BLD-BD: A Field-Based Image Dataset of Bean Leaf Diseases for Deep Learning & Computer Vision Research

Description

Overview: This dataset contains high-resolution images of bean leaves collected directly from agricultural fields in Kulchara village, located in the Jhenaidah district of Bangladesh. It is specifically curated for computer vision tasks such as plant disease detection, classification, and deep learning model benchmarking (e.g., ResNet, MobileNet, EfficientNet). The dataset comprises three distinct classes: 1. Angular Leaf Spot 2. Bean Common Mosaic Virus 3. Healthy All images in both sub-datasets have been standardized to a resolution of 512 × 512 pixels in RGB format, ensuring clean background contrast and clear presentation of individual leaf disease symptoms. Dataset Structure & Content The dataset is divided into two separate compressed ZIP files: 1. Original_Bean_Leaf_Disease_Dataset.zip • Total Images: 672 original images • Distribution: Balanced across 3 classes (224 images per class) • Resolution: 512 × 512 pixels • Purpose: Serves as the raw, baseline field dataset for evaluation and baseline model benchmarking. 2. Augmented_Bean_Leaf_Disease_Dataset.zip • Total Images: 3,360 images • Distribution: 1,120 images per class (224 original + 896 augmented images) • Resolution: 512 × 512 pixels • Augmentation Techniques Applied: Controlled rotation (±15°), horizontal flipping, subtle random zoom/crop (5 - 10%), light exposure/brightness adjustments, and minimal Gaussian blur. • Note: Color signatures (hue/saturation) were strictly preserved to maintain the visual integrity of brown spots and green mosaic patterns vital for classification. Data Collection & Methodology The primary image acquisition was conducted under natural environmental lighting conditions in the field to reflect real-world agricultural scenarios. Data preprocessing included strict filtering to remove noisy background artifacts and non-target plant species. • Location of Collection: Village: Kulchara, District: Jhenaidah, Bangladesh • File Format: .jpg Potential Applications • Computer Vision and Agricultural AI Research • Single-label Leaf Disease Classification • Transfer Learning and Hybrid Deep Learning Model Training • Comparative Performance Analysis of CNN and Vision Transformer Models

Steps to reproduce

1. Image Acquisition & Preprocessing: • Raw field images of bean leaves across three classes (Angular Leaf Spot, Bean Common Mosaic Virus, and Healthy) were collected from agricultural fields in Kulchara village, Jhenaidah, Bangladesh. • All raw images were resized to a standardized resolution of 512x512 pixels using OpenCV with INTER_AREA interpolation (Original_Bean_Leaf_Disease_Dataset.zip). 2. Sequential Cluster Renaming: • Images were renamed using class-specific prefixes (ALS_, BCMV_, HLT_) to maintain 1:1 cluster traceability between original and augmented samples. 3. Data Augmentation: • To build the augmented dataset (Augmented_Bean_Leaf_Disease_Dataset.zip), each resized original image generated 4 corresponding augmented images using the Albumentations library (1:4 ratio, totaling 3,360 images). • Augmentation Pipeline: • Horizontal Flip (p=0.5) • Random Rotation (±15°, p=0.7) • Random Resized Crop (5–10%, p=0.6) • Random Brightness & Contrast Adjustment (±10%, p=0.5) • Gaussian Blur (3×3 kernel, p=0.3) 4. Environment: • Code executed in Python 3 using OpenCV, Albumentations, NumPy, and Google Colab.

Institutions

• Daffodil International University

  Dhaka Division, Dhaka

Categories

Licence