-plant recognition, AI-based botany, image segmentation
Institutions
- American International University Bangladesh
- North South University
- Chittagong University of Engineering and Technology
- Sher-e-Bangla Agricultural University