Preprocessed Brinjal Leaf Image Dataset for Healthy and Diseased Leaf Detection (Phomopsis Blight and Little Leaf) Using Deep Learning

Published: 12 July 2026| Version 2 | DOI: 10.17632/c39p3k87g2.2
Contributors:
,
,
,

Description

This dataset contains 13,600 preprocessed images of Solanum melongena (Brinjal) leaves collected at Vidhani Village, Jaipur District, Rajasthan, India (26°46'22.4"N, 75°52'18.2"E) for automated plant disease detection using deep learning. VERSION 2 CHANGES (from v1): - Resolution upgraded: 224×224 → 512×512 pixels - Raw and Augmented images stored in separate directories - master metadata.csv added (split assignments, device, lighting) - Data splitting done at original-image level before augmentation DATASET STRUCTURE: Brinjal_Final_Preprocessed/ ├── Raw/ → 850 originals at 512×512 │ ├── Healthy_Leaves/ (400 images) │ ├── Phomopsis_Blight/ (250 images) │ └── Little_Leaf/ (200 images) ├── Augmented/ → 12,750 variants at 512×512 │ ├── Healthy_Leaves/ (6,000 images) │ ├── Phomopsis_Blight/ (3,750 images) │ └── Little_Leaf/ (3,000 images) └── metadata.csv → 13,600 rows master ledger CLASSES: - Healthy_Leaves: uniform green, intact margins, no lesions - Phomopsis_Blight: necrotic lesions, brown/yellow spots (Phomopsis vexans) - Little_Leaf: stunted, curled, pale/white leaves (Leaf Curl) IMAGE ACQUISITION: Captured using Google Pixel 8 (50MP) and Samsung Galaxy M52 5G (64MP) under natural outdoor light (08:00–17:00 IST). 50 images excluded after quality screening; pHash deduplication confirmed no near-duplicates among 850 finals. AUGMENTATION (15 techniques per original): V1 Grayscale | V2 CLAHE (clipLimit=2.0, tile=8×8) | V3 Gamma (γ=1.5) | V4 HSV (sat×1.2) | V5 Brightness (×1.3) | V6 Contrast (×1.4) | V7 Sharpening (3×3 kernel) | V8 Gaussian Blur (5×5,σ=1) | V9 Median Blur (k=3) | V10 Bilateral (d=5,σ=75) | V11 Top-Hat (15×15) | V12 Black-Hat (15×15) | V13 Rotation (+25°) | V14 Horizontal Flip | V15 Unsharp Mask (9×9,σ=10) NAMING CONVENTION: Raw → ClassName_SerialNumber_Original.jpg Augmented → ClassName_SerialNumber_Technique.jpg METADATA.CSV FIELDS: filename | class_label | original_image_id | preprocessing_technique | data_split | device_used | lighting_condition Split 70:15:15 assigned at original-image level. All 15 variants of any leaf stay within the same partition. REAL-WORLD DEPLOYMENT: ResNet50V2 trained on this dataset powers Fasal Saathi, a real-time brinjal disease detection app by MaxBrain Technologies (Google Play Store), confirming production-level robustness. Tools: Python, OpenCV, NumPy, PIL, TensorFlow | Google Colab License: CC BY 4.0

Files

Steps to reproduce

STEPS TO REPRODUCE — Version 2 1. DOWNLOAD AND EXTRACT Download the ZIP. It contains Raw/, Augmented/, metadata.csv. 2. READ METADATA FIRST — DO NOT RANDOM SPLIT Load metadata.csv to get pre-assigned train/val/test splits: import pandas as pd meta = pd.read_csv('metadata.csv') train = meta[meta['data_split'] == 'train'] val = meta[meta['data_split'] == 'val'] test = meta[meta['data_split'] == 'test'] Random splitting on files risks placing augmented variants of the same leaf into different partitions (data leakage). 3. LOAD IMAGES All images are 512×512 RGB JPG. Resize at training time if your model requires a different input size: import cv2 img = cv2.imread('path/to/image.jpg') # shape: (512,512,3) img = img / 255.0 # normalize to [0,1] 4. CHOOSE EXPERIMENTAL SETUP - Raw only (850 imgs): unaugmented baseline experiments - Augmented only (12,750 imgs): preprocessing impact studies - Full dataset (13,600 imgs): combined training pipeline Use metadata.csv file_path column to build your loader. 5. CLASSIFICATION OPTIONS - Binary: Healthy_Leaves vs {Phomopsis_Blight + Little_Leaf} - Multi-class: 3 separate labels (Healthy, Phomopsis, Little) 6. RECOMMENDED TRAINING CONFIGURATION Batch size: 32 | Optimizer: Adam (lr=0.0001) Loss: Categorical Cross-Entropy Callbacks: EarlyStopping + ReduceLROnPlateau Input shape: (512, 512, 3) 7. BENCHMARK RESULTS ON THIS DATASET Normal CNN: 89.84% accuracy (F1: 0.895) MobileNetV2: 90.63% accuracy (F1: 0.903) ResNet50V2: 92.97% accuracy (F1: 0.928) 8. REPRODUCE AUGMENTATION (OPTIONAL) All 15 techniques implemented in Python (OpenCV, NumPy, PIL) in Google Colab. Parameters for each technique are documented in the Description above and in Table 3 of the accompanying Data in Brief article (DOI: 10.17632/c39p3k87g2.2). 9. FOLDER PATHS REFERENCE Raw originals: Raw/Healthy_Leaves/Healthy_Leaves_001_Original.jpg Raw/Phomopsis_Blight/Phomopsis_Blight_001_Original.jpg Raw/Little_Leaf/Little_Leaf_001_Original.jpg Augmented variants: Augmented/Healthy_Leaves/Healthy_Leaves_001_CLAHE.jpg Augmented/Phomopsis_Blight/Phomopsis_Blight_023_Gamma.jpg Augmented/Little_Leaf/Little_Leaf_001_Gaussian_Blur.jpg

Institutions

Categories

Artificial Intelligence, Computer Vision, Image Processing, Machine Learning, Plant Pathology, Precision Agriculture, Deep Learning

Licence